# State-Level Determinants of Rhetorical Position-Taking on Terrorism Author: Charlie Carter Status: Working paper; undated ## Abstract The United Nations has not agreed on a single comprehensive definition of terrorism despite decades of debate, and existing accounts generally attribute this stalemate to either cynical posturing or genuine contestation over meaning. Yet we still lack broad, text‑sensitive evidence showing how states actually use the term and which mechanisms, particularly those concerning material and strategic factors, structure variation in rhetoric across time and place. I assemble every identifiable English‑language use of “terrorism” in United Nations General Debate speeches from 1947–2022 (20,109 instances) and extend conText embedding regression with multiple imputation, fixed‑effect removal in embedding space, and projection onto three spatial dimensions that capture states positions on salient issues in the international politics of terrorism—dimensions pertaining to terrorism’s ‘Causes’, ‘Responses’, and ‘Perpetrators’. Between countries, state attributes and politico‑geographic regions strongly predict terrorism rhetoric; within countries, changes explain little movement, with the operation of a legislature the only robust signal in baseline models, and with higher democracy associated with movement toward terrorism rhetoric invoking “root causes” and state violence, and higher domestic terrorism associated with movement toward state perpetrators. These patterns imply that international rhetoric on terrorism is highly structured between countries but its evolution within countries finds limited explanations in state-level covariates. The findings refine debates over ambiguity versus contestation in the international politics of terrorism and provide an adaptable measurement framework for inference on meaning in large political text corpora. ## Links - [PDF (full text)](https://www.charlieacarter.com/papers/paper1.pdf) - [HTML landing page](https://www.charlieacarter.com/research/paper1.html) - [BibTeX](https://www.charlieacarter.com/research/paper1.bib) ## Version and source This source-derived Markdown corresponds to the published PDF with SHA-256 `ce9b52f1d8dbd7fbbc7faa82e55d11c49532e57edc849151706b505c82d95c1a`. Equations retain LaTeX notation; citation formatting and layout differ from the PDF. Use the PDF for page numbers. ## Introduction What are the state-level determinants of how states use the term “terrorism” in international fora? The starting point of this enquiry is the assertion that the features which constitute an act of terrorism are contested. This is true in both semantic and legal terms: disparities exist in what is considered the proper meaning of “terrorism”, who can be considered a “terrorist”, and in how terrorism is defined in sub-national, national, and international formal bodies of law. Further, this contestation has been noted on a number of salient dimensions. Terrorism may be more or less political, rational or irrational, a weapon of the weak or instrument of the state, a tactic or an identity, and completely illegitimate or sometimes justified. This semantic ambiguity permits states flexibility in how they use the term “terrorism”. Understanding how state interests, attributes, and policies shape the rhetorical usage of “terrorism” in light of this flexibility would shed light on, firstly, how meaning is negotiated by disparate actors in international politics and, secondly, whether or not these constructions can plausibly influence international outcomes. Terrorism’s contested meaning has been evident in international politics. Attempts to define terrorism in international law span nearly a century, “suggesting that the \[international community\] attaches considerable importance to definition” (Saul 2005, 58). Member-states of the United Nations (UN) have made perennial attempts to produce a comprehensive, consensus definition of terrorism within the UN General Assembly since 1972. These attempts have not yet been successful and, notably, their failure has been regularly attributed to international disagreement over the precise content of proposed definitions and the actions and actors to which they should refer. In addition, states articulate and apply the concept of terrorism differently in practice. Stark differences exist amongst even “formally allied countries with considerable records of cooperation” in who is considered a “terrorist” (Jarvis and Legrand 2018, 201) and the groups which are formally proscribed as “foreign terrorist organisations” on national lists (Haspeslagh 2013; El Masri and Phillips 2021). Such disparate uses of the term terrorism may reflect genuine international disagreement over its proper semantic content, but alternative explanations remain plausible. States may articulate terrorism in ways which are consistent with the preferences and understandings of actors in domestic politics or seek to reconcile domestic and international counterterrorism agendas. Alternatively, semantic contestation may emerge not as a result of differences in states’ interests, but precisely because these interests are not jeopardised by strategically inconsequential and costless “cheap talk”. Or, if “talk” can be persuasive and modify how others perceive key issues, states may contest the meaning of terrorism in an effort to make international attitudes comport with their private interests: they might malign their adversaries as “terrorists”, insulate their allies from a similar charge, align their rhetoric with that of their allies while maintaining private reservations, or spoil efforts to produce an international definition of terrorism — which may place limits on state autonomy over the exercise of violence or production of counterterrorism policies — by feigning semantic objections. It should, however, be stressed that each of these three explanations suggest different data-generating processes. Thus, they offer substantively distinct theoretical expectations of the patterns of speech one would expect to observe being used by states conditional on their attributes. Which contention is correct is a consequential matter for International Relations theory, more broadly. Whether speech is all epiphenomenal “cheap talk”, just a reflection of domestic politics, or a means to persuade international counterparts fundamentally a question of causality and bears on issues of inference. Adjudicating between these various explanations requires, first, a robust approach to measuring how states use the term “terrorism” which, in turn, can produce empirical evidence against which theory can be tested. For this, recent methodological advances in text analysis provide a solution. The use of quantitative text analysis, or “text-as-data” techniques, by social scientists has grown significantly with the attendant increases in computing power, access to large corpora of political and social significance, and the development of specialised programming packages for processing and analysing text. Text analysis provides a potent means for the description and measurement of corpora which are too large for manual coding (Grimmer and Stewart 2013). For example, scholars have derived the policy positions of political parties from their manifestos (Laver et al. 2003) and those of states from their UN General Debate speeches (Baturo et al. 2017). The utility of these techniques also extends beyond the categorisation of texts into coarse topics or categories: text analysis methods can be sufficiently sensitive and flexible to aid in causal identification. In these cases, text has been treated as a confounder to be controlled for (Roberts et al. 2020), an outcome generated by some treatment in “text-as-outcome” studies, or as a causal determinant in its own right in “text-as-treatment” studies (Grimmer et al. 2022, 272–83). To this end, this paper first attempts to quantitatively measure how state representatives use the concept of terrorism in the annual United Nations General Debate, which has been a key international forum for debates over terrorism and its proper meaning. Specifically, this measurement is achieved through the use of *à la carte* (Khodak et al. 2018) feature embedding induction. This text analysis method allows each observed use of the word “terrorism” to be assigned a unique low-dimensional numeric vector — also known as a word embedding — which approximates its semantic meaning from the words in its immediate context. Word embeddings have desirable properties: words with similar meanings are closer together in the geometric space they encode. Further, this geometric space captures important socio-cultural axes that can be operationalised as interpretable semantic directions (see Kozlowski et al. 2019; Stoltz et al. 2025, for a technical review). Then, this paper treats these derived word embeddings as dependent variables and employs the *à la Carte on Text* ($conText$) embedding regression inferential modelling framework (Rodriguez et al. 2023) to assess the distance — and the statistical significance thereof — between states’ uses of the word “terrorism” with respect to speaker- and country-level covariates. Finally, this paper quantitatively locates states’ uses of the word “terrorism” on meaningful, and more readily interpretable, semantic dimensions to better substantiate its results. This paper finds that state-level attributes explain robust, systematic differences between countries in how they speak about terrorism once common time trends are controlled. Effects for politico-geographic regions are pronounced, and long-run expected government ideology corresponds to persistent differences in rhetoric. In contrast, within-country changes in state attributes explain comparatively little of the temporal movement in a state’s language. Baseline within estimators detect only limited change; focusing attention on theoretically motivated semantic dimensions reveals modest but informative within-country associations. These patterns indicate that while international rhetoric on terrorism is structured and not merely epiphenomenal, it exhibits relative inertia and lacks acute responsiveness to changes in domestic attributes. Methodologically, I extend $conText$ by integrating multiple imputation, removing fixed effects in embedding space to enable tractable within estimators, and projecting predicted embeddings onto semantic directions in order to test covariates against specific aspects of meaning with jackknife estimation and residual-permutation inference. ## Literature Review ### Terrorism Rhetoric: International Politics The United Nations has been the site of contestation between state representatives over what terrorism is and how it should be formally defined in international law (Saul 2006; Brulin 2013; Herschinger 2013). However, debate over the meaning of terrorism extends beyond its bounds. Following Schmid (1992), disagreement over *who* can be considered a terrorist and *what* acts or classes of acts can rightly be considered acts of terror abound in at three other “discursive arenas”: a public debate on terrorism which (in “open societies”) “is largely co-extensive” with the positions expressed by those in the media, an actor-centred arena populated by those who commit acts of political violence, and an academic arena populated by scholars working on issues of terrorism and political violence. Such a view highlights the many possible sources of variation in how states speak about terrorism. The prevailing literature offers accounts which, to varying degrees, consider these in relation to the state-level production of terrorism rhetoric. Scholars within critical terrorism studies (CTS) suggest that the term is used by the powerful to delegitimise acts of violence and actors that oppose their interests and, accordingly, offers more political than analytical utility (Jackson 2011). Terrorism is one strategy of political violence among the many available to state and non-state actors; it is neither a “single causally coherent phenomenon” (Tilly 2004, 12) nor a “fixed or essential identity” (Jackson 2007, 5). Yet, through fundamentally contingent social processes, the signifier “terrorism” has been assigned meanings that serve “to entrench certain hierarchies of power and exclude alter-native, counter-hegemonic forms of knowledge and practice” (Jackson 2007, 3). CTS scholars reject such hierarchies and definitions of terrorism which necessarily accord legitimacy to some actors and not others (Jackson 2007, 4). Although characterising this canon of work succinctly risks oversimplification, some of its distinct claims can be noted. Terrorism is political, but terror-afflicted states benefit from and employ constructions of terrorism as pathological, senseless, and disconnected from any actual political struggle (Stohl 2008). On the other hand, states actively exempt their own terroristic violence, i.e., state terrorism (Jackson 2005, 23), and that of actors which are perceived to advance their national interests (Stohl 2008) from the label of “terrorism”. Western democracies, in particular, use “terrorism” to stablise their national identities against non-Western others; the West is good, civilized, and peace-loving, whereas the non-West is none of these (Jackson 2005, 50). Moreover, Western states may rhetorically invoke terrorism primarily to legitimate their wealth and power advantages over the Global South (Herring 2008). States which rely on public support for their policies invoke terrorism to justify political projects that appear disproportionate in relation to the empirical incidence of acts of terror (Breen Smyth et al. 2008) and to acquire new powers (Zenn 2023). Conventional terrorism studies scholars more readily accept the qualitative distinction between non-state terrorism and other forms of political violence: such a distinction is not arbitrary or fully discursive and terrorist tactics are uniquely illegitimate (Hoffman 2006). However, they note reasons for why rhetorical constructions of terrorism vary across states. Principally, “terrorism” is a pejorative term. Those who employ terrorist tactics would rather be labelled “freedom fighters” or “soldiers” (Hoffman 2006). To this end, state-sponsors of terrorism and states which formed as a result of violent struggles for national liberation deliberately mislead others to believe that they are decidedly not terrorists (Ganor 2002). Additionally, they accept that the popular understandings of terrorism have shifted over time (Jenkins 1980), that identifying acts of terrorism is a partly subjective exercise (Jenkins 1980), and that the features of the acts which comprise the category of “terrorism” are heterogeneous (Schmid 2004). These accounts, however insightful, are not sufficiently comprehensive or granular to understand the factors that give rise to state rhetoric on terrorism in international fora. Terrorism is undoubtedly a powerful label to wield against one’s enemies. This recognition does not shed light on how a particular state will articulate the issue of terrorism given its security challenges, national history, institutions, or internal politics. To produce a more complete account of the state-level determinants of terrorism rhetoric, this paper first identifies a non-exhaustive set of salient semantic dimensions along which the meaning of terrorism has varied. Then, from existing theory, considers how state attributes might map rhetoric onto them. Early debates over the definition of terrorism in the 1970’s were paralyzed by two key issues of contention. Firstly, there was the question of whether states could commit terrorism, especially when a state maintained a colonial regime through repressive measures ((Peterson 2004, 179); (Acharya 2009, 654)). Along this dimension, states aligned with the USSR throughout the Cold War and leftist governments in the Global South were the most vocal proponents of considering acts of exceptional state violence, repression, and domination as instances of “state terrorism” (Acharya 2009). Secondly, there was the question as to whether the political violence of “freedom fighters” and anticolonial militants could be exempted from the definition of terrorism. Calls for such an exemption relied on the perceived legitimacy of these actors’ struggles against unjust forms of domination and the asymmetry in capabilities between foreign oppressors and the oppressed ((Jenkins and Johnson 1975, 2); (Saul 2006, 71); (Acharya 2009, 678)). These issues are reflective of an underlying semantic dimension with two opposed poles. At one extreme, terrorism may be the product of underlying “root-causes”, grievances, and motivations. These sources of discontent may confer terrorism legitimacy. Or, their resolution might be required to bring an end to terrorism. At the other extreme, terrorism has no justification; it must be categorically condemned, and any overtures to terrorists’ purported motivations is inappropriate. Another semantic dimension — legal versus repressive responses to terrorism — presents itself in relation to the problem of how states should respond to terrorism. Member states of the UN have questioned whether terrorism can be legally regulated and curtailed in the fashion of international humanitarian law or domestic policing (Peterson 2004). For some, terrorism is distinct from criminality, and the only appropriate recourse for states is violent suppression or outright war (Edel and Josua 2018). Which state-level features, then, explain difference and similarity in how state representatives use the term “terrorism”? And, more substantively, how do these features map onto the salient semantic dimensions pertaining to terrorism’s purported causes and policy responses it demands? In the subsequent sections, the extant literature is surveyed for theory-driven explanations. ### Domestic Politics, Public Opinion, and Threats An important explanation for contestation over the definition of terrorism and the outcomes of counterterrorism policymaking efforts emphasizes domestic politics, public opinion, and the nature of internal terrorism threats. This position is explicated here with reference to Foreign Policy Analysis (FPA) and Selectorate Theory. ### Two-Level Games While most theories of IR treat states as their primary units-of-analysis, the FPA literature on “two-level games” focuses on the interplay between international negotiations, state leaders, and domestic politics (Conceição-Heldt and Mello 2017). Built on the work of Robert Putnam (Putnam 1988), the metaphor of two-level games observes that, in international negotiations, state representatives seek to reach an agreement in their win-set, or one which would be supported by the necessary majority of their domestic constituents. The determinants of win-set size are many, but some particularly relevant determinants can be identified: Win-sets expand when the failure to secure an agreement is more costly to constituents, an issue is less politicised, decision-makers are relatively more autonomous from domestic pressures, and when a negotiator has more secure domestic political support (Putnam 1988). Larger win-sets make international agreements more likely given that a negotiator can readily accept a greater range of agreements. But this flexibility also comes at a cost: one’s relative bargaining position is weakened by possessing a large win-set because international counterparts can seek greater concessions (Putnam 1988). Formal negotiations over an international definition of terrorism resemble a two-level game. Diplomats at the UN who are tasked with agreeing upon an international definition of terrorism act as agents of a principal, their national government and its constituents. The apparent failure to agree upon a single definition of terrorism at the international level, then, may actually result from the irreconcilability of the optimal set of international outcomes and the demands placed on state leaders by intragovernmental policy coalitions and domestic populations. Critically, less formal negotiation over the meaning of terrorism takes place even when no immediate agreement is sought. If diplomats seek to negotiate an international consensus on the nature of terrorism and the policy response it demands, then the relative size of win-sets may explain why this has not been achieved in practice. A consensus may be possible when win-sets are large: domestic constituencies would accept many different rhetorical constructions of terrorism so consensus is more readily found. Conversely, small win-sets make reaching consensus more challenging. Larger dissimilarity between actor’s rhetoric, then, is expected along covariates which modify win-set size. Covariates which are orthogonal to win-set size would not be associated with meaningful differences in actor’s rhetoric. An empirical example illustrate this tension between international and domestic negotiations. Kenya’s purported initial status as a “reluctant partner” in the twenty-first century international counter-terrorism regime (Whitaker 2010, 2008) and the increase in its resolve to engage in counterterrorism efforts since 2011 (Whitaker 2014) have been partly attributed to domestic factors. Domestic constituencies in Kenya had initially resisted international counterterrorism cooperation and the passage of domestic counterterrorism legislation on various grounds. Constituents feared that increased government surveillance and policing powers would enable repression (Whitaker 2008, 260). Terrorism was perceived as a minor threat to Kenyans and was largely seen as an American or Israeli problem (Whitaker 2014, 93). Further, there existed a broad reluctance to acquiesce to pressure from the US or the United Nations Security Council Counter-Terrorism Committee (CTC) when the newfound freedoms of its transitional democratic institutions afforded Kenyans an opportunity to assert their key interests (Whitaker 2010, 2008). These factors are suggestive of a smaller win-set; domestic constituents would not accept strong rhetorical support by the government for a Global War on Terror. As a result, a pattern of “private cooperation and public contention” (Whitaker 2008, 268) characterised Kenya’s engagement with the United States and the international counterterrorism regime. In 2011, domestic and international contexts began to shift: Kenya invaded Somalia to fight the al-Qaeda-affiliated group al-Shabaab, reprisals by al-Shabaab heightened the public’s threat-perceptions of terrorism, and engaging in salient counterterrorism efforts became electorally and politically advantageous for Kenya’s then-divided government (Whitaker 2014). These factors are suggestive of a larger win-set. This granted the Kenyan government greater latitude in, firstly, how they responded to terrorism and, secondly, the extent to which they could rhetorically align themselves with the Global War on Terror. Vulnerability to terrorism may expand a state’s win-set by increasing the costs of non-agreement. Terrorism presents a number of collective actions problems (Sandler 2005). Thwarting terror attacks often requires joint international efforts on problems including intelligence sharing, extradition, the freezing of assets, and the disruption of financial flows (Sandler 2005). States which are vulnerable to terrorism suffer greatly when coordination on these issues fails, whereas less vulnerable states do not. To the extent that coordination first requires common ground over what constitutes terrorism and how it should be targeted, the rhetoric of vulnerable states may converge. Less vulnerable states may have more latitude to take idiosyncratic positions on terrorism and to bargain for rhetorical concessions that uniquely benefit them. **H1a:** Differences in vulnerability to terror attacks are associated with dissimilarity in state uses of “terrorism”. Authoritarian leaders and hereditary monarchs are more autonomous from domestic pressures than democratic leaders and have larger win-sets. Accordingly, their representatives may more readily adopt terrorism rhetoric that is consistent with international counterparts. **H1b:** Differences in representative autonomy from domestic pressures are associated with dissimilarity in state uses of “terrorism”. ### Selectorate Theory Selectorate theory is premised on a similar focus on both international and domestic levels of politics (Bueno de Mesquita et al. 2003). However, it seeks to explain the political strategies that leaders employ to remain in power. Democratic and authoritarian leaders alike rely on the support of some subset of the population, or their “winning coalition”, to remain in power (Bueno de Mesquita et al. 2003, 7–8). For the former, the winning coalition is “the group of voters who elect the leader” while, for the latter, the winning coalition chiefly consists of a small group of military and political elites (Bueno de Mesquita et al. 2003). Leaders will avoid articulating “terrorism” in ways that alienate the members of their winning coalition or jeopardise their continued political survival. Higher prevalence of terror attacks may threaten a leader’s political survival by revealing foreign policy deficiencies, triggering the misallocation of government resources, and harming the society’s well-being (Park and Bali 2017). Robust material and rhetorical commitment to counterterrorism, whether through outright repression or legal means, may represent a strategy by terror-affected leaders to offset these threats to their continued rule. Indeed, the domestic threat of terrorism increases support for the implementation of domestic and international measures to combat terrorism (Huddy et al. 2005). Moreover, states ratify international counterterrorism treaties, firstly, when they experience a high frequency of domestic terrorist attacks and, secondly, in order to signal their commitment to fighting terrorism to both international and domestic audiences (Milton et al. 2021). However, material and rhetorical commitments to counterterrorism efforts will not necessarily remedy the challenges terrorism poses to a leader’s political survival. Instead, they may increase the threat that terror attacks occur domestically and *compound* the threat to a leader’s political survival. Coercive measures taken by states against domestic insurgent groups are “robustly associated with later rebel use of terrorism” against civilian targets (Asal et al. 2019). Policy change which occurs in response to international pressure may provoke violent responses from domestic terrorist organisations which benefited from existing laws (Khan 2005). Repressive counterterrorism measures may place an undue economic and social burden on aggrieved populations, again compromising support for the government (Bueno De Mesquita and Dickson 2007; Park and Bali 2017). Finally, states which either sponsor terrorist groups (see Hoffman 2006, Ch.9) or those in which terrorist groups enjoy public support risk backlash from penalizing these actors. States with low state capacity, *ceteris paribus*, are more prone to terrorism (Piazza 2008) and may suffer more greatly from such reprisals. Yet limits in a state’s administrative, surveillance, and monitoring capacities render preventive measures less effective and selective means of addressing grievances less feasible (Blankenship 2018). For these states, domestic repression may be both a particularly inefficient, costly strategy and still the only realisable approach to counterterrorism (Blankenship 2018). Any significant redistribution of already scarce resources to counterterrorism bureaucracies or selective outreach may alienate a leader’s winning coalition and would be avoided. In constrast, high capacity states may more readily endorse legal, bureaucratic, and uncompromising modes of counterterrorism; these approaches are effective (Hendrix and Young 2014), the costs and risks can be shouldered, and one’s winning coalition is insulated from the effects of repression. Authoritarian states face a similar dilemma. Authoritarians rely on a combination of public repression and private rewards to stay in power. Conventional terrorism, or acts of violence perpetrated by radical groups which are relatively weak and few in number, disrupts the distribution of private rewards to pro-regime elites and further threatens the political survival of its leadership by communicating leader weakness. Accordingly, authoritarian leaders might rationally seek to counter terrorists with repression to ensure regime survival. However, repression is a costly and risky strategy for authoritarian states (Haber, cited in Aksoy et al. 2015). Severe repression risks mobilising domestic moderates against the regime and drawing international condemnation (ibid.; (Abrahms 2007)). Further, while conventional terrorism may be readily managed with the selective targeting of a few radical actors, terrorism perpetrated by the sustained insurgent movements that are more common in authoritarian states can require more severe and widespread repression (Abrahms 2007; Aksoy et al. 2015). Countering sustained insurgencies also often requires the mobilisation of military forces. Empowered military forces in authoritarian states may possess a greater capacity to attempt a coup if discontented (Aksoy et al. 2015). Repression and military mobilisation in authoritarian states may invite, rather than resolve, challenges to regime survival. Authoritarian states may recognise the threat repression entails and seek to co-opt opposition before they can engage in terror. Strictly autocratic authoritarians may have little scope to politically accommodate or pacify opposition, whereas hybrid regimes may be able to do so within nominally democratic legislatures (Aksoy et al. 2012). Partisan legislatures have existed within most states with non-elected executives at least since the late nineteenth century (Bonvecchi and Simison 2017) and it has been argued that autocrats use these legislatures precisely to nullify threats which are external to their ruling elite (Gandhi and Przeworski 2007). Strong rhetorical commitments to repression may undermine authoritarian attempts at terrorist co-optation and draw condemnation in spite of repression’s primacy as an authoritarian tactic of political survival. Full democracies may be relatively more insulated from backlash effects and reprisals to counterterrorism measures. Democratic institutions afford individuals peaceful means of expressing grievances, so terrorism is viewed not only as completely illegitimate, irrespective of one’s cause, but also as an affront to liberal ideals. Thus, motivation to counter terrorism will remain high among the winning coalition in the face of significant public costs and compromises (Abrahms 2007). Further, it has argued that public costs are likely to be curtailed by democracies’ commitment to civil liberties; democracies will be less likely than authoritarian states to overreact to terrorism or resort to widespread, unbridled repression (Abrahms 2007). However, democracies face immense “political and emotional pressure to retalitate” against terrorists from their publics (Adler 2010) and have been willing to suspend civil liberties to do so (Magen 2018). Democracies, then, may feel emboldened by their relative security and threaten repression rather than double down on liberal restraint. To summarise, these constraints on states’ responses to terrorism may condition the content of their rhetoric in the following ways. High state capacity is associated with a greater willingness to endorse legal-bureaucratic and uncompromising solutions to terrorism, as opposed to more repressive and accommodating solutions. Democracies will use more repressive, condemnatory rhetoric given public demands to do so and the more assured continuation of support from their winning coalition. Authoritarian states will refrain from such rhetoric to facilitate the co-optation of adversaries and to avoid condemnation for tactics of repression. Finally, a greater incidence of terror attacks is associated with a stronger condemnation of terrorism. **H1c:** Greater state capacity is associated with semantic proximity to “legal solution” and “senseless and unjustified”. **H1d:** Greater levels of democracy are associated with semantic proximity to “repressive solution” and “senseless and unjustified”. **H1e:** A higher incidence of terrorism is associated with semantic proximity to “senseless and unjustified”. The ideology of political parties may mediate terrorism’s relationship to political survival, and consequently condition a representative’s rhetoric. Enquires into whether the increased incidence of terrorism increases support for right-wing (Berrebi and Klor 2008; Getmansky and Zeitzoff 2014) or left-wing parties (Bali 2007; Gould and Klor 2010) have produced mixed results (for null effect see Balcells and Torrats-Espinosa 2018). However, the literature does support the notion that terrorism rhetoric will vary by political ideology. Issues of personal security, law and order, and threats of violence tend to be salient for those holding right-wing attitudes (Osborne et al. 2023). Right-wing authoritarian attitudes, which entail greater acceptance of illiberal responses to threats and the derogation of norm-violating groups, might also preclude the consideration of the purported motivations of terrorists and produce support for punitive countermeasures to terrorism (Cohrs et al. 2005). State representatives from right-wing parties, then, may condemn terrorism in unequivocal terms, endorse harsh responses to terror, and dismiss any purported motivations for or “causes” of terrorist violence to appeal to their winning coalition, among which supportive attitudes are expected to be prevalent. In contrast, representatives from left-wing parties may use rhetoric which draws attention to the fundamental causes, grievances, and motivations that lead to terror and the formal policy responses which may alleviate these (Krebs and Lobasz 2007). However, differences by ideology may be attenuated if left-wing parties seek to challenge right-wing “ownership” over the issue of terrorism (see Aytaç and Çarkoğlu 2021) and speak of it in similarly stark terms. **H1f:** Rhetoric of representatives from right-wing political parties will be semantically proximate to “repressive solution” and “senseless and unjustified”. **H1g:** Rhetoric of representatives from left-wing political parties will be semantically proximate to “legal solution” and “root-causes and motivations”. ### Cheap Talk and Costless Signalling Realists maintain that the rhetoric of state representatives is “cheap talk” (Mearsheimer 2010): talk does not incur costs, inescapably commit its speaker to take some action, fundamentally reconfigure state interests or identities, or alter the payoffs of bargaining outcomes. The costless nature of such communication renders rhetoric a poor signal of state preferences. States recognise this, and will not be persuaded to change their behaviour in response to costless signals. Accordingly, rhetoric is disregarded by realists as uninformative and epiphenomenal on these grounds. However, works on “cheap talk” in the IR bargaining theory literature and economics suggest an outright dismissal is misguided: cheap talk, while costless, can be made credible and reveal states’ true preferences if incentives and the nature of state interaction permit. In this section, talk’s “cheapness” is assumed and the implications thereof are considered in relation to debates over the meaning of terrorism. In typical bargaining situations, two states disagree over some issue and seek, through private diplomacy, to reach a settlement which is mutually-preferable to a costly and risky war. War is used to resolve the disagreement when no settlement can be reached. Under these conditions, states face incentives to misrepresent their willingness to fight, i.e., their resolve, as being higher than it is in order to extract greater concessions from war-averse opponents. Such misrepresentation renders cheap talk uninformative with respect to actual state preferences and might preclude states from reaching an agreement which both parties prefer to war (Ramsay 2004). However, international crises — or *public* bilateral disputes which both state parties have suggested might be resolved militarily — may occur precisely because they provide the means to make cheap talk credible and informative (Fearon 1994). They do so by placing exogenous constraints on the incentives to misrepresent: audience and reputation costs. Leaders may verbally escalate an international crisis by expressing their willingness to fight in the hopes that their opponent will back down before resorting to war. However, if these leaders were to eventually back down and renege on their otherwise costless rhetorical commitments, they may be electorally penalised by discontented domestic audiences or suffer a tarnished reputation in the eyes of international allies and adversaries (Fearon 1994; Ramsay 2004). These audience and reputation costs, respectively, are held to linearly increase as a crisis continues to escalate. As costs mount, states revise their estimate of their opponent’s resolve upwards because low resolve states are increasingly expected to have conceded. Otherwise, the mounting costs of conceding would eventually exceed those of war, making war inevitable. Thus, states which can generate audience and reputational costs through verbal commitments (most notably, democracies) are able to credibly communicate their true resolve to an opponent by continuing to escalate in the face of mounting costs (Ramsay 2004). General Debate speeches, too, are public statements which are observed by international and domestic audiences. International counterparts are addressed directly. Media delegations frequently accompany world leaders to the debate to report on the speech for domestic audiences (Kentikelenis and Voeten 2021, 726), and recorded speeches can be accessed on the United Nations website. Member states reserve a “sovereign right... to solely and entirely determine the content of their general debate statements” (United Nations General Assembly 2003) and, unlike the high level of coordination observed in state voting behaviour in the UN (Iida 1988; Laatikainen 2017; Chelotti et al. 2022), states generally do not coordinate on the content of their General Debate speeches (Chelotti et al. 2022). Audience and reputational costs, then, can theoretically be imposed on leaders that back down from (or fail to deliver on) verbal commitments made during a General Debate. However, whether or not credible and informative cheap talk is expected in the General Debate depends on the preferences states hold and the ability of observers to accurately monitor the conformity of states’ behaviour to their rhetorical commitments. Consider that states which self-select into international crises prefer a non-military agreement to a costly war but are uncertain of their opponent’s willingness to fight. They might instrumentalise audience and reputation costs to resolve this indeterminacy but, having self-selected into a crisis, are guaranteed to face one of these two potential outcomes. Granting concessions represents a signal that a leader has not delivered on their commitments. War, or winning concessions, signals that they have. Rhetorical uses of terrorism in the UN might resemble such a scenario if demands are placed on adversaries and backed by the threat of force. The United States’ call for Iraq to “immediately end all support for terrorism and act to suppress it” or face “action” is one such case (Bush 2002). However, more typical scenarios are not characterised by similarly deterministic potential outcomes, unambiguous fulfillment of commitments, nor such clear preferences. For one, to the extent that terrorism has been constructed as the world’s “absolute enemy”, states may seek to be perceived as part of an international collective of states which oppose it (Herschinger 2013) irrespective of their true commitment to counterterrorism. To this end, states may undertake costly signalling. For instance, states may quit behaviours which undermine this perception, such as supporting terrorist organisations financially, politically, or materially. Or, they may engage in behaviours which support it, such as material investment or active cooperation in international counterterrorism efforts. While the cost of such actions is precisely what affords them credibility, states may nevertheless be unwilling to suffer such costs to develop such a reputation. In contrast, cheap talk in such a context enables states to lie or misrepresent their preferences with relative impunity. Unlike international crises, in which states use cheap talk instrumentally to extract concessions and avoid war, cheap talk in this context is of less material consequence. States which fail to convince others of their opposition to terrorism on account of their behaviour may risk sanctions or lose access to foreign aid. But this is likely independent of, or even mitigated by, cheap talk which expresses support for international counterterrorism efforts. Additionally, while backing down in a public international crisis is a unambiguous departure from a leader’s rhetorical commitments to fight, it is less clear what would constitute a similar offense in this context. Thus, audience costs and reputational costs may be less applicable. Accordingly, one would expect that states which endorse or enact illegitimate forms of violence would unequivocally condemn terrorism, much as would states which do not. Thus, one would expect the former class of states to use “terrorism” in essentially equivalent ways to the latter class of states. **H2a:** States’ uses of the term “terrorism” are not semantically dissimilar along covariates and are semantically proximate to “senseless and unjustified”. However, the contention that states seek to be perceived as part of an international collective of states which oppose the “absolute enemy” of terrorism is more analytically sustainable for periods after the end of the Cold War and the inception of the Global War on Terrorism. These later periods are ostensibly when the “absolute enemy” frame was most firmly established (Herschinger 2013). An alternative “cheap talk” account which is potentially more tenable across time might stress the symbolic, *theatrical* divisions in international politics. States may simply utilise rhetoric as if they were actors from one of “the two major troupes” — North versus South or East versus West — in the “global theatre” of UN politics, symbolically distinguishing its camp from the other even when “actual country positions have virtually nothing to do with these categories” (Weiss 2009). The recognition that terrorism is imprecisely defined, relates directly to the legitimacy of violence in international politics, and that talk is “cheap” may embolden states to use “terrorism” to act out their role in this global drama. Accusations that an opposing bloc supports or perpetrates acts of terror, for instance, serve this oppositional politics while incurring no costs on the sender and surviving objections because no international definition exists to definitively rule out such a charge. The widely noted debates over whether terrorism can be committed by states or, separately, whether freedom fighters are, by definition, not terrorists on account of the legitimacy of their struggle, may be residual artifacts of this theatrical form of politics and not the consequence of true differences in countries’ positions. In other words, such a cheap talk approach would argue that symbolic geopolitical divisions are causally prior to semantic contestation over terrorism; domestic politics and strategic communication approaches may anticipate similar patterns of semantic contestation to emerge. However, they would maintain that state attributes and relations are causally prior to both a state’s selection into an international bloc and the dimensions of semantic contestation which become salient. This theatrical account of “cheap talk” would expect insignificant or only marginal differences between states’ rhetoric on “terrorism” after controlling for each state’s politico-geographic region. **H2b:** States’ uses of the term “terrorism” are not semantically dissimilar after controlling for their politico-geographic region. ### Strategic Social Construction Speech may also be persuasive. This is true in a narrow sense: speech may be the medium to issue threats or to make commitments. But, in these cases, persuasion is not a property of language one uses but the *speech act* that language performs (Waever 1995). A state might avoid attacking an adversary if issued with a threat of reprisals by third party states. This does not imply that the semantic content of those threats sparked self-reflection or attitudinal change on the part of the target state. These threats commit the issuing states to some course of action, and it is that action, not the language it is issued with, which might convince the target state to refrain from attacking. Conventional constructivist theorists — a designation which is meant to exclude “critical” (Hopf 1998) or “poststructuralist” (Onuf 2012) constructivists, though only because their works are not considered in this paper — offer accounts of speech which take its persuasive capacity further. For these theorists, subjective beliefs and “objective” national interests (Wendt 1999, 231–43) jointly motivate state behaviour. Diplomatic communication is unlikely to convince a state to seriously redefine its objective interests — as these are behavioural imperatives placed on states by the hard realities of the international system and the nature of corporate statehood — but it may shift state attitudes and beliefs in more limited, yet nevertheless consequential, ways. In the words of Alastair Iain Johnston (2001), “the goal of diplomacy is often the socialization of others to accept in an axiomatic way novel understandings about world politics.” This section considers the implications of treating talk as persuasive with reference to Finnemore and Sikkink’s (1998) notion of “strategic social construction”. They define strategic social construction as processes “in which actors strategise rationally to reconfigure preferences, identities, or social context” (Finnemore and Sikkink 1998, 888). Their invoking of rationality and, thus, rational choice theories, is purposeful. Finnemore and Sikkink argue that ideational factors and rational choice logics operate simultaneously and condition one another. Thus, strategic social construction also involves actors “making detailed means-ends calculations to maximise their utilities, but the utilities they want to maximize involve changing the other players’ utility function in ways that reflect \[their\] normative commitments” (Finnemore and Sikkink 1998, 910). The actors in question are, for Finnemore and Sikkink, primarily “norm entrepreneurs” who seek to convince states to accept new norms. Norm entrepreneurs which successfully do so have maximised their utility. In contrast, the actors of interest for this paper are state representatives. The object of their rhetorical efforts is not the adoption of discrete norms. Instead, this paper assumes that these actors seek to transform the “common knowledge” and “intersubjective understandings” about terrorism which pervade international politics. Their utilities are held to be both ideational *and* material; they seek to garner support for constructions of terrorism which lead to a permissive international context for the pursuit of their material preferences and one which regards them as legitimate. Of course, if one accepts that strategic social construction occurs, then a search for strictly “state-level determinants” of speech becomes problematic. A state’s observed rhetoric may be determined by their state-level interests, attributes, and policies. Yet their rhetoric will be determined by international factors if another state has successfully modified their utility function through strategic social construction. Nevertheless, a strategic social construction account of state rhetoric offers observable implications that can be evaluated without having to fully resolve the issues of interdependence and influence. First, this paper accepts the contention that “normative claims become powerful and prevail by being persuasive” and that “being persuasive means grounding claims in existing norms in ways that emphasize normative congruence and coherence” (Finnemore 1996, 141). This provides an observable implication of strategic social construction, conditional on one further assumption: states whose interests are more directly impacted by the international ideational context will be more likely to engage in strategic social construction efforts. States with a greater incidence of terrorism, state sponsors of terrorism, state which engage in targeted politics killings of domestic opposition, and states which decolonised through violent means have more at stake in the contestation over the meaning of terror. The international ideational context which emerges from this contestation determines whether others regard their behaviour — such as their counterterrorism efforts or the acts of political violence they and their allies perpetrate — as legitimate. In expectation, these states will be more likely to engage in strategic social construction and, in order to be persuasive, will use more appeals to established international principles and norms in their rhetoric. **H3 (Responses):** States with higher terrorism incidence, domestic political killings, or violent decolonisation legacies will be more semantically proximate to the “law-enforcement and co-operation” pole (and more distant from the “repressive and militaristic” pole) on the Responses dimension. Consider that “terrorism” is a term with a particularly negative connotation that is difficult to shed once acquired (Jenkins 1980). Those who are labelled “terrorists” risk losing international legitimacy and are likely to find their ability to pursue their strategic interests significantly impaired (Jenkins 1980). States may rationally attempt to frame the issue of terrorism in ways that best serve their interests by imposing such costs on their adversaries or by shielding themselves and their allies. For instance, Turkey and Russia successfully framed their domestic ethno-nationalist separatist movements — the Kurdistan Workers’ Party and Chechen separatists — as “terrorists” by echoing the normative appeals of the post-9/11 War on Terror (Pokalova 2010): sovereign states reserve the exclusive right to lawful belligerency (Ralph 2007) and significant counterterrorism operations are legitimated on the grounds of self-defense (Dinstein and Domb 2003). These frames resonated, and these groups were included on several lists of international terrorist groups (Pokalova 2010). While Serbian attempts to cast separatist, ethnic Albanians in Kosovo as terrorists fell apart after severe repression by Serbian forces in their 1998 Drenica offensive drew international condemnation (Pokalova 2010), their appeals were also replete with appeals to international principles and norms. Letters by the Government of the Republic of Serbia to the UN Secretary-General and the President of the General Assembly (2003) expressed the government’s readiness to protect “the rule of law, full equality and life in peace and security of all citizens, regardless of their nationality”. They affirmed their commitment to means of resolution “in accordance with... European and international standards, OSCE principles, Paris Charter and the Charter of the United Nations” (Government of the Republic of Serbia 2003). Equally, states may preempt charges that their forces or allies are “terrorists” and seek to insulate them from such a label and its consequences. For instance, Boaz Ganor (2002) notes that states which sponsor terrorism — a designation in which he includes Syria, Libya, and Iran — have lobbied the international community to exempt “freedom fighters”, citing their internationally-recognised right to self-determination, from an international definition of terrorism. In doing so, he argues, these states believe that any resulting definition will “absolve them from all responsibility for supporting terrorism” (Ganor 2002). However, the issue of whether freedom fighters can be considered terrorists has been a concern of a much broader set of states than those noted by Ganor. And, critically, this includes states with no immediate involvement in wars of national liberation or insurgencies. The EU Council initially sought to exclude freedom fighters from the 2002 EU Framework Decision; their chief concern was to avoid retroactively criminalising the asymmetric violence of World War II resistance fighters who “acted in the interest of preserving or restoring... democratic values” (EU Council, cited in Saul 2006). A group of non-aligned nations that argued to exempt national liberation fighters from an international definition of terrorism in the 1970s included states that decolonised violently and those that did not (see Jenkins and Johnson 1975, 2). The alternative definition they suggested condemned violence and “repressive acts by colonial, racist and alien regimes against peoples struggling for their liberation” as a form of international terrorism which contravened those peoples’ “inalienable right to self-determination and independence” (Jenkins and Johnson 1975). This pattern is suggestive of strategic social construction; states seek to persuade others to adopt beliefs about terrorism which render it inapplicable to acts of political violence which are critical to their national histories and present-day perceptions of their legitimacy. Such constructions were not inevitable or determined. A number of early national liberation movements openly self-identified as “terrorists”, though this practice of self-identification largely disappeared once “the word terrorist became so repugnant” that engaging in terrorism compromised one’s international legitimacy (Rapoport 2017). ## Data and Methods The primary dataset that I use in this study is the *United Nations General Debate Corpus* (UNGDC) (Jankin Mikhaylov et al. 2017), expanded to include all identifiable English-language transcriptions of speeches delivered during the United Nations General Debates held between 1947 and 2022. The annual General Debate of the UN General Assembly offers a forum for representatives of UN member-states to appear before their peers and deliver a speech on the issues they deem relevant. These speeches need not address terrorism and regularly do not. However, the General Debate’s regularity, temporal breadth, inclusive membership, and standardised format of speech delivery makes it an appropriate source for detecting meaningful semantic differences in how states speak about terrorism. As a robustness check, I also estimate models on a subset from 1970 to 2022, reflecting the view that the sustained, modern form of UN attention to terrorism began during the 1970s (Zulaika and Douglass 1996; Stampnitzky 2011). Prior to my own processing of the data, the speeches are converted into text files and cleaned by the dataset authors. Translation is performed by United Nations staff when a speech is delivered in a language other than English. Where possible, the authors record the name and position of the state representative delivering a speech. While position data are sparser in early years, coverage is strong from the 1990s onward. This position data is used to infer the political party with which a speaker is affiliated, if any, and the features of that party. Accordingly, analyses that employ party- or speaker-level predictors leverage the widest feasible temporal range given party data availability. For this analysis, the speeches are tokenised, or separated into individual words. Stopwords are not removed; the retention of stopwords does not significantly affect the results of conText embedding regression (Rodriguez et al. 2023, 15). Next, uses of the word “terrorism” (sometimes with modifiers, see [Appendix](#sec:appendix)) are identified. For each use of the word “terrorism”, a new observation is created which consists of the ten tokens on either side of the focal word. These tokens represent the term’s context. Following the “distributional hypothesis” that “you shall know a word by the company it keeps” (Firth, cited in Rodriguez et al. 2023, 3) or, alternatively, “a feature is characterized by the words in its context” (Harris, cited in Khodak et al. 2018, 1), these words can be used to infer the focal word’s meaning. Each context word in this tokenised corpus is paired with their corresponding word embedding estimated on the full United Nations General Debate corpus (all speeches used to train the embedding model, not only terrorism windows). The primary analysis uses 100-dimensional locally trained embeddings. Two robustness checks are employed: (i) a 100-dimensional retrofitted Mittens model that compromises between local UN embeddings and pretrained GloVe vectors (Dingwall and Potts 2018; Pennington et al. 2014); and (ii) 100-dimensional fastText embeddings (Mikolov et al. 2018). This combination captures domain-specific usage while borrowing strength from large, general corpora. Such an approach is computationally expensive and produces a single embedding for each unique word. One typically needs to observe a word many times, across many different contexts, to produce a high-quality measure of its *average* meaning. To examine differences in how states use the word “terrorism”, one needs a unique embedding for each individual appearance of the word in the corpus. For this purpose, word embeddings for individual uses of the word “terrorism” are derived through *à la carte* embedding, or “a linear transformation of the average word embeddings *in the feature’s available contexts*” (Khodak et al. 2018, 1). In sum, the semantic information learned from the entire contents of the UNGD corpus is leveraged to inductively infer the meaning of each specific use of the word “terrorism” within it. The $conText$ modelling framework’s approach to measurement over text leverages the ability of à la carte embeddings — the theory and practice of which were most developed by Arora et al. (2015) and Khodak et al. (2018), respectively — to create high-quality embeddings of individual words simply from their immediate contexts. By treating these embeddings as variations in a focal word’s meaning or usage and casting them as dependent variables in a multivariate regression framework, one is able to, intuitively, “make statements about how embeddings (our Y) differ as covariates (our X) change” (Rodriguez et al. 2023, 4). My implemention of this method with the UNGD corpus is effective: The top-five nearest neighbours of the corpus-wide average of the à la carte embeddings are “terrorism”, “terrorist”, “crime”, “extremism”, “acts”. This suggests that the à la carte embedding has been performed successfully. That is, the process of deriving these embeddings is blind to the fact that the focal word is “terrorism” but, on average, it was nonetheless able to recover an embedding representing the concept purely from the language used around the focal word. Success, then, means that resulting vectors capture the semantic meaning of the terror-tokens inductively from their immediate contexts. Independent variables are produced at the country-year level. These include each country’s government ideology (left, centre, or right), level of democracy, real gross domestic product (GDP) per capita, whether a country had decolonised and whether this was through violent or non-violent means, whether a state uses tactics of torture or political killings domestically, a state’s politico-geographic region, and the logged number of terror attacks which had taken place domestically between the debate in question and the session which had preceded it. To retain maximal information, I handle missing covariate data via multiple imputation. Year and country fixed effects are also used. Government ideology categories (left, centre, right) are drawn from Herre (2023), which provides government ideology classifications over a wide temporal range. Real GDP per capita (in constant 2011 \$) estimates are drawn from *The Maddison Project* dataset (Bolt and Luiten Van Zanden 2020) as a proxy for state capacity (see Fearon and Laitin 2003). Variables for each state’s level of democracy and their tendency to use torture and political killings domestically are drawn from the *V-Dem* dataset (Coppedge et al. 2022; Pemstein et al. 2022). Data on whether each country had been colonised and, if so, the degree of violence which took place during decolonisation are drawn from the *ICOW Colonial History* dataset (Hensel 2018). Finally, the logged number of terror attacks in the year preceding each session are derived from the *Global Terrorism Database* (START 2022). I extend the published conText workflow in two key ways to accommodate longitudinal missingness and design-based constraints. First, I perform multiple imputation using the Amelia package in R, which is well suited to time-series cross-sectional data. I generate $m=5$ imputed datasets using country identifiers, year, politico-geographic region, and other observables to inform the imputation model. For each imputed dataset, I estimate a conText-style multivariate multiple regression. For a given covariate, I obtain its $K$-dimensional coefficient vector by concatenating the coefficients across the $K$ embedding dimensions (with $K=100$ in the main analyses), and I then average these coefficient vectors across the five imputations to produce a single point estimate. I compute uncertainty using leave-one-out jackknife standard errors within each imputed dataset and then average these standard errors across imputations. Because Rubin’s rules are not appropriate here — the original conText model assesses significance via empirical, residual-based permutation rather than parametric $t$-tests — I extend conText’s residual permutation logic to the multiply imputed setting: for each permutation index $i\in\{1,\ldots,1000\}$, I compute the permuted coefficient vector within each imputed dataset, average these $m$ permuted vectors to obtain a single permuted vector for index $i$, and compare the ratio of its squared Euclidean norm to the squared Euclidean norm of the primary averaged point estimate. As in conText, empirical $p$-values are derived from how often this ratio falls below one across the 1,000 permutations; ratios at or exceeding 1 undermine claims of significance. Second, I allow manipulation of the document embedding matrix prior to regression. Specifically, I implement linear projections that (a) partial out fixed effects directly in embedding space (year fixed effects for between-country analyses, and year plus country fixed effects for within-country analyses), yielding a within estimator for time-varying covariates; and (b) project document embeddings onto user-defined semantic subspaces before estimation. The latter permits inference on targeted, theoretically motivated dimensions of meaning while continuing to conduct significance tests in the original embedding space via the residual-permutation procedure. To identify the state-level determinants of states’ uses of “terrorism”, I employ the inference strategy of the framework “$conText$”: the embedding regression model (Rodriguez et al. 2023). The model returns $\beta$ coefficients that can predict how “terrorism” will be used, given the values of one’s independent variables. That is, the model predicts the expected word embedding representation of how an actor’s terrorism rhetoric in the UN given the values of one’s independent variables. This embedding possesses meaning through its location in the embedding space and distance from other word embeddings or concepts therein. First, I use this model to assess whether variation on a covariate produces any significant differences in how states’ use “terrorism”. This essentially assesses whether the expected embeddings at different levels of a covariate, *ceteris paribus*, differ significantly from each other. This is the conventional use of $conText$ embedding regression, and substantive interpretation of any observed differences across covariates is typically performed via a post-estimation analysis of a model’s predicted embeddings. Following recent guidance on distances in high-dimensional spaces, I use the debiased squared Euclidean norm of the model’s $\beta$ coefficients as the measure of *distance* between predicted embeddings across covariate values, rather than raw norms (Green et al. 2025). Uncertainty is quantified via jackknife; empirical $p$-values are obtained via residual-based permutation (1,000 repetitions), extended to multiple imputation by averaging permuted coefficient vectors across imputations. Pairing these measures of magnitude and statistical significance allows one to ask: “do state representatives significantly differ in how they use the word ‘terrorism’?” and “if so, by how much?” For discussion of bootstrap-related bias noted by the package authors that was not included in the original conText paper, see the update note.[^1] As alluded to above, the base $conText$ model does not immediately convey *how* a state’s rhetoric might change as a result of changes in a covariate. Post-estimation analyses convey some information by, for instance, showing whether the predicted embedding becomes semantically closer or further away from a particular concept of interest, via cosine similarity or other distance metrics. This is a useful starting point. However, it is somewhat opaque and renders hypothesis testing whether a covariate is associated with change on particular dimensions of meaning difficult. To address this, I adapt the approach described in Kozlowski et al. (2019) and reviewed technically by Stoltz et al. (2025) to construct three substantively interpretable semantic directions that correspond to some of the most contested issues in the international politics of terrorism at the United Nations (see [Appendix](#sec:appendix) for the concept word-pairs). First, a “root-causes and motivations” to “senseless and unjustified” direction (Causes) captures the contrast between states which recognise that terrorists are motivated by political grievances and seek to remedy them, and those which condemn them outright. Second, a “law-enforcement and cooperation” to “repressive and militaristic” direction (Responses) captures the contrast between states which call for unrestrained, violent strategies of counterterrorism, and those which suggest legal-bureaucratic solutions. Third, a “non-state perpetrators” to “state perpetrators” dimension (Perpetrators) is designed to capture the contrast between states which consider non-state actors such as insurgents and freedom fighters to be terrorists, and those which consider state actors such as repressive regimes to be terrorists. Given the nature of word embeddings and the limits of construction of dimensions in the embedding space, a more strictly technical interpretation of this dimension might be understood as capturing the contrast between language that relates to non-state actors perpetrating political violence versus language that is semantically closer to forms of state violence and repression. I modify the conText logic to permit pre-estimation modifications to its dependent variable: the document embedding matrix. I project each document embedding onto my dimensions, effectively parcelling out the components of their language that are unrelated to the dimension of interest while recovering the document’s direction, or position, on the dimension. I then assess whether changes in covariates are associated with corresponding movement along these dimensions and conduct hypothesis tests with the original $conText$ framework. Briefly, semantic directions point from one pole of a juxtaposed concept set toward the other. They are defined using opposing anchor terms that “pin” each side of the axis; averaging anchors on each side and taking their offset yields a direction vector that points, for example, towards poverty and away from wealth. Directions explicitly encode the contrast of interest and permit straightforward projection of word or document embeddings (e.g., via cosine with the direction) to recover positions along the axis (Stoltz et al. 2025, for an accessible review). This is directly analogous to spatial models familiar in political science: a semantic direction functions like a left–right policy axis, with two poles (left and right) and texts (or speakers) positioned along the line. Here, however, the poles are defined by opposed anchor terms drawn from theory (e.g., “root-causes” vs. “senseless”), and any relevant juxtaposition can be used to construct such a spatial scale. In this paper, I specify anchor pairs from domain theory and project document embeddings onto these directions prior to estimation; significance testing remains in the original embedding space through the residual-permutation procedure. For estimation, I run four sets of models that leverage these extensions: (1) a base conText regression with year fixed effects and no semantic projection; (2) three separate models—one per semantic dimension (Causes, Responses, Perpetrators)—with year fixed effects; (3) a base model with year and country fixed effects to identify within-country, time-varying associations; and (4) the dimension-specific models with both year and country fixed effects. The models with strictly year fixed effects permit the inclusion of time-invariant covariates that are of substantive interest; nonetheless, the research design cannot preclude the possibility that these are confounded by unobserved country-level characteristics. On the other hand, the models with country and year fixed effects effectively control for all time-invariant country-level characteristics and common shocks over time, rendering a causal interpretation of within-country associations relatively more robust. ## Results The base conText model with year fixed effects reveals pronounced between-country differences in how states use the term “terrorism”. Table 1 reports debiased squared-norm coefficients, which measure distances in the original embedding space; larger values indicate greater semantic differences across covariate values. Politico-geographic regions exhibit the largest magnitudes. Colonisation history, electoral democracy, terrorism incidence (logged), GDP per capita (logged), physical integrity, the operation of a legislature, and government ideology are all associated with statistically significant differences at conventional levels using residual-permutation $p$-values with jackknife uncertainty. These results indicate substantial cross-national structure in rhetoric once common time trends are controlled. Table 1. Deflated squared Euclidean norms with jackknife SEs and residual-permutation p-values (Base). Embedding: local; Corpus: Full corpus; FE: yrFE
Coefficient ∥β∥2 SE p-value
Region: Western Europe and North America (ref)
Region: The Pacific 0.244* * * (2.030) 0.000
Region: The Caribbean 0.180* * * (1.082) 0.000
Region: Eastern Asia 0.306* * * (2.411) 0.000
Region: Eastern Europe and post-Soviet Union 0.232* * * (0.055) 0.000
Region: Latin America 0.234* * * (0.052) 0.000
Region: North Africa and Middle East 0.090* * * (0.046) 0.000
Region: Southern Asia 0.091* * * (0.065) 0.000
Region: South-Eastern Asia 0.128* * * (0.102) 0.000
Region: Sub-Saharan Africa 0.068* * * (0.048) 0.000
Government Ideology: Center (ref)
Government Ideology: Left 0.021* * * (0.011) 0.000
Government Ideology: Right 0.015* * * (0.010) 0.000
Colonisation History: No colonisation (ref)
Colonisation History: Decolonized 0.017* * * (0.036) 0.000
Colonisation History: Violent Decolonisation 0.033* * * (0.034) 0.000
Terrorist Attacks (log) 0.083* * * (0.014) 0.000
Electoral Democracy Index 0.173* * * (0.052) 0.000
GDP per Capita (log) 0.025* * * (0.022) 0.000
Legislature in Operation 0.019* * * (0.014) 0.000
Physical Integrity Index 0.046* * * (0.029) 0.000
Model Information
Fixed Effects Year Year Year
Observations 20,109 20,109 20,109
Years Covered 1947-2022 1947-2022 1947-2022
*Note.* Euclidean norms measure absolute distances in embedding space and should be strictly positive. Negative values result from bias corrections applied to already small norms (after year/country FE removal) that shift them across zero; these do not indicate effect directionality. Significance stars reflect empirical p-values from residual permutation tests: $^{*}$ $p<0.05$, $^{**}$ $p<0.01$, $^{***}$ $p<0.001$. Standard errors from jackknife resampling. First, the existence of notable differences in states’ rhetoric which persist after controlling for year fixed effects and differences between geopolitical regions runs contrary to simple expectations that public rhetoric is uninformative “cheap talk”. This does not preclude talk from being costless to send, as the concept requires; it suggests, rather, that intuitions derived from the treatment of “cheap talk” in crisis-bargaining scenarios are ill-suited here. Had these been appropriate, one would expect to see few or minimal between-country differences in how states use the term “terrorism” after controlling for common time trends; this would parallel the pressures on *all* states, irrespective of their type, weak or strong, unresolved or resolved, to “talk tough” and present themselves as possessing high resolve and military capability when facing the looming threat of war. Results reported use multiple imputation, residual-permutation $p$-values adapted to imputation, jackknife standard errors averaged across imputations, and, where appropriate, embedding-level projections to remove fixed effects prior to estimation. Introducing year and country fixed effects to identify within-country, time-varying associations yields a different picture. In the baseline within-country model (), only the indicator for a legislature in operation is statistically significant. Other covariates do not reach significance in this specification, consistent with the view that short-run changes in attributes seldom produce wholesale changes in how a given state uses the term “terrorism”. Table 2. Deflated squared Euclidean norms with jackknife SEs and residual-permutation p-values (Base). Embedding: local; Corpus: Full corpus; FE: yrCntryFE
Coefficient ∥β∥2 SE p-value
Government Ideology: Center (ref)
Government Ideology: Left -0.000 (0.008) 0.999
Government Ideology: Right -0.007 (0.007) 1.000
Terrorist Attacks (log) 0.004 (0.004) 0.936
Electoral Democracy Index 0.011 (0.014) 1.000
GDP per Capita (log) 0.001 (0.004) 1.000
Legislature in Operation 0.008* * * (0.007) 0.000
Physical Integrity Index 0.001 (0.011) 1.000
Model Information
Fixed Effects Year+Country Year+Country Year+Country
Observations 20,109 20,109 20,109
Years Covered 1947-2022 1947-2022 1947-2022
*Note.* Euclidean norms measure absolute distances in embedding space and should be strictly positive. Negative values result from bias corrections applied to already small norms (after year/country FE removal) that shift them across zero; these do not indicate effect directionality. Significance stars reflect empirical p-values from residual permutation tests: $^{*}$ $p<0.05$, $^{**}$ $p<0.01$, $^{***}$ $p<0.001$. Standard errors from jackknife resampling. Government ideology (left, centre, right) (see Herre 2023) is used to characterise the prevailing government’s ideological orientation. In the between-country model with year fixed effects, left and right governments differ from centrist governments in how they use “terrorism”, with statistically significant and substantively non-trivial distances (Table 1). Given that country-level fixed effects are absent in these specifications, it is most appropriate to view these as long-run differences in terrorism rhetoric according to the average political orientation of a state’s government. In contrast, changes in government ideology within a country are not associated with significant shifts in language in the baseline within-country specification (Table 2). This pattern is consistent with durable ideological sorting across countries and with limited short-run responsiveness of a state’s rhetoric to government turnover. Table 3. Effect Sizes on Semantic Dimensions given $conText$ Model (Year FE’s) Predictions.
Causes Responses Perpetrators
Coefficient Effect Size Effect Size Effect Size
Region: Western Europe and North America (ref)
Region: The Pacific -0.327* * * 0.363* * * -0.540* * *
Region: The Caribbean 0.037 0.223** -0.034
Region: Eastern Asia -0.462* * * 0.011 -0.410* * *
Region: Eastern Europe and post-Soviet Union 0.042 -0.125* 0.173* * *
Region: Latin America -0.245* * * 0.397* * * -0.552* * *
Region: North Africa and Middle East -0.070 -0.061 -0.078
Region: Southern Asia 0.048 -0.057 -0.271* * *
Region: South-Eastern Asia 0.272* * * -0.155 -0.006
Region: Sub-Saharan Africa 0.053 -0.006 -0.096
Government Ideology: Center (ref)
Government Ideology: Left -0.107** -0.075* 0.055
Government Ideology: Right -0.156* * * -0.017 0.028
Colonisation History: No colonisation (ref)
Colonisation History: Decolonized 0.046 0.091** 0.016
Colonisation History: Violent Decolonisation 0.007 0.088 -0.047
Terrorist Attacks (log) 0.018 -0.179* * * 0.422* * *
Electoral Democracy Index 0.287* * * -0.032 -0.026
GDP per Capita (log) -0.038 0.116* * * 0.028
Legislature in Operation -0.080 0.065 0.046
Physical Integrity Index -0.118* -0.078 -0.117*
Model Information
Fixed Effects Year Year Year
Observations 20,109 20,109 20,109
Years Covered 1947-2022 1947-2022 1947-2022
Rhetoric Dimension Interpretation
Higher ↑ Root Causes Legalistic State
Lower ↓ Condemnation Repressive Non-State
*Note.* Rescaling of effect sizes akin to Cohen’s d. Effect sizes represent $\displaystyle d = \frac{s(\mathbf{\hat{y}}_{x_j + 2\sigma_j}) - s(\mathbf{\hat{y}}_{\mathrm{baseline}})}{\sigma_s}$, where $s(\cdot)$ denotes scalar projection onto the semantic dimension, $\mathbf{\hat{y}}_{x_j + 2\sigma_j}$ is the predicted embedding with covariate $j$ increased by 2SD, $\mathbf{\hat{y}}_{\mathrm{baseline}}$ is the prediction at intercept with covariates at means, and $\sigma_s$ is the SD of scalar projections from the original document embeddings. Significance stars reflect empirical p-values from residual permutation tests: $^{*}$ $p<0.05$, $^{**}$ $p<0.01$, $^{***}$ $p<0.001$. Table 4. Effect Sizes on Semantic Dimensions given $conText$ Model (Year+Country FE’s) Predictions.
Causes Responses Perpetrators
Coefficient Effect Size Effect Size Effect Size
Government Ideology: Center (ref)
Government Ideology: Left 0.011 -0.020 0.029
Government Ideology: Right 0.008 -0.012 0.029
Terrorist Attacks (log) 0.024 -0.046 0.074**
Electoral Democracy Index 0.132* * * -0.034 0.106**
GDP per Capita (log) 0.048 -0.048 0.053
Legislature in Operation -0.004 0.044 0.037
Physical Integrity Index -0.047 0.011 -0.083
Model Information
Fixed Effects Year+Country Year+Country Year+Country
Observations 20,109 20,109 20,109
Years Covered 1947-2022 1947-2022 1947-2022
Rhetoric Dimension Interpretation
Higher ↑ Root Causes Legalistic State
Lower ↓ Condemnation Repressive Non-State
*Note.* Rescaling of effect sizes akin to Cohen’s d. Effect sizes represent $\displaystyle d = \frac{s(\mathbf{\hat{y}}_{x_j + 2\sigma_j}) - s(\mathbf{\hat{y}}_{\mathrm{baseline}})}{\sigma_s}$, where $s(\cdot)$ denotes scalar projection onto the semantic dimension, $\mathbf{\hat{y}}_{x_j + 2\sigma_j}$ is the predicted embedding with covariate $j$ increased by 2SD, $\mathbf{\hat{y}}_{\mathrm{baseline}}$ is the prediction at intercept with covariates at means, and $\sigma_s$ is the SD of scalar projections from the original document embeddings. Significance stars reflect empirical p-values from residual permutation tests: $^{*}$ $p<0.05$, $^{**}$ $p<0.01$, $^{***}$ $p<0.001$. Dimension-specific effects are reported in Tables 3 and 4. Between countries (year fixed effects), higher electoral democracy is associated with movement toward the Causes pole and with modest movement toward the legalistic side of Responses; terrorism incidence is associated with movement toward condemnation on Causes and toward non-state perpetrators on Perpetrators; GDP per capita is associated with movement toward legalistic Responses. Within countries (year and country fixed effects), higher electoral democracy is positively related to movement toward the Causes pole and the State pole of Perpetrators, while higher terrorism incidence is positively related to movement toward the State pole of Perpetrators. Other covariates show limited within-country associations. These results affirm that particular aspects of meaning move differently over time and across countries. ## Discussion and Limitations This paper applies novel text analysis methods to a comprehensive dataset of UN speeches to identify state-level determinants of state rhetoric on terrorism. Moreover, it develops its expectations with reference to a range of International Relations theories that, while highly relevant, are rarely applied to international contestation over terrorism. In doing so, it finds support for important determinants of rhetoric which originate in domestic politics and in states’ efforts at strategic social construction. Domestically, leaders are attentive to political survival. They frame terrorism to mobilise their winning coalition, match available counterterrorism tools, and blunt home‑front threats. Strategically, when prevailing constructions risk delegitimising their practices, states stress lawful co‑operation and co‑ordinated enforcement on the Responses dimension. Although “cheap talk” is often thought to encourage misrepresentation, the evidence here rejects two common claims: feigned support for global counterterrorism and performative UN bloc politics. Formal theory frequently predicts that cheap talk lacks independent effects (Spaniel 2023), yet other work shows messages can endogenously generate costs. Trager (2010) demonstrates that threats become costly when targets can choose to invest in military capabilities in response, weakening the position of the threatening state relative to the counterfactual where no threat was made. Likewise, terrorism rhetoric that triggers suspensions of co‑operation or aid can impose endogenous costs beyond second‑order audience or reputational effects. These costs call for a theory of cheap talk that considers the material and strategic consequences of otherwise signals that are otherwise costless to send in states’ international networks and domestic politics. In addition, the research opens up new avenues of enquiry by examining the impact of political ideology and partisanship on elite rhetoric in international institutions. Notably, it shows that partisanship variables can be more impactful on rhetoric than some notable state-level indicators. Moreover, differences in a speaker’s left-right political ideology correspond to substantively significant differences on relevant semantic dimensions. Future work could deepen this attention to speaker-level attributes and intrastate domestic competition. Although this paper limits its scope to state-level determinants of rhetoric, the research question and the paper’s attention to UN politics naturally invites questions at a higher level-of-analysis: how do international relations affect the ways states articulate “terrorism”? Does rhetoric converge within international alliances? Or can states “buy” rhetorical concessions from other states with foreign aid, material threats, or other quid pro quos? The remaining chapters in this thesis explore some of these questions more closely, particularly in relation to foreign aid provision and how it might operate as a form of political leverage through political conditionality or related forms of coercion and/or incentivisation. Additionally, whilst the models include fixed-effects of time, there is also reason to believe that the effect of some covariates vary across time, and temporal interactions may be warranted. For instance, the inception of the Global War on Terror may have shifted the relationship between political ideology and rhetoric, or made the relationship between democracy and “repressive solutions” to terror more pronounced. Whilst such temporal dynamics are generally beyond the scope of the present analysis, Paper 2 examines a narrower period from the 1990s to the early 2010s which may elucidate some of these temporal patterns, and the third empirical paper presents an experiment conducted with contemporary foreign policy professionals, thus capturing dynamics in the present period. ## Conclusion I analyse every identifiable English-language use of the term “terrorism” in UN General Debate speeches from 1947 to 2022 (20,109 observations). Between countries and controlling for common time trends, state-level attributes are associated with robust differences in how states use the term. These include politico-geographic region, colonisation history, government ideological orientation, democracy, GDP per capita, institutional operation, physical integrity, and terrorism incidence. Politico-geographic region effects are large. In contrast, within-country changes in attributes explain comparatively little of the movement in language over time. In the baseline within estimator, only the operation of a formal legislature reaches significance, and even many autocratic states and hybrid regimes satisfy this criterion. When attention is restricted to semantic subspaces that focus on specific dimensions of meaning and are known to be salient in the international politics of terrorism, limited within-country movement appears: higher electoral democracy is associated with movement toward “root causes” on the Causes dimension and toward the State pole of Perpetrators, and higher domestic terrorism incidence is associated with movement toward the State pole of Perpetrators. Substantively, the results indicate that international rhetoric on terrorism is structured by enduring cross-national differences and by regionally patterned discourse. Long-run ideological orientation corresponds to systematic differences across states, but government ideology turnover does not reliably shift rhetoric within states in the short run. The effects vary across semantic dimensions, which underscores the need to disaggregate language into theoretically motivated facets of meaning. Fortunately, such operations are easy to perform with embeddings (Stoltz et al. 2025) and I have shown that they can be used to powerfully extend the $conText$ framework. On this note, I methodologically extend $conText$ in three ways: (i) multiple imputation to retain the full corpus and propagate imputation uncertainty; (ii) fixed-effects removal directly in embedding space to obtain tractable within estimators without large dummy sets; and (iii) projection of predicted embeddings onto semantic directions with inference in the original embedding space using residual-permutation tests and jackknife standard errors. These extensions provide a general approach for testing whether covariates are associated with targeted aspects of meaning in large-scale political text. Taken together, the findings support accounts that treat rhetoric as consequential rather than merely epiphenomenal. Talk is not uniform: it reflects durable cross-national structure and moves selectively along particular dimensions of meaning. And yet, the relative fixity of terrorism rhetoric within-states in the face of within-country changes in state-level attributes is surprising; this thesis turns to higher levels of analysis in the following chapters to explore how international counterparts can spur changes in another state’s terrorism rhetoric when even significant internal institutional change could not. ## References
Abrahms, Max. 2007. “Why Democracies Make Superior Counterterrorists.” *Security Studies* 16 (2): 223–53. .
Acharya, Upendra D. 2009. “War on Terror or Terror Wars: The Problem in Defining Terrorism.” *Denver Journal of International Law and Policy* 37 (4): 653–80.
Adler, Emanuel. 2010. “Damned If You Do, Damned If You Don’t: Performative Power and the Strategy of Conventional and Nuclear Defusing.” *Security Studies* 19 (2): 199–229. .
Aksoy, Deniz, David B. Carter, and Joseph Wright. 2012. “Terrorism In Dictatorships.” *The Journal of Politics* 74 (3): 810–26. .
Aksoy, Deniz, David B. Carter, and Joseph Wright. 2015. “Terrorism and the Fate of Dictators.” *World Politics* 67 (3): 423–68. .
Arora, Sanjeev, Yuanzhi Li, Yingyu Liang, Tengyu Ma, and Andrej Risteski. 2015. *A Latent Variable Model Approach to PMI-Based Word Embeddings*. arXiv. .
Asal, Victor, Brian J Phillips, R Karl Rethemeyer, Corina Simonelli, and Joseph K Young. 2019. *Carrots, Sticks, and Insurgent Targeting of Civilians*. 63 (7): 1710–35. .
Aytaç, S Erdem, and Ali Çarkoğlu. 2021. “Terror Attacks, Issue Salience, and Party Competence: Diagnosing Shifting Vote Preferences in a Panel Study.” *Party Politics* 27 (4): 755–66. .
Balcells, Laia, and Gerard Torrats-Espinosa. 2018. “Using a Natural Experiment to Estimate the Electoral Consequences of Terrorist Attacks.” *Proceedings of the National Academy of Sciences* 115 (42): 10624–29. .
Bali, Valentina A. 2007. “Terror and Elections: Lessons from Spain.” *Electoral Studies* 26 (3): 669–87. .
Baturo, Alexander, Niheer Dasandi, and Slava J. Mikhaylov. 2017. “Understanding State Preferences with Text as Data: Introducing the UN General Debate Corpus.” *Research and Politics* 4 (2). .
Berrebi, Claude, and Esteban F. Klor. 2008. “Are Voters Sensitive to Terrorism? Direct Evidence from the Israeli Electorate.” *American Political Science Review* 102 (3): 279–301. .
Blankenship, Brian. 2018. “When Do States Take the Bait? State Capacity and the Provocation Logic of Terrorism.” *Journal of Conflict Resolution* 62 (2): 381–409. .
Bolt, Jutta, and Jan Luiten Van Zanden. 2020. “The Maddison Project: Maddison style estimates of the evolution of the world economy. A new 2020 update.” .
Bonvecchi, Alejandro, and Emilia Simison. 2017. “Legislative Institutions and Performance in Authoritarian Regimes.” *Comparative Politics* 49 (4): 521–44. .
Breen Smyth, Marie, Jeroen Gunning, Richard Jackson, George Kassimeris, and Piers Robinson. 2008. “Critical Terrorism Studies–an Introduction.” *Critical Studies on Terrorism* 1 (1): 1–4. .
Brulin, Remi. 2013. “Defining ‘Terrorism ’: The 1972 General Assembly Debates on ‘International Terrorism ’ and their coverage by the New York Times.” *New York Times Examiner*, 16. [www.nytexaminer.com/2013/09/defining-terrorism](https://www.nytexaminer.com/2013/09/defining-terrorism).
Bueno de Mesquita, Bruce, Alastair Smith, Randolph M. Siverson, and James D. Morrow. 2003. *The Logic of Political Survival*.
Bueno De Mesquita, Ethan, and Eric S. Dickson. 2007. “The propaganda of the deed: Terrorism, counterterrorism, and mobilization.” *American Journal of Political Science* 51 (2): 364–81. .
Bush, George. 2002. *A/57/PV.2*. New York. .
Chelotti, Nicola, Niheer Dasandi, and Slava Jankin Mikhaylov. 2022. “Do Intergovernmental Organizations Have a Socialization Effect on Member State Preferences? Evidence from the UN General Debate.” *International Studies Quarterly* 66 (1): sqab069. .
Cohrs, J. Christopher, Sven Kielmann, Jurgen Maes, and Barbara Moschner. 2005. “Effects of Right-Wing Authoritarianism and Threat from Terrorism on Restriction of Civil Liberties.” *Analyses of Social Issues and Public Policy* 5 (1): 263–76. .
Conceição-Heldt, Eugénia da, and Patrick A. Mello. 2017. “Two-Level Games in Foreign Policy Analysis.” *Oxford Research Encyclopedia of Politics*, no. April: 1–25. .
Coppedge, Michael, John Gerring, Carl Henrik Knutsen, et al. 2022. *V-Dem Country-Year/Country-Date Dataset v11*. Varieties of Democracy (V-Dem) Project. .
Dingwall, Nicholas, and Christopher Potts. 2018. “Mittens: An Extension of GloVe for Learning Domain-Specialized Representations.” *Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers)* (New Orleans, Louisiana), 212–17. .
Dinstein, Yoram, and Fania Domb, eds. 2003. “Counter-Terrorism and the Use of Force in International Law.” In *Israel Yearbook on Human Rights, Volume 32 (2002)*. Brill \| Nijhoff. .
Edel, Mirjam, and Maria Josua. 2018. “How Authoritarian Rulers Seek to Legitimize Repression: Framing Mass Killings in Egypt and Uzbekistan.” *Democratization* 25 (5): 882–900. .
El Masri, Mirna, and Brian J. Phillips. 2021. “Threat Perception, Policy Diffusion, and the Logic of Terrorist Group Designation.” *Studies in Conflict & Terrorism*, December, 1–24. .
Fearon, James D. 1994. “Domestic Political Audiences and the Escalation of International Disputes.” *American Political Science Review* 88 (3): 577–92. .
Fearon, James D., and David D. Laitin. 2003. “Ethnicity, Insurgency, and Civil War.” *The American Political Science Review* 97 (1): 75–90. .
Finnemore, Martha. 1996. *National Interests in International Society*.
Finnemore, Martha, and Kathryn Sikkink. 1998. “International Norm Dynamics and Political Change.” *International Organization* 52 (4): 887–917. .
Gandhi, Jennifer, and Adam Przeworski. 2007. “Authoritarian Institutions and the Survival of Autocrats.” *Comparative Political Studies* 40 (11): 1279–301. .
Ganor, Boaz. 2002. “Defining Terrorism: Is One Man’s Terrorist another Man’s Freedom Fighter?” *Police Practice and Research* 3 (4): 287–304. .
Getmansky, Anna, and Thomas Zeitzoff. 2014. “Terrorism and voting: The effect of rocket threat on voting in Israeli elections.” *American Political Science Review* 108 (3): 588–604. .
Gould, Eric D., and Esteban F. Klor. 2010. “Does Terrorism Work?” *The Quarterly Journal of Economics* 125 (4): 1459–510. .
Government of the Republic of Serbia. 2003. *Statement Issued by the Government of the Republic of Serbia on 10 March 1998 Concerning the Situation in Kosovo and Metohija*. .
Green, Breanna, William Hobbs, Sofia Avila, Pedro L. Rodriguez, Arthur Spirling, and Brandon M. Stewart. 2025. “Measuring Distances in High Dimensional Spaces: Why Average Group Vector Comparisons Exhibit Bias, And What to Do about It.” *Political Analysis* 33 (3): 266–73. .
Grimmer, Justin, Margaret E. Roberts, and Brandon M. Stewart. 2022. *Text as Data: A New Framework for Machine Learning and the Social Sciences*. Princeton University Press.
Grimmer, Justin, and Brandon M. Stewart. 2013. “Text as Data: The Promise and Pitfalls of Automatic Content Analysis Methods for Political Texts.” *Political Analysis* 21 (3): 267–97. .
Haspeslagh, Sophie. 2013. “"Listing Terrorists": The Impact of Proscription on Third-Party Efforts to Engage Armed Groups in Peace Processes-a Practitioner’s Perspective.” *Critical Studies on Terrorism* 6 (1): 189–208. .
Hendrix, Cullen S., and Joseph K. Young. 2014. “State Capacity and Terrorism: A Two-Dimensional Approach.” *Security Studies* 23 (2): 329–63. .
Hensel, Paul R. 2018. *ICOW Colonial History Data Set, Version 1.1*. .
Herre, Bastian. 2023. “Identifying Ideologues: A Global Dataset on Political Leaders, 1945–2020.” *British Journal of Political Science* 53 (2): 740–48. .
Herring, Eric. 2008. “Critical Terrorism Studies: An Activist Scholar Perspective.” *Critical Studies on Terrorism* 1 (2): 197–211. .
Herschinger, Eva. 2013. “A battlefield of meanings: The struggle for identity in the UN debates on a definition of international terrorism.” *Terrorism and Political Violence* 25 (2): 183–201. .
Hoffman, Bruce. 2006. *Inside Terrorism*. Rev. and expanded ed. Columbia University Press.
Hopf, Ted. 1998. “The Promise of Constructivism in International Relations Theory.” *International Security* 23 (1): 171–200. .
Huddy, Leonie, Stanley Feldman, Charles Taber, and Gallya Lahav. 2005. “Threat, Anxiety, and Support of Antiterrorism Policies.” *American Journal of Political Science* 49 (3): 593–608.
Iida, Keisuke. 1988. “Third World Solidarity: The Group of 77 in the UN General Assembly.” *International Organization* 42 (2): 375–95. .
Jackson, Richard. 2005. *Writing the War on Terrorism: Language, Politics, and Counter-Terrorism*. Manchester University Press.
Jackson, Richard. 2007. “The Core Commitments of Critical Terrorism Studies.” *European Political Science* 6 (3): 244–51. .
Jackson, Richard. 2011. “In Defence of ‘Terrorism’: Finding a Way Through a Forest of Misconceptions.” *Behavioral Sciences of Terrorism and Political Aggression* 3 (2): 116–30. .
Jankin Mikhaylov, Slava, Alexander Baturo, and Niheer Dasandi. 2017. “United Nations General Debate Corpus.” Version V6. Harvard Dataverse. .
Jarvis, Lee, and Tim Legrand. 2018. “The Proscription or Listing of Terrorist Organisations: Understanding, Assessment, and International Comparisons.” *Terrorism and Political Violence* 30 (2): 199–215. .
Jenkins, Brian Michael. 1980. *The Study of Terrorism: Definitional Problems*. The RAND Corporation. .
Jenkins, Brian Michael, and J. J Johnson. 1975. *International Terrorism: A Chronology, 1968-1974*. R-1597-DOS/ARPA. RAND. .
Johnston, Alastair Iain. 2001. “Treating International Institutions as Social Environments.” *International Studies Quarterly* 45 (4): 487–515. .
Kentikelenis, Alexander, and Erik Voeten. 2021. “Legitimacy Challenges to the Liberal World Order: Evidence from United Nations Speeches, 1970–2018.” *The Review of International Organizations* 16 (4): 721–54. .
Khan, Aarish Ullah. 2005. *The Terrorist Threat and the Policy Response in Pakistan*. Policy Paper No. 11. Stockholm International Peace Research Institute.
Khodak, Mikhail, Nikunj Saunshi, Yingyu Liang, Tengyu Ma, Brandon Stewart, and Sanjeev Arora. 2018. *A La Carte Embedding: Cheap but Effective Induction of Semantic Feature Vectors*. arXiv. .
Kozlowski, Austin C., Matt Taddy, and James A. Evans. 2019. “The Geometry of Culture: Analyzing the Meanings of Class Through Word Embeddings.” *American Sociological Review* 84 (5): 905–49. .
Krebs, Ronald R., and Jennifer K. Lobasz. 2007. “Fixing the Meaning of 9/11: Hegemony, Coercion, and the Road to War in Iraq.” *Security Studies* 16 (3): 409–51. .
Laatikainen, Katie Verlin. 2017. “Conceptualizing Groups in UN Multilateralism: The Diplomatic Practice of Group Politics.” *The Hague Journal of Diplomacy* 12 (2-3): 113–37. .
Laver, Michael, Kenneth Benoit, and John Garry. 2003. “Extracting Policy Positions from Political Texts Using Words as Data.” *American Political Science Review* 97 (02). .
Magen, Amichai. 2018. “Fighting Terrorism: The Democracy Advantage.” *Journal of Democracy* 29 (1): 111–25. .
Mearsheimer, J. J. 2010. “The Gathering Storm: China’s Challenge to US Power in Asia.” *The Chinese Journal of International Politics* 3 (4): 381–96. .
Mikolov, Tomas, Edouard Grave, Piotr Bojanowski, Christian Puhrsch, and Armand Joulin. 2018. “Advances in Pre-Training Distributed Word Representations.” *Proceedings of the International Conference on Language Resources and Evaluation (LREC 2018)*.
Milton, Daniel, Amira Jadoon, and Jason Warner. 2021. “Needs or Symbols? The Logic of United Nations Counterterrorism Treaty Ratification.” *International Studies Quarterly*, 1–13. .
Onuf, Nicholas. 2012. *World of Our Making: Rules and Rule in Social Theory and International Relations*. Zeroth. Routledge. .
Osborne, Danny, Thomas H. Costello, John Duckitt, and Chris G. Sibley. 2023. “The Psychological Causes and Societal Consequences of Authoritarianism.” *Nature Reviews Psychology* 2 (4): 220–32. .
Park, Johann, and Valentina Bali. 2017. “International Terrorism and the Political Survival of Leaders.” *Journal of Conflict Resolution* 61 (7): 1343–70. .
Pemstein, Dan, Kyle Marquardt, Eitan Tzelgov, et al. 2022. “The v-Dem Measurement Model: Latent Variable Analysis for Cross-National and Cross-Temporal Expert-Coded Data.” *SSRN Electronic Journal*, ahead of print, January. 7th Edition Editions. .
Pennington, Jeffrey, Richard Socher, and Christopher Manning. 2014. “GloVe: Global Vectors for Word Representation.” In *Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)*, edited by Alessandro Moschitti, Bo Pang, and Walter Daelemans. Association for Computational Linguistics. .
Peterson, M. J. 2004. “Using the General Assembly.” Chap. 8 in *Terrorism and the UN: Before and After September 11*, edited by Jane Boulden and Thomas G. Weiss. Indiana University Press.
Piazza, James A. 2008. “Incubators of Terror: Do Failed and Failing States Promote Transnational Terrorism?” *International Studies Quarterly* 52 (3): 469–88. .
Pokalova, Elena. 2010. “Framing Separatism as Terrorism: Lessons from Kosovo.” *Studies in Conflict & Terrorism* 33 (5): 429–47. .
Putnam, Robert D. 1988. “Diplomacy and Domestic Politics : The Logic of Two-Level Games.” *International Organization* 42 (3): 427–60.
Ralph, Jason. 2007. “International Society and America’s War on Terrorism.” In *Defending the Society of States*, First. Oxford University PressOxford. .
Ramsay, Kristopher W. 2004. “Politics at the Water’s Edge: Crisis Bargaining and Electoral Competition.” *Journal of Conflict Resolution* 48 (4): 459–86. .
Rapoport, David C. 2017. “Terrorism as a Global Wave Phenomenon: Anti-Colonial Wave.” In *Oxford Research Encyclopedia of Politics*. Oxford University Press. .
Roberts, Margaret E., Brandon M. Stewart, and Richard A. Nielsen. 2020. “Adjusting for Confounding with Text Matching.” *American Journal of Political Science* 64 (4): 887–903. .
Rodriguez, Pedro L., Arthur Spirling, and Brandon M. Stewart. 2023. “Embedding Regression: Models for Context-Specific Description and Inference.” *American Political Science Review*, January, 1–20. .
Sandler, Todd. 2005. “Collective Versus Unilateral Responses to Terrorism.” *Public Choice* 124 (1/2): 75–93. .
Saul, Ben. 2005. “Attempts to Define ‘Terrorism’ in International Law.” *Netherlands International Law Review* 52 (1): 57–83. .
Saul, Ben. 2006. *Defining Terrorism in International Law*. Oxford Scholarship Online. .
Schmid, Alex P. 1992. “The Response Problem as a Definition Problem.” *Terrorism and Political Violence* 4 (4): 7–13. .
Schmid, Alex P. 2004. “Terrorism - The Definitional Problem.” *Case W. Res. J. Int’l L.* 36 (2): 375–419. .
Spaniel, William. 2023. *Formal Models of Crisis Bargaining: Applications in the Politics of Conflict*. 1st ed. Cambridge University Press. .
Stampnitzky, Lisa. 2011. “Disciplining Terror: How Experts Invented ‘Terrorism’.” *Disciplining Terror: How Experts Invented "Terrorism"*, 1–244. .
START. 2022. *Global Terrorism Database 1970 - 2020*. .
Stohl, Michael. 2008. “Old Myths, New Fantasies and the Enduring Realities of Terrorism.” *Critical Studies on Terrorism* 1 (1): 5–16. .
Stoltz, Dustin S., Marshall A. Taylor, and Jennifer S. K. Dudley. 2025. “A Tool Kit for Relation Induction in Text Analysis.” *Sociological Methods & Research* 54 (2): 565–604. .
Tilly, Charles. 2004. “Terror, Terrorism, Terrorists.” *Sociological Theory* 22 (1): 5–13. .
Trager, Robert F. 2010. “Diplomatic Calculus in Anarchy: How Communication Matters.” *American Political Science Review* 104 (2): 347–68. .
United Nations General Assembly. 2003. *Revitalization of the Work of the General Assembly*. .
Waever, Ole. 1995. “Securitization and Desecuritization.” In *On Security*, edited by Ronnie D. Lipschutz. Columbia University Press.
Weiss, Thomas G. 2009. “Moving Beyond North–South Theatre.” *Third World Quarterly* 30 (2): 271–84. .
Wendt, Alexander. 1999. “Social Theory of International Politics.” *Foreign Affairs* 78 (6): 142. .
Whitaker, Beth Elise. 2008. “Reluctant Partners: Fighting Terrorism and Promoting Democracy in Kenya.” *International Studies Perspectives* 9 (3): 254–71. .
Whitaker, Beth Elise. 2010. “Compliance Among Weak States: Africa and the Counter-Terrorism Regime.” *Review of International Studies* 36 (3): 639–62. .
Whitaker, Beth Elise. 2014. “African Reactions to U.S. Counter-Terrorism Policies: Kenya, From Reluctance to Resolve.” In *U.S. Strategy in Africa*, edited by Maya Kandel. Études de l’IRSEM N°36.
Zenn, Jacob. 2023. “War on Terror 2.0: Threat Inflation and Conflation of Far-Right and White Supremacist Terrorism After the Capitol ‘Insurrection’.” *Critical Studies on Terrorism* 16 (1): 62–97. .
Zulaika, Joseba, and William Douglass. 1996. *Terror and Taboo: The Follies, Fables, and Faces of Terrorism*.
## Appendix Table A.1. Table of opposed word-pairs used to construct the “Legal and Law-Enforcement” to “Repressive and Militaristic” Responses dimension | **Legal and Law-Enforcement** | **Repressive and Militaristic** | |:------------------------------|:--------------------------------| | illegal | military | | enforcement | war | | justice | militarily | | penalty | fight | | imprison | security | | rehabilitate | force | | police | defeat | | identify | win | | monitor | victory | | crime | lose | | criminal | combat | | criminality | campaign | | law | offensive | | convict | battle | | prosecute | warfare | | courts | | | proscribe | | Table A.2. Table of opposed word-pairs used to construct the “Root-Causes” to “Consequences and Condemnation” Causes dimension | **Root-Causes** | **Consequences and Condemnation** | |:-----------------|:----------------------------------| | root | senseless | | cause | needless | | desperation | unjustified | | despair | renounce | | manifestations | condemn | | underlying | indefensible | | freedom-fighters | | Table A.3. Table of opposed word-pairs used to construct the “Non-State Perpetrators” to “State Perpetrators” Perpetrators dimension | **Non-State Perpetrators** | **State Perpetrators** | |:---------------------------|:-----------------------| | non-state | tyrannical | | militants | regime | | guerrillas | despotic | | extremists | repression | | separatists | genocide | | combatants | internment | | fighters | disappearances | | warlords | | Table A.4. Terrorism Tokens Table. Only lists tokens that appeared five or more times. | **Token** | **Number of Instances in Corpus** | |:-----------------|:----------------------------------| | terrorism | 14283 | | terrorist | 4015 | | terrorists | 1731 | | terroristic | 23 | | counterterrorism | 22 | | bioterrorism | 8 | | terrorism’s | 6 | | antiterrorism | 5 | [^1]: