2.1 Research design

We analyze all the 13,773 unique speeches held by member states' representatives before the First Committee of the UNGA between 1993 and 2019. Manually encoding these amounts of text would be a very labor- and time-intensive enterprise. Therefore, we rely on automated text analysis. Although the range of available methods has increased dramatically in recent years (Mitrani et al., Reference Mitrani, Adams and Noy2022), the most prominent approaches for scaling political texts remain Wordscores (Laver et al., Reference Laver, Benoit and Garry2003) and Wordfish (Slapin and Proksch, Reference Slapin and Proksch2008). Yet both algorithms are restricted to scaling positions on a single latent dimension and thus may provide an incomplete picture of states' preferences on a multifaceted issue such as arms control and disarmament.

Lauderdale and Herzog (Reference Lauderdale and Herzog2016) address this limitation. In a first step, they estimate Wordfish scores for debates over individual topics. In a second step, they map the debate-level scores via factor analysis to a lower-dimensional space. Unfortunately, it is not always easy to identify and delimit political debates. And where it is possible to delimit debates, they are not always structured along a single conflict dimension. Both challenges are particularly relevant for texts that cover multiple topics under a single headline, such as many of the speeches before the UNGA. The most obvious way to solve this challenge is by reading the documents and assigning sections and paragraphs to individual topics. Yet, with large bodies of text, this approach does not appear feasible.

Here, we propose to overcome these challenges in three steps. In the first step, we estimate an STM on all 13,773 unique speeches in our corpus. In the second step, we combine these speeches at the country-year level. This results in 3117 observations for which we can estimate member states' annual positions. To delimit debates, we sample terms by their relevance for each of the latent topics identified in step 1 and then estimate Wordfish models for 1000 samples per topic. We argue and demonstrate empirically, that the mean values across these samples constitute topic-specific positions. In the third step, we apply PCA to reduce the topic-specific positions to a lower-dimensional latent conflict space. Figure 1 offers a summary of our three-step approach and guides the reader through our analysis.

Figure 1. Unfolding the rhetorical conflict space in the UNGA's First Committee in three steps.

2.2 Exploration of latent topics

At the beginning, the corpus was stemmed, cleared of numbers, stop words, and rare words that appear in less than five documents. In addition, we deleted all country names, as these appear primarily in the speaking country's own speeches, which may bias the results and complicate the interpretation. Here, our main objective is to explore the topic structure underlying the corpus of UNGA speeches. Therefore, we opt for an unsupervised topic model which is ideal for exploring a corpus without requiring prior knowledge of all relevant topics. Specifically, we apply STM as implemented in the corresponding R package (Roberts et al., Reference Roberts, Stewart, Tingley, Lucas, Leder-Luis, Gadarian, Albertson and Rand2014). Compared to other latent topic models, STM improves the assignment of words to latent topics by exploiting meta data such as the UNGA session and the country-affiliation of the speaker.

There is no unique answer to choosing the optimal (let alone “correct”) number of topics, but the choice must be justified by (i) the predictive accuracy of the model and (ii) the interpretability of the resulting latent topics. Optimizing predictive accuracy and interpretability can result in a trade-off (Chang et al., Reference Chang, Gerrish, Wang, Boyd-Graber and Blei2009). Appendix A reveals a first elbow effect for semantic coherence, predictive accuracy, and residuals at k = 5 topics and a second one at k = 8 topics. In the following, we present the results for k = 8 topics, but our overall substantive findings are robust against choosing any number of topics between k = 5 and k = 15.

STM provides us with information on the relevance of each term for each topic. More precisely, the topic model offers two types of conditional probabilities: first, how frequently a term appears in debates on a given topic (frequency score), and second, how exclusive a term is to debates on this topic (exclusivity score). For example, the term “nuclear” has high frequency and exclusivity scores for debates over nuclear weapons (topic 6, see below). In contrast, “issu” has a high frequency, but a low exclusivity, while the opposite is the case for “plutonium.” The term “landmin” has both a low frequency and a low exclusivity for this topic. The intuition is that terms can have high-frequency scores for all topics. By contrast, terms cannot have high exclusivity scores for all topics. The commonly used FREX scores weigh the exclusivity of a term with its frequency within a topic (Bischof and Airoldi, Reference Bischof and Airoldi2012). Appendices B and C contain the 50 terms with the highest FREX scores for the k = 5 and k = 8 models.

Here, we focus on the k = 8 model (Table 1). Two of the eight topics (topics 2 and 5) do not substantively deal with arms control, but primarily cover procedural matters. This is illustrated by terms such as paragraph, vote, draft, and sponsor (topic 2), or session, agenda, chairman, and membership (topic 5).

Table 1. Latent topics structuring the debates in the UNGA's First Committee

Among the six more substantive topics, one covers nuclear arms control (topic 6)—nuclear, npt, ctbt, and iaea are among the terms with the highest FREX scores. Terms such as chemic, opcw, biolog, and mass indicate that topic 4 focuses on WMD other than nuclear weapons, in particular chemical and biological weapons (CBWs).

The issue of conventional weapons is split into two topics. Topic 3 focuses on indiscriminate weapons (e.g., landmin, ottawa, indiscrim, ccw), in particular anti-personnel mines. In contrast, topic 8 primarily covers small arms and light weapons and their proliferation (e.g., trade, light, small, att).

Topic 1 encompasses matters of regional cooperation and conflicts (e.g., region, latin, africa, mediterranean). Interestingly, this does not include the Middle East, which is part of the nuclear topic. This, however, is not too surprising, given that the Middle East conflict has generally been delegated to the Fourth Committee, whereas the First Committee only deals with issues related to the establishment of a nuclear-weapon-free zone in this region.

Finally, topic 7 primarily includes armament of (cyber-)space (space, outer, missil, telecommun), but also terms related to transparency and confidence-building measures.

2.3 Scaling of topic-specific positions

The next step is to estimate states' positions for each of the eight topics. To this end, we aggregate the individual speeches to the country-year level. This results in 3117 observations for which we can estimate member states' annual positions. Given the observation period (26 years) and the number of UN member states (188 in 1993, 193 in 2019), it is apparent that many smaller states do not speak before the First Committee in every year.

As described above, the STM provides us with information on how exclusively each term has been used in connection with specific topics. More precisely, a term's exclusivity score measures the probability for a topic given the occurrence of the term. We exploit the exclusivity scores to estimate Wordfish models, because, unlike Wordscores, this algorithm is more exploratory and does not require us to identify the conflict dimensions a priori. More precisely, we estimate 1000 Wordfish models for each topic, drawing subsamples of all terms for every single model. Here, the exclusivity scores determine the probability for each term to be sampled. For example, the terms “nuclear” and “plutonium” are part of many subsamples for the nuclear topic, whereas “issu” and “landmin” are only rarely sampled for this topic.

Estimating Wordfish models on each subsample provides us with topic-specific distributions of all the relevant parameters, most importantly the positions for every country-year and the term discrimination parameters. Similar to factor scores in standard factor analysis, the latter enable us to interpret the latent rhetorical dimension. The mean values of these distributions are the expected topic-specific positions and the expected topic-specific discrimination parameters. The variations of these distributions give us information on the uncertainty of our topic-specific estimates (Appendix D).

Importantly, scaling models like Wordfish (Slapin and Proksch, Reference Slapin and Proksch2008) and Wordscores (Laver et al., Reference Laver, Benoit and Garry2003) are based on saliency theory (Budge, Reference Budge2001). A latent conflict only emerges if one speaker emphasizes one alternative, whereas another speaker emphasizes another alternative. However, some debates are not characterized by conflicting positions but rather by different levels of salience. For example, with regard to the regulation of space armament, our Wordfish models estimate speakers' salience rather than their conflicting positions.

In the following, we do not discuss all topics in detail, but primarily focus on the nuclear and landmines topics as illustrative examples. While we briefly describe the four remaining substantive topics, we refrain from interpreting the topics that merely deal with administrative issues.

For the nuclear topic, Wordfish identifies a conflict between those states that seek to preserve the current nuclear order and those that challenge the status quo. Terms with negative term discrimination parameters are related to non-proliferation measures and bilateral agreements, including but not limited to those between Russia and the United States (fmct, verif, start, bilat). Both types of agreements strengthen the status quo by design; the former because they prevent the spread of weapons and technology to states that do not yet possess them (Müller et al., Reference Müller, Becker-Jakob, Seidler-Diekmann, Müller and Wunderlich2013a), the latter because they exclude less powerful states (Johnston, Reference Johnston1996; Fehl, Reference Fehl2014). In addition, terms associated with critical views of nuclear disarmament (steadi, realist, pragmat, nato) are also found at the negative end of the scale. The other extreme is characterized by terms indicating opposing views and criticizing, for example, the unbalanced nature of the existing nuclear regimes (double, discrimin, imbal, inalien) and the slow pace of disarmament (pressure, failure, insist). Accordingly, we find states that benefit from the status quo on the negative side—in particular the permanent members of the Security Council (P5) and their allies. On the positive side, we find challengers to the status quo, especially from the Middle East, but also states such as Vietnam and Venezuela.

Regarding the landmines topic, Wordfish identifies a latent conflict between humanitarian and security framings of these weapons. The positive side includes terms emphasizing the harm these weapons cause, especially to civilians (contamin, survivor, displac, disabl, devast). The negative side mainly discusses indiscriminate weapons in security-related terms, emphasizing for example their military utility (occupi, troop, aggress, armament). In addition, it focuses on agreements adopted before the “humanitarian turn” in arms control that took place in the mid- and late-1990s (cfe, register) (Wisotzki, Reference Wisotzki, Müller and Wunderlich2013). In line with this, the negative end is occupied by states suffering from the unrestricted influx of landmines and other types of indiscriminate weapons. These include, for instance, Bosnia, Cambodia, and various states in Latin America and sub-Saharan Africa. States with low position estimates include those that either oppose stricter regulations (e.g., Russia, Syria, United States) or are at least not affected by landmines (e.g., Germany, France, United Kingdom).

Among the other topics, the main conflict structuring debates over CBW separates (alleged) users and possessors of these weapons justifying and defending themselves (fals, accus, occupi, baseless, aggress) from those that stand up for the CBW taboos (coordin, coalit, outreach, ratifi). With regard to the arms trade topic, terms with high positive discrimination parameters indicate the immediate negative impact of conventional weapons (traffic, demobil, displac, scourg, refug, violenc), while procedural terms (sponsor, coalit, regist, streamlin, discuss) yield negative term discrimination parameters. The regional topic captures disputes over separate territorially limited conflicts and arms control efforts (africa, caribbean, sahel, mercosur) from those of transregional relevance (western, atlant, caucasus). Regarding the space topic, Wordfish effectively measures the salience of debates on great powers' space armament (orbit, sky, satellit, nato, abm) versus other, seemingly unrelated, topics (traffic, social, climat, inalien).

2.4 Identification of the rhetorical conflict space

In this step, we map our estimates of topic-specific positions into a lower-dimensional rhetorical conflict space using PCA. We expect that the rhetorical conflicts over several topics are aligned along the same latent dimensions (Lauderdale and Herzog, Reference Lauderdale and Herzog2016). Specifically, we opt for varimax rotation to enhance the interpretability of the results by reducing cross-factor loadings. Note that factor analysis leads to very similar substantive results.

The results are unambiguous (see Table 2): the rhetorical conflict in the First Committee is structured along two latent dimensions, which explain approximately 78 percent of the variation observable in the topic-specific positions. The first latent conflict deals with conventional weapons, including landmines as well as the arms trade, and their use in regional conflicts. Positions on administrative and organizational matters are correlated with this conflict, too. The second latent conflict deals with WMDs, including nuclear arms, CBW, and new weapons in (cyber-)space.

Table 2. Principal components underlying the topic-specific positions

Bold numbers indicate assignment of topics to components.

We can interpret the latent components using the topic-specific term discrimination parameters estimated in step 2. However, in doing so we must take two aspects into account. First, we need to rotate the term discrimination parameters using the component scores. This is standard practice in the application of PCA. Second, we must consider each term's relevance for the estimation of each latent component. For this purpose, we aggregate the weighted exclusivity scores estimated in step 1, weighted by the topic's absolute component score. The latter represents each topic's relative importance for each component. The higher the aggregated exclusivity, the more important a term has been for estimating member states' positions on the respective latent component.

Figure 2 depicts the rotated term discrimination parameters for the latent rhetorical contestation over conventional arms. Large positive discrimination parameters indicate local violent conflicts and cooperation (sahel, african, caribbean, mercosur), the humanitarian impact of conventional weapons (violenc, displac, existenti, contamin), and the need for international assistance to overcome these issues (unregul, assist, donor, att). By comparison, large negative discrimination parameters primarily point at procedural matters and bureaucratic language (paragraph, comment, organiz), a limited substantive interest in the regulation of conventional arms (abstain, reserv, realist), and a focus on security matters (defens, aggress, strike). The underlying conflict dimension thus delineates states that frame arms control as a humanitarian issue from states that see arms control as a matter of national security.

Figure 2. Rotated term discrimination parameters (betas) and aggregated exclusivity scores for the latent conflict over conventional weapons (terms mentioned in the text depicted in bold as illustrative examples).

Figure 3 depicts the rotated term discrimination parameters for the latent conflict over WMDs. Large negative term discrimination parameters indicate support for the current arms control regimes and their reinforcement as agreed during and after the end of the Cold War. Typical terms either refer to bilateral and multilateral treaties (abm, fmct, start, bwc) and intentions to strengthen these further (enlarg, underpin, reinforc, deepen) but also to restrain overambitious disarmament efforts (steadili, ration, realist, pragmat). By contrast, terms with large positive discrimination parameters point toward critique of and challenges to the established order (reject, failur, weak), and in particular double standards and the discriminatory nature of arms control regimes (doubl, discrimin, imbal). This also includes references to the “inalienable right” to use nuclear energy (inalien, instal, energi) and defensive language addressing accusations of violating established rules and regulations (baseless, unjustifi, innoc). Hence, the second conflict over WMDs delineates supporters of the status quo from states that are challenging the status quo.

Figure 3. Rotated term discrimination parameters (betas) and aggregated exclusivity scores for the latent conflict over WMD (terms mentioned in the text depicted in bold as illustrative examples).

In sum, our data and methodological approach helps us to understand states' conflict over arms control and disarmament in two crucial ways. First, we establish that states' disagreements in this policy field center around two dimensions: “humanitarian vs. security” and “status quo vs. challenging.” While the latter dimension resembles conflict dimensions previously identified through other types of data (see, e.g., Bailey et al., Reference Bailey, Strezhnev and Voeten2017; Barnum and Lo, Reference Barnum and Lo2020), our work is, to our knowledge, the first to systematically demonstrate the relevance of the former dimension for debates around conventional weapons.Footnote ¹ Second, our approach allows us not only to identify the two conflict dimensions, but also the specific types of arms control each of them is associated with. The “humanitarian vs. security” dimension is primarily relevant for discussions around conventional arms control and matters of regional conflict and cooperation, whereas disagreements concerning WMD and new technologies evolve around the “status quo vs. challenging” dimension. This further enhances our knowledge of states' arms control preferences.

3.1 Face validity

3.1.1 Conventional weapons

To explore the validity of our estimates for the first conflict dimension, we first inspect the cross-country variation (see Figure 4) before we examine changes over time. In line with our interpretation, the “humanitarian side” of the rhetorical conflict over conventional arms is primarily occupied by states such as the D.R. Congo (DRC), Cambodia (CAM), and Honduras (HON). On the “security side” of the conflict we find Western democracies, but also the MENA region and Brazil (BRA), China (CHN), Russia (RUS), and North Korea (PRK). In between we find states with mixed interests. This includes, among others, South Africa (SAF) and Turkey (TUR) as well as Western states that are not affected by armed conflicts themselves but incorporate the humanitarian aspects of arms restrictions into their foreign policy considerations—for example, Denmark (DEN).

Figure 4. Member states' average positions in the UNGA First Committee's two-dimensional rhetorical conflict space between 1993 and 2019.

When assessing positional changes (see Figure 5), it is notable that the average scores become progressively more positive over time. This is in line with the focus of conventional arms control shifting more and more to the containment of intrastate violence and human suffering (Wisotzki, Reference Wisotzki, Müller and Wunderlich2013). This is illustrated by the negotiation and adoption of various humanitarian arms control treaties in the late 1990s and 2000s—such as the Ottawa Treaty, the Convention on Cluster Munitions, and the Arms Trade Treaty.

Figure 5. Development of positions on the conventional weapons dimension. Note that the graphs in Figures 5 and 6 have been smoothed with the geom_smooth function implemented in the R package ggplot2.

Figure 6. Development of positions on the WMD dimension.

Next, we selected four states that are known to have changed their positions on conventional arms control over time: the United States, Colombia, Russia, and Ukraine. Being the largest exporter of conventional weapons and a nation with exceptionally liberal gun laws, the United States has consistently shown hesitancy toward conventional arms control measures (Wisotzki, Reference Wisotzki, Müller and Wunderlich2013; Wunderlich et al., Reference Wunderlich, Hellmann, Müller, Reuter, Schmidt, Müller and Wunderlich2013). While we do not observe a substantial shift between the presidencies of Clinton and Bush, it is important to note that the global rhetoric changed substantively during that period. Accordingly, the Bush administration, justifying its particularly negative position toward arms control with security concerns, deviated much more from the “average” rhetoric and “opposed new global norms in humanitarian arms control” (Wunderlich et al., Reference Wunderlich, Hellmann, Müller, Reuter, Schmidt, Müller and Wunderlich2013, 172). Under Obama, the US position underwent a major change, embracing arms control measures and putting a stronger emphasis on the humanitarian dimension than it had before (Stohl, Reference Stohl2010). This positional shift was subsequently halted after rising international tensions from 2014—and then completely reversed by the Trump administration, which actively dismissed humanitarian issues in its conventional arms control policy (Stohl, Reference Stohl2021).

By contrast, the Colombian government has experienced a decades-long civil war and been exposed to the negative consequences of conventional weapons proliferation. As a reaction to the war's most severe phase in the late 1990s and early 2000s (UCDP, 2023a), the government began to engage in peace negotiations and disarmament, demobilization, and reintegration efforts (Angelo, Reference Angelo2016). Moreover, it has largely followed—and helped to shape—the humanitarian approach to conventional arms control (Zughni, Reference Zughni2012). Ever since, Colombia has laid a primary focus on humanitarian aspects and assistance to affected countries and has been a strong proponent of conventional arms control (Bromley and Malaret, Reference Bromley and Malaret2017).

Ukraine also largely followed the humanitarian rhetorical shift on conventional weapons—especially after democratization efforts and shifts in its foreign policy orientation in the early 2000s (Åslund, Reference Åslund2009). However, the government's rhetoric was completely reversed after Russia annexed Crimea and subsequently backed separatist militias in Donetsk and Luhansk (UCDP, 2023b). This reflects our interpretation that involvement in conflicts does not necessarily lead to a stronger emphasis on humanitarian issues. Instead, it depends on the specific type of conflict and its relevance for international politics and the global order. More precisely, interstate tensions that are important beyond these states' borders—as is the case with the Russo–Ukrainian conflict—bring security matters back into focus. By contrast, intrastate conflicts with a limited geographical scope, such as the Colombian civil war, rather shift states' focus to humanitarian assistance and the negative consequences of conventional weapons.

Finally, similar to the United States and in line with our expectations, Russia remains on the security side of the conflict dimension. Nevertheless, we observe some relevant changes in its rhetoric over time. While the Russian government largely followed the general trend in the late-1990s and early-2000s and developed a somewhat more humanitarian focus, it reversed its position back a decade later. This reflects Russia's increasingly aggressive and expansive foreign and military policies during this period (Götz and MacFarlane, Reference Götz and MacFarlane2019), which led to an even greater focus on security matters than at the beginning of the observation period.

3.2 Validity: existing measures

To further validate the rhetorical conflict space underlying debates in the First Committee, we compare them to existing measures of states' positions on arms control. Specifically, we assess how our position estimates—aggregated to the country levelFootnote ²—correlate with the measure of state support for arms control introduced by Risse (Reference Risse2023), also averaged to the country level (see Table 3). Appendix E provides further correlation analyses with other position estimates that also support our substantive interpretation of the two dimensions.

Table 3. Correlations with support for arms control (Pearson's r)

Note: Significance levels: ***p < 0.001; **p < 0.01; *p < 0.05; ^†p < 0.1.

Risse (Reference Risse2023) measures states' arms control preferences through manually coded UNGA resolutions on arms control passed between 1994 and 2016 and the corresponding voting records. In contrast to our conflict dimensions, he defines his conflict dimension—support versus non-support for arms control—a priori. His coding of resolutions' topics allows us to construct sub-indices corresponding to the topics of our two conflict dimensions.

We argue, first, that our position estimates on the regulation of conventional weapons should be positively correlated with Risse's measure, because proponents of conventional arms control tend to frame it as a humanitarian issue rather than a matter of security (Wisotzki, Reference Wisotzki, Müller and Wunderlich2013). Second, we expect a weak correlation between support for arms control and the WMD dimension, as the latter reflects conflicts over the type of arms control (e.g., non-proliferation vs. disarmament) rather than the degree of arms control.

The correlation analyses are in line with our interpretations of both dimensions. Our conventional weapons dimension is, as expected, positively and moderately correlated with Risse's measure. In contrast, the position estimates on WMDs are only weakly correlated. While the focus here is on the degree rather than the direction of the correlation, the negative coefficient indicates that status quo-oriented states are somewhat more supportive of arms control than challengers. This is also not too surprising, given that several important WMD control agreements have been criticized for favoring more powerful countries—which are also more status quo-oriented (Müller et al., Reference Müller, Becker-Jakob, Seidler-Diekmann, Müller and Wunderlich2013a).

3.3 Validity: exogenous variables

As a final step, we explore two relevant exogenous explanations for states' rhetorical positions in the UNGA's First Committee. First, we assess the effect of arms exports and imports (SIPRI, 2023) on positions in UNGA debates over conventional weapons. We expect that a greater economic stake in the arms trade is associated with a stronger focus on security matters rather than humanitarian issues and thus more negative scores on the conventional weapons dimension. Second, we regress the estimates on the WMD conflict dimension on two dummy variables indicating the possession (or alleged possession) of nuclear weapons and chemical weapons (Bleek, Reference Bleek2017; Arms Control Association, 2022). Here, we expect more ambiguous results. While nuclear weapons and deterrence remain an integral part of the post-Cold War global order (Krause and Latham, Reference Krause and Latham1998), chemical weapons are under a “taboo” and banned by the widely recognized Chemical Weapons Convention (Price, Reference Price2019). Thus, chemical weapons possession constitutes a challenge to the status quo and should be positively related to our estimates, whereas we expect the opposite relationship for nuclear weapons possession.Footnote ³

All models control for regional affiliation (Gleditsch et al., Reference Gleditsch, Wallensteen, Eriksson, Sollenberg and Strand2002; Davies et al., Reference Davies, Pettersson and Öberg2023), level of electoral democracy (Coppedge et al., Reference Coppedge, Gerring, Knutsen, Lindberg, Teorell, Altman, Bernhard, Cornell, Fish, Gastaldi, Gjerløw, Glynn, Hicken, Lührmann, Maerz, Marquardt, McMann, Mechkova, Paxton, Pemstein, von Römer, Seim, Sigman, Skaaning, Staton, Sundström, Tzelgov, Uberti, Wang, Wig and Ziblatt2023), GDP per capita (UNSD, 2023), involvement in armed conflicts (Gleditsch et al., Reference Gleditsch, Wallensteen, Eriksson, Sollenberg and Strand2002; Davies et al., Reference Davies, Pettersson and Öberg2023), and national material capabilities (CINC) (Singer et al., Reference Singer, Bremer, Stuckey and Russett1972; Singer, Reference Singer1988). Due to the clustered data structure, we include random effects for country and year. Models that exclude the control variables and that include all four independent variables are reported in Appendix F and largely produce similar results.

The effect of arms trade turns out as expected. Table 4 reveals that speeches by states with high arms trade volumes see conventional arms control from a perspective of national security. Furthermore, the possession of nuclear weapons is significantly related to a stronger status quo orientation. By contrast, states with chemical weapons stockpiles are significantly more likely to challenge the status quo. Thus, the regression analyses confirm our expectations and further validate our substantial interpretation of both conflict dimensions.

Table 4. Regression analysis

Note: Standard errors in parentheses.

Significance levels: ***p < 0.001; **p < 0.01; *p < 0.05; ^†p < 0.1.