Entrenching positions? The dynamics of Brexit negotiations mirrored in British, Irish, and EU executives’ speeches

ABSTRACT Brexit constitutes an unprecedented crisis for the European Union and its member states. Three affected parties – the United Kingdom, the EU and Ireland – worked hard to protect their interests, making the negotiations difficult, especially regarding the Irish land border. While the key interests of the negotiating partners are clear, it is less known how and to what extent these interests were communicated by the political representatives of the three countries. We approach Brexit as a critical juncture and base our work on the framing literature. We use a unique dataset including political speeches made by EU Commission members and the UK and Irish government officials during a period from January 2016 until January 2020, relying on the innovative automated text analysis approach of keyword-assisted topic modelling. Beyond insisting on their interests, communication reflects the socio-political context. We find indications that speeches reflect the flow of the negotiations. Finally, our analysis reveals topics that largely flew under the media radar, such as research and protection.


Introduction
Brexit negotiations have been described as difficult and 'very serious', oscillating between entrenched positions where all sides try to defend their own key interests and characterised by profound domestic political quarrels on the British side and an exacerbated legiti'macy crisis on the EU's side (e.g. Braun 2020). In such a crisis situation, political actors enter political contests, not only by trying to find solutions which favour the constituencies they represent (e.g. Boin, Hart, and McConnell 2009) but also by strategizing to secure that their stakes in the process be considered (Olsson, Nord, and Falkheimer 2015). One of the most heatedly debated issues during negotiations was the border between Northern Ireland and the Republic of Ireland (i.e. the Irish land border, henceforth referred to as the 'border'). The prospect of a reintroduced 'hard' border, separating the Republic of Ireland from Northern Ireland as a part of the United Kingdom, makes Ireland an especially affected EU member state.
To understand the dynamics of Brexit negotiations, our study analyses the strategic communication of executive actors of the UK and the EU as well as Ireland. It advances the extant political science discussion about Brexit which has, until now, very often focused on political attitudes to explain motivations in the Brexit referendum. Naturally, studies have often dealt with the case of the UK (Hobolt 2016;Vasilopoulou and Talving 2019), the consequences for the EU (e.g. Hobolt and Rodon 2020) and in some cases Ireland (Gormley-Heenan and Aughey 2017; Hayward and Murphy

Brexit in political speeches
Brexit can be seen as one of the most fundamental tests for the EU in terms of political resilience. Brexit is a protracted crisis or a critical juncture in the development of the EU and the UK. Theoretically rooted in Historical Institutionalism (e.g. Capoccia and Kelemen 2007), critical junctures are characterised by a high degree of contingency where future developments are hard to predict. They open up windows of opportunities for actors pushing differing interests, seeing their chance to realise them. New paths, in our case, may arise out of Brexit for the UK, other EU member states like Ireland, as well as the EU as such. This is because of the externalities that the UK exit poses on the EU and its member states, especially Ireland. Brexit as a crisis of the EU is -as the EU itself -a multi-level phenomenon with consequences for the whole polity as well as each member state individually (e.g. Rose 2020).
While such situations of crises present opportunities for elites to push their agendas, democratic leaders are under pressure to find solutions that best serve the interests of those they represent (Boin, Hart, and McConnell 2009). Framing literature suggests that it is of crucial importance for them to convince the public of their qualities as crisis managers, and that they are adequate representatives to tackle crises in the public's interest. In other words, crises can challenge decision-makers and end in political failure, but they are also a chance to raise political support in institutions and bolster office-holders' profiles (Olsson, Nord, and Falkheimer 2015). They are a setting for political contests where 'contestants manipulate, strategize and fight to have their frame accepted as the dominant narrative' (Boin, Hart, and McConnell 2009, 81).
The House of Commons, the Dáil Éireann, and the European Parliament are prime fora for contesting different Brexit issues amongst executives and their (in-house) political opponents. Yet, Brexit negotiations as intergovernmental negotiations were less debated on parliamentary floors and were, instead, mostly conducted behind closed doors. For the EU, the Article 50 task force led by Chief Negotiator Michel Barnier represented the interests of the 27 remaining member states (European Commission 2020), whereas the British government's Department for Exiting the EU was responsible for negotiations on the UK's side (Government of the United Kingdom 2020). Their public statements and speeches are important sources to understand the political signals that executives send to the general public. Such signals are indicators also to affected parties regarding the central stakes decision-makers cannot disavow (e.g. Rauh, Bes, and Schoonvelde 2019). We, therefore, explore what aspects of Brexit are (made) more salient in executives' speeches, shedding light on the strategic interests brought forward as the negotiations progressed. We expect political executives' Brexit communication to be influenced by the respective vital interests of decision-makers and the political entity they represent (Boin, Hart, and McConnell 2009;Olsson, Nord, and Falkheimer 2015). Relating back to Brexit as a critical juncture and time of heightened contingency and unpredictability, it is, therefore, hard to theoretically predict what general issues would be more or less prominent in speeches. Against this backdrop, it is necessary to empirically inform expectations by understanding the macro contexts of the three cases included in our analysis.

The macro context or defending vital interests
The UK and the EU are the two main opponents in the political contests revolving around Brexit, where the former was faced with the dilemma to implement the 'will of the people' without having a viable concept (Gamble 2018). The EU's legitimacy, already a heatedly discussed topic long before Brexit (Follesdal and Hix 2006), was further questioned. An easy Brexit could incentivise other member states to follow the example whereas a hard Brexit would endanger a whole country's economy and thereby the welfare of its people. The role of Ireland is best described as an affected bystander caught in the middle. On the one hand, Ireland is somewhat bound by historical bonds to the UK which have been characterised by UK dominance. On the other hand, the EU is an economic and political shelter which Ireland as a small member state had depended upon already during the hardships of the financial crisis (Thorhallsson and Kirby 2012).

'Taking back control'
While the UK always had a turbulent relationship with the EU (Bulmer and Quaglia 2018), it is argued that immigration policy and anti-elitism formed the core of the macro-level context of Brexit (e.g. Hobolt 2016). Following the 2004 and 2007 EU enlargement rounds, the UK, as well as Ireland and Sweden, witnessed a significant rise in migration from new member states (Constant, Kahanec, and Zimmermann 2009). As a result of the Labour Government's decision not to introduce restrictions on the freedom of movement, migration became an important policy issue as the economic and financial crisis exploded towards the end of the 2000s. The migration crisis of 2015 was, not surprisingly, predominantly framed negatively, partially also through the engagement of the Leave campaigners (Hobolt 2016). In general, it has been argued that an 'anti-system' and 'anti-elite' backlash resulted in rising support for parties such as the United Kingdom Independence Party (UKIP) (Hopkin 2017).
Immigration and intra-EU mobility are thus seen as the main points of contention, motivating the pro-Brexit vote. However, it is also of fundamental economic importance to the United Kingdom to maintain access to the EU's Single Market as one of the most crucial trading partners in the world (Sampson 2017). In that sense, economic interests are somewhat at odds with the political goal to 'take back control' and regain political sovereignty (Hobolt 2016;Gamble 2018). As Schimmelfennig (2018) argues, this has led to an increasing readiness to make concessions in the course of negotiations due to the stronger bargaining position on the EU. As the UK turns away from the EU, it is, moreover, in its interest to strengthen and emphasise its relationships with countries and trading partners outside the EU, particularly the countries of the Commonwealth (Hearne, De Ruyter, and Davies 2019).
Furthermore, the topic of the border and, more specifically, the desire to avoid a hard land border between Northern Ireland and the Republic of Ireland was one of UK's negotiating aims. Considering the great importance of the border aspect for Ireland (and therefore also for the EU), we outline this aspect in more detail when discussing communication expectations for Ireland. Overall, (H1) we expect the topics of immigration and the economy to be most prominent in the speeches of the UK government's representatives.

Caught in the middle
Ireland is a small EU member state which was severely hit by the financial crisis. As a result of the banking sector collapse, a budgetary crisis in Irish state finances developed, which eventually led to the Irish government's acceptance of an EU/IMF rescue package in 2010 (Thorhallsson and Kirby 2012). As a result, the relationship with the EU, then, became more and more characterised by economic and political dependencies. However, the EU had also already been a crucial facilitator of the 'Good Friday Agreement' which, by means of an international agreement, established peace in Northern Ireland (Hayward and Murphy 2018). Regarding the UK, then, Ireland has always had strong political ties with the Kingdom which were, however, also characterised by UK dominance (Coakley 2018).
The border on the Irish island became the most crucial point on the Brexit negotiation agenda. Ireland threatened to veto any results that would introduce a hard border (Schimmelfennig 2018). On the one hand, this border has long been a site of terroristic activities that were largely settled with the Good Friday Agreement, an agreement between two sovereign states binding in international law. A hard border between EU member state Ireland and non-EU member state UK would risk this conflict being reinvigorated. In addition, breaching this agreement would entail legal obligations and adaptations of different legal frameworks at domestic, European, and international levels (McCrudden 2021). On the other hand, the economic consequences of having to implement EU customs policies would represent a significant difficulty, in terms of setting up new infrastructures such as cameras and border posts across a land border of almost 500 kilometres. Open borders with Northern Ireland, therefore, are an extremely important issue in economic and political terms (Burroughs 2015). In other words, the negative externalities of Brexit affect Ireland most heavily of all the (remaining) EU countries.
Ireland will, also in the future, be affected by the externalities of UK politics and therefore will, also due to historical bonds, have an interest in maintaining a good relationship with the UK. At the same time, it will have a strong interest in demonstrating unity with the EU which has proved to be a strong protective economic and political umbrella. Against this background, (H2) we expect the border issue to be most prominent in the speeches of the Irish government's representatives.

Maintaining legitimacy and credibility
For the European Union, Brexit questioned the rationale of supranational governance as such. In that sense, Brexit could mark the beginning of the end of the EU, if, for example, other countries with similarly negative attitudes towards the EU might follow the example of the UK (Hobolt 2016(Hobolt , 1259. The legitimacy of the EU has been a matter of debate already for a long time (Follesdal and Hix 2006), and ten years of crisis have led to a loss of public trust in many member states (Dotti Sani and Magistro 2016). Thus, the EU needs to remain politically united to maintain the legitimacy and credibility of EU politics as such. In that light, the EU emphasised an 'internal cohesiveness' approach to Brexit, without granting the possibility of politicising single issues or cherry-picking from the UK side to strengthen its own bargaining position (Šimunjak and Caliandro 2020).
Regarding concrete priorities for Brexit negotiations, a crucial interest of the EU was to guarantee the respect of existing rights for EU citizens in the UK as well as the settlement of the UK's financial assets and obligations towards the EU (Schimmelfennig 2018). The most crucial point on the agenda, mirroring the interests of its member state Ireland, was to avoid a hard Irish border. Overall, (H3) we expect that the Irish border and EU citizens, as well as the normative aspect of the Brexit agreement, as far as the 'internal cohesiveness' approach allows, will be a prominent topic in the EU executives' speeches.

Over time: UK Leadership and triggering Article 50
Finally, (H4) we expect to see changes over time regarding the issues politicians emphasise in their communication to citizens -i.e. the prominence of different topics in speeches. Our expectations are two-fold. First, for the UK in particular, shifts may be induced by changing leadership (see Figure 1), which may, in turn, affect the positions of parties affected by Brexit. Not only do individual politicians have different leadership styles (Andeweg 2014), which define how politics and government work, but their coming to a position also reflects societal and political changes to which communication adapts.
Second, the period before the triggering of Article 50 of the Lisbon Treaty especially was preceded by the insecurity of whether the UK's leadership would act on the results of the referendum. For such events, the literature agrees that a new event within a crisis will be accompanied by efforts to define the problem and shape public understanding (Baumgartner and Jones 2010). Therefore, the activation of Article 50 may be anticipated to affect not only what the UK's executives would decide to emphasise, but also how their EU and Irish counterparts communicate about Brexit. This action, representing significant punctuation in an already fragile equilibrium (Capoccia and Kelemen 2007), can be expected to result in significant interruptions of the temporarily established framing of the key interests. As a result, we assume that particularly core economic and political issues might come to be re-examined after changes in leadership or following an event such as the invocation of Article 50.
Based on the existing literature and the macro-context of each involved party, we formulated a number of preliminary expectations. In addition to these, we also engage in an inductive, qualitative exploration of how each of the affected parties discussed the most important Brexit-related issues. In this endeavour, then, we look inductively into data to identify which topics, beyond those referring to the key interests of the affected parties, were represented in direct communication by political executives.

Data and methodology
Our analysis first focuses on the topics emphasised within political executives' Brexit communication. We look into additional topics that may have been important in executives' speeches and were not highlighted as strategic interests in the public discourse. We compiled a dataset containing Brexitrelated speeches (n = 814) to different audiences, i.e. parliament and the general public, by all European Commissioners, the Chief Brexit Negotiator, and the President of the European Council, as well as all ministers/secretaries of state including the head of government for Ireland and the UK. The dataset includes speeches from 2016 (the year of the referendum) to the end of January 2020 (see  All speeches were scraped from the respective webpages using the R package rvest (Wickham 2016). Due to very different search routines in the respective databases, we decided to download all speeches and subset the complete dataset for our purposes later on. To identify Brexit-relevant speeches, we constructed a Brexit dictionary (see Appendix Table A2) based on Gadd (2019) and included only speeches mentioning at least one word from our dictionary. The resulting datasets include 183 (22.48% of all), 187 (22.97% of all), and 444 (54.55% of all) speeches for the European Union, Ireland, and the United Kingdom, respectively, ranging from 237 to 8918 words with an average length of 1613 words. Both Irish and UK speeches are longer on average (1771 words and 2214 words respectively) than EU speeches. A list of speakers and descriptive statistics reflecting the number of speeches per speaker is available in Table A1 in the Appendi For the analysis, we relied on the key-word assisted topic modelling (keyATM) as a semisupervised machine-learning approach implemented in the R package keyATM (Eshima, Imai, and Sasaki 2020). In topic modelling approaches, each word in a document has a certain probability of belonging to a topic while each document represents a mixture of topics (Roberts, Stewart, and Tingley 2017). Theoretically speaking, it is hard to define what a 'topic' is, as the model output only provides clusters of words that co-occur in documents. Topic models, thus, tell us 'what is being talked/written about' (Brown and Yule 1983, 73). Recent literature suggests that, under certain conditions, topic models can be used to identify frames in a text (e.g. Walter and Ophir 2019; Ylä-Anttila et al. 2018). We follow Ylä-Anttila et al. (2018) who argue that if topics are operationalised as connections between concepts and subject-specific data are analysed, topic modelling can render valid representations of frames in the sense of sub-issues. Following this reasoning, we suggest that executives use their speeches to communicate issues to citizens, but also to indirectly send a message to other parties to the crisis negotiations. Topics derived from the topic modelling approaches, if coherent and validated, can provide an insight into the frames, i.e. sub-issues, politicians use to attempt to shape a protracted crisis in their favour.
The keyATM is categorised as a probabilistic topic model which also allows an a priori definition of topics by feeding keywords to the model. Thus, on the one hand, the researcher should already have an idea of what to search for as this beforehand definition of topics makes the interpretation of results much easier. On the other hand, keyATM also allows the inductive exploration of the corpus, working like a 'standard' topic model. Thus, keyATM lends itself to our analysis where some key topics can be expected to turn up in executives' speeches (see earlier section) but also allows exploring data without prior expectations.
To define the keywords for our expected topics, we relied on the Lexicoder Topic Dictionaries (Albugh, Sevenans, and Soroka 2013) as well as the Brexit Glossary (Gadd 2019) (see Supplementary Material A2 for details). Derived from this Brexit glossary, we defined keywords for the three major political topics -immigration, the Irish border, and the Brexit agreement. From the Lexicoder Dictionaries, we additionally used the economic (foreign trade, finance, and labour) and international affairs topics that we expected to be discussed in speeches as well (Albugh, Sevenans, and Soroka 2013). 1 Keywords were defined as n-grams, thus not only taking single terms like immigration but also combinations of words into account, e.g. 'Good Friday Agreement'. In addition to these seven predefined topics, we also allowed three additional topics that were identified inductively by the model. While the seven predefined topics encompass the topics we hypothesiseto be present in speeches, the three additional topics allow us to research which other topics may have been directly communicated to citizens. We run a model on the aggregated corpus (including EU, Irish and UK speeches) as we are interested in the overall Brexit discourse and the contests between the three affected sides. Following Grimmer and Stewart (2013), the number of topics was selected in a two-stage process. We first estimated a model with a varying number of inductively defined issues, ranging from 0 to 10. In the second stage, we selected the final model based on the internal consistency of the topics. Finally, we assessed the suitability of the model from the perspective of statistical criteria (see Appendix Figure A1).

Results and discussion
Before discussing differences across actors and time, we briefly outline the ten topics our keyATM approach delivered (see Table 1). Seven topics were defined deductively based on predefined keywords as the topics we anticipate (see discussion on the Macro Context), to dominate the Brexit discourse. We distinguish between two types of dictionaries -Lexicoder Dictionaries (immigration, labour, international affairs, foreign trade, and finance) and Brexit dictionaries (border and Brexit agreement).

Deductive modelling approach
Turning to the topics predefined based on Lexicoder dictionaries, the immigration topic seems mostly associated with words such as work, support, government and development. Out of the keywords used to pre-define the immigration topic (marked by * in Table 1), the term 'work' features prominently among the top terms defining the topic. The labour topic emphasises the overall labour system and is occupied with future developments. Similarly, as for the immigration topic, from the keywords used to define the labour policy, only 'working' features prominently among the top terms. The international affairs topic, in turn, is dominated by British-EU relationships, where the bi-gram 'European Union' is the sole term from our list of keywords.
Foreign trade and finance are the two topics that seem to work best in terms of the predefined keywords -these selected keywords often appear and co-occur in the speeches thereby coherently defining the respective topics. The emphasis of the foreign trade is on 'global', 'markets', with an interesting focus on 'opportunities', while the finance topic has a rather straightforward focus on 'finance', 'markets' and 'investments'.
What about the topics for which we developed Brexit-specific dictionaries (based on Gadd 2019)the Irish border and Brexit Agreement? Not surprisingly, the Irish border topic is very concerned with Ireland and Northern Ireland in particular. The topic covering the actual Brexit agreement demonstrates the willingness to find a 'clear' and the 'best' possible deal.

Inductive modelling approach
As discussed in the methodology section, the keyATM allows for both deductive and inductive identification of topics in the corpus. After introducing the seven topics covering issues we expected as relevant for Brexit discourse, we now turn our attention to the three inductively defined topicstopics which were identified by the model as jointly co-occurring. Relying on the word clusters and the most representative speeches for each topic, we subsequently labelled topics as protection, the EU, and research (see Table 1). The protection topic is concerned with security and defence, but also human rights and values such as solidarity. The terms 'united', 'together', and 'friends' lead us to conclude that this topic deals not only with security but also with common values and unity and may reflect the use of diplomatic language as an (indirect) negotiation strategy. The second topic deals with EU politics generically, not focusing on Britain but rather on the Union itself and its member states. The research topic, finally, could be an indicator of the importance of international networks in academia and the innovation industry which could be affected by Brexit.

Topic prevalence across actors and time
Based on the EU, Irish and UK macro context, the British executives (H1),can be expected to focus on economic issues, as well as on regaining political sovereignty. Irish executives, we expect (H2), are predominantly concerned with the border. We also expect that the focus of EU executives will be on the Union itself as well as on the Irish border (H3). Finally, we expected changes in leadership and the triggering of Article 50 to affect the Brexit discourse, particularly on the key economic and political * marks predefined keywords, all terms lowercased as used in the modeling approach topics (H4). We turn our focus to these four expectations but also engage in an exploratory analysis of the identified topics which we discuss relating to the broader framing and crises management literature.
Looking into the distribution of topics by actors (Figure 2), we see that British, EU and Irish executives each emphasise different topics. As expected, British executives focus on economic topics (foreign trade and labour) but also emphasise international affairs, protection, and research more than their counterparts from the EU and Ireland. Foreign trade and reinforcing ties with the Commonwealth was an explicit focus of Leavers during the referendum campaign (Owen and Walter 2017), explaining the executives' speeches focus on it. For the Irish incumbents, immigration, the Irish border, labour, protection, and research were the topics to which they largely contributed. The EU executives did not concentrate their efforts too much on a single policy issue but rather focused on the EU itself, financial matters, the Irish border, and the normative aspect of the Brexit agreement. Overall, then, this provides supportive evidence for Hypotheses 1-3. However, some of our expectations were not met -Irish political elites discussed the topic of immigration more than their UK counterparts. The literature agrees that the politicisation of immigration was of quintessential importance, not just as a probable cause of the Brexit vote, but also as an important topic for the UK during negotiations. We look for a possible explanation for this by looking into what the British and Irish executives emphasise when discussing immigration.
We find two inductively defined topics, protection, and research, which certainly merit further attention. While the impacts of Brexit on research are very relevant for the academic community, the media has compared to other issues, rather disregarded this aspect (Gorman 2016). Considering the excellence of the UK universities in global comparison, the innovation and technological implications, it perhaps comes as no surprise that the UK has been a prominent contributor to this topic. The topic of protection, with a specific emphasis not just on security, but also on values, rights, togetherness, and the future, while initially puzzling, might refer to the executives' efforts to attempt to protect the British community from Brexit divisions and instead emphasise the need to unify the divided nation while maintaining a friendly EU relationship. Finally, the British executives discussed the normative aspect of the agreement less, overshadowing instead these issues with a strong focus on foreign trade. The border topic was somewhat underemphasised, potentially as a result of May's and Johnson's unwillingness to draw attention to it -May agreed to the backstop solution (Morris and Kibasi 2018) and Johnson signed a relatively unfavourable border deal (Freedland 2019).
Against this background, we now turn our attention to understanding the dynamics of Brexitrelated communication by political executives over time. For that purpose, we partitioned our period of analysis into four periods, based on the changes in the British leadership and the triggering of Article 50 (see Figure 1). Overall, on the one hand, we find quite large shifts in topic emphasis for the key political and economic topics -immigration, agreement, border, foreign trade, and finance, as well as for the topics of international affairs and research. On the other hand, we see more stability in the salience of labour, protection, and EU topics.
As expected, shifts in topic prevalence can be identified mostly in UK speeches. Immigration (with the focus on work aspect) is continuously becoming more prominent and so does research, particularly after the invocation of Article 50. Foreign trade becomes more important during May's mandate than in the six months preceding the Brexit referendum, but it weakens in relevance in the period post-Article 50, a similar trend to the one of the 'international affairs' topic. This might be a result of the inability to negotiate trade deals while still being a member of the EU (Dhingra, Ottaviano, and Sampson 2017). The EU's and Irish discourses show less over-time dynamic, but speeches related to the Irish border issue and the Brexit agreement topic are strongly affected by the activation of Article 50. In addition, the EU's emphasis on finance was drastically reduced in the period post-Article 50. Overall, these results suggest that our modelling strategy indeed allows an exploration of the Brexit discourse's temporal dynamic. Finally, we find some interesting shifts in topic prevalence between parties across time which almost resemble reactions of one actor to the changes in the discourse of the other -e.g. Irish executives decrease their emphasis on the border topic at the same time as the EU's team augments it. There is a similar dynamic going on also when discussing the topic of the Brexit agreement. We see this as a potential indication of how the negotiation process influenced the parties affected by Brexit.
So far, we focused on the ten topics which have dominated executives' speeches. We turn next to the qualitative exploration of how each of the three involved actors discussed these topics by analysing which terms they used when discussing the two key economic and political topics (see Table 1). Starting first with the economic topics, for the EU, the finance topic is discussed along the lines of the Euro project highlighting 'support' -emphasising the maintenance of the EU's legitimacy. For the UK, the topic of finance is rather about one of its key sectors, financial services. For Ireland, the finance topic seems rather generic, not giving us a distinct Irish agenda. In terms of foreign trade, then, the UK's interests here seem focused on future relationships with trading partners, indicating its efforts to strengthen its economic post-Brexit ties. EU interests, here, are focused on fairness, trade wars, and its relationship with China. Ireland, then, discusses the topic more in terms of opportunities and investments.
We focus next on the two most prevalent political topics, immigration, and border (see Table 2). Immigration in the UK and Irish speeches were discussed with an emphasis on the future, institutions, procedures, and development, whereas the EU's interests here seem more genuinely tied to the actual immigration topic and refugees, whereby the aspect of 'sharing' comes up prominently. While this interpretation needs further inspection of the relevant speeches, we might assume that this is related to sharing 'the burden' in terms of distributing refugees amongst all member states. Regarding the unanticipated emphasis that Irish executives place on the topic of immigration (which we rather expected from their British counterparts), we suggest this may be due to the way the issue has been discussed. Irish executives focus on specific programmes and activities relating to immigration, while the British tend to use a more diplomatic and rather neutral vocabulary. This may be because the UK is in the middle of negotiations to leave the EU, for which it needs to follow general diplomatic conventions, while Ireland focuses on the de facto implementation of immigration policies in its national context, potentially explaining the additional emphasis placed on this topic in the speeches. We provide this as a possible explanation, but future research should look more carefully into this aspect of Brexit-related speeches by executive members.
Most interestingly, then, the border topic draws a distinct picture for all three groups of executives. We see that the whole topic is heavily dominated by Ireland and Northern Ireland in particular. For Ireland, arguably a very affected country in this respect, peace is a key term, reflecting the fear of a hard border jeopardising the Good Friday Agreement. For the EU, the focus is on citizensconfirming our speculations of the EU's emphasis on its citizens in the UK in relation to border regimes, whereas UK speeches feature the term 'control', lending additional support to our expectations that executives' speeches would focus on sovereignty and regaining control over its territory.
Beyond substantive findings, we use this opportunity to discuss the usefulness of our methodological approach for understanding executives' communication. This analytical effort offers an insight into a large corpus of speeches, and the specific technique of key-word assisted topic modelling allows a combination of deductive and inductive analytical approaches. More specifically, we are able to model the data in a way that allows us to test assumptions derived from the literature, while simultaneously looking for topics that this literature may have so far overlooked. This has allowed us to test whether speeches reflect the strategic interests of the most affected parties, and how new events within a protracted crisis affect communication through political speeches. It has also pointed to two topics which flew under the radar of the media but were emphasised by political executives: research and protection. Our approach has aimed to understand what these topics mean across a large number of speeches. However, this approach does not deliver an in-depth understanding of the specific issues emphasised by each of the three parties. Future studies, with additional emphasis on human coding of speeches and qualitative analysis of representative speeches for each topic, would greatly contribute to a better understanding of what (and how) political executives communicate during protracted crisis periods. The following section provides a summary of the most important findings and outlines avenues for future research.

Conclusion
This paper set out to understand the direct communication agendas of political executives of the three parties most affected by Brexit -the EU, Ireland, and the UK. The key interests of these three parties have been considered very entrenched, with neither side seemingly willing to budge. Can these established key interests also be observed in political executives' communication? Using the keyword-assisted topic modelling approach, we explored in which ways top-ranked politicians decided to discuss Brexit in their public speeches. Our results offer four contributions to the evergrowing literature on the Brexit crisis and communication.
First, our results show that executive members mostly emphasised issues relating to vital interests of the parties they represent -thereby focusing on economic aspects, which indirectly permeated political topics as well. This is very much in line with the existing literature (Bulmer and Quaglia 2018), but we add to existing knowledge by presenting evidence that even political topics were interpreted through an economic lens. Second, confirming the existing insights from the literature (Baumgartner and Jones 2010), important milestones such as the triggering of Article 50 also had a measurable effect, particularly affecting communication about key topics such as finance, foreign trade and the border. Third, we find some evidence that political speeches at moments of crisis can be seen as an interface for the flow of the crisis negotiations. We see that progress in the negotiations can lead one party to decrease attention dedicated to a specific topic, and another to increase it. This is an important finding and certainly merits further analysis, as it shows that affected parties, while often appearing to be entrenched and irreconcilable in their positions (e.g. Frenhoff Larsén 2020), may have sought compromise regarding certain issues after all. Fourth, beyond the hypotheses we have formulated and tested, we have also set out to inductively research whether topics outside the media-driven Brexit discourse were represented in political executives' speeches. These inductively identified topics shed further light on the contents of executives' Brexit-related communication, illuminating the stakes of the negotiations beyond what is predominantly discussed in the media Beyond these substantive contributions to the literature on Brexit and framing, the article provides a methodological contribution. We demonstrated how a semi-supervised machine learning approach can be used to examine a large number of political speeches in a manner that allows both deductive and inductive research. On the methodological front, future research could certainly combine this quantitative and automated approach with a more in-depth qualitative analysis of the speeches most associated with each topic.
In conclusion, our paper facilitates a better understanding of executives' speeches as a means of crisis communication. Our approach has been to treat Brexit as a critical juncture with an unpredictable outcome, that continually puts the political resilience of the EU, Ireland and the UK to an unprecedented stress test. In this sense, our paper outlines not only how these actors chose to discuss the specificities of the Brexit crisis but also provides more general insights into communication during protracted, high-risk crisis periods. Note 1. In developing our expectations regarding Brexit negotiations, we rely on the ability of the Lexicoder Dictionaries, which have been widely tested and validated on various types of speeches (e.g. Albaugh et al. 2013) to identify policy fields in textual data. To augment the original dictionaries, we have further adjusted them to suit the European context. Therefore, we assume dictionaries are able to equally well identify all of the selected policy fields but note that our endeavour to formulate expectations and subsequently test them relies on the ability to identify and capture topics. As a result of this limitation, we establish only exploratory expectations.

Disclosure statement
No potential conflict of interest was reported by the author(s).

Funding
This work was supported by the Austrian Science Fund (FWF) [T-989 G27].