Parliamentary speeches in big data and natural language processing research

With the growing availability of digitized textual data, political science researchers seeking to assess policymakers' ideological positions can now draw on a wide range of empirical sources, from social media posts to electoral manifestos. Consequently, text data have also been widely used to analyze Italian politics. For instance, Ceron (Reference Ceron2024) employs a text analysis of Italian presidents' investiture speeches and television addresses to explore how Italian presidents' ideological leanings manifest themselves in these settings. Quantitative text analysis has also been employed to investigate Italian social media texts and have helped address a variety of different research questions. These include exploring levels of intra-party conflicts (Ceron, Reference Ceron2017), analyzing immigration-related discussions (De Rosa et al., Reference De Rosa, Bocci, Bonito and Salvati2021), and forecasting electoral outcomes (Caldarelli et al., Reference Caldarelli, Chessa, Pammolli, Pompa, Puliga, Riccaboni and Riotta2014). Additionally, scholars of Italian politics have employed newspaper corpora as a resource to analyze how policymakers adopted different policy narratives, for example during the Euro crisis (Bobba and Seddone, Reference Bobba and Seddone2018) or the COVID-19 pandemic (Crabu et al., Reference Crabu, Giardullo, Sciandra and Neresini2021).

In this context, researchers have shown growing interest in leveraging collections of parliamentary speeches to study political actors' policy preferences. Using parliamentary speeches makes good sense as parliaments are crucial venues for the formulation of policy agendas, as they constitute the formal institution where parties compete to fulfill their representative mandates in between elections and thus a forum in which policymakers can not only signal policy positions toward other parties, but also to the electorate as a whole (Vliegenthart et al., Reference Vliegenthart, Walgrave and Zicha2013; Proksch and Slapin, Reference Proksch and Slapin2015). However, researchers' ability to draw inferences on policymakers' positions on a range of political questions is predicated upon the ease with which it is possible to access and analyze these texts. Parliamentary speeches have proven to be a rather challenging source of textual data as parliamentary records are often not adequately digitized and data quality frequently deteriorates the further back in time one goes. As noted by Sebők et al. (Reference Sebők, Proksch, Rauh, Visnovitz, Balázs and Schwalbach2025) for several European democracies, there exist significant hurdles in obtaining data that can then be readily parsed and analyzed using commonly used programming languages. This means that researchers must invest significant efforts in scraping parliamentary debates from national parliaments' websites and even so, scraped files can often not be readily processed and analyzed because of poor data quality or inconsistent data infrastructures and naming conventions.

However, when accessible, legislative data from parliamentary corpora have become an increasingly valuable resource, widely applied in diverse studies utilizing various methodological designs. Thus, parliamentary corpora have been used to assess the salience of policy issues (Greene and Cross, Reference Greene and Cross2017; Cova and Schmitz, Reference Cova and Schmitz2024), conduct sentiment analyses (Proksch et al., Reference Proksch, Lowe, Wäckerle and Soroka2019), and investigate speech complexity in elite rhetoric (Osnabrügge et al., Reference Osnabrügge, Hobolt and Rodon2021), for example. Additionally, quantitative text analyses of legislative debates have explored how different parliamentary rules shape legislative behavior, both in comparative contexts and specifically for Italy (Giannetti and Pedrazzani, Reference Giannetti, Pedrazzani, Back, Debus and Fernandes2021). Researchers have also applied text analysis to corpora of parliamentary debates to investigate how the socio-demographic characteristics of parliamentarians influence the likelihood of certain policy themes being discussed. Numerous studies have thus illustrated the extent to which personal and socio-demographic factors, such as gender and socio-economic background, shape legislative behavior, including variations in speaking time and topic selection (Bäck and Debus, Reference Bäck and Debus2019; O'Grady, Reference O'Grady2019).

Considering the numerous research questions that can be addressed using parliamentary speeches as well as the different automated text analysis methods that can be applied, recent years have seen increased efforts to publish machine-readable collections of parliamentary speeches, which have taken the form of country-specific as well as cross-national projects. For instance, Remschel and Kroeber (Reference Remschel and Kroeber2022) released a comprehensive dataset on German parliamentary proceedings that includes not only parliamentary speeches but also other types of parliamentary data, such as bills, written responses, communications, requests, and replies. Beelen et al. (Reference Beelen, Thijm, Cochrane, Halvemaan, Hirst, Kimmins, Lijbrink, Marx, Naderi, Rheault and Polyanovsky2017) have harmonized and standardized Canadian parliamentary archives, dating back to the 19th century, into machine-readable formats. Recent years have also marked the emergence of comparative projects, which have sought to map out and harmonize digital infrastructures, including parliamentary corpora (Erjavec et al., Reference Erjavec, Ogrodniczuk, Osenova, Ljubešić, Simov, Pančur, Rudolf, Kopp, Barkarson, Steingrímsson, Çöltekin, de Does, Depuydt, Agnoloni, Venturi, Pérez, de Macedo, Navarretta, Luxardo, Coole, Rayson, Morkevičius, Krilavičius, Darǵis, Ring, van Heusden, Marx and Fišer2023). Most notably, the widely used ParlSpeech dataset (Rauh and Schwalbach, Reference Rauh and Schwalbach2020) contains full-text corpora of parliamentary speeches from various advanced democracies over the past two to three decades and has been extensively employed in comparative analyses of legislative activities and parliamentary behavior. However, Italy is not represented in the dataset.

As far as the Italian case is concerned, as part of the broader CLARIN project, which aims to collect machine-readable and annotated corpora of European countries' parliamentary proceedings, Agnoloni et al. (Reference Agnoloni, Bartolini, Frontini, Montemagni, Marchetti, Quochi, Ruisi and Venturi2022) have collected 79,000 speeches containing ca. 31 million words for the Senato for 2013–2020. While this constitutes an important achievement, the limited timespan offered by the database limits the possibility of conducting historical analyses. Within this context, it is also important to mention the Italian Legislative Speech Dataset: a historical collection (1946–2022) of investiture speeches delivered in Parliament during votes of confidence.Footnote ¹ The texts are segmented into quasi-sentences to facilitate classification of parliamentary interventions based on the salience of different policy themes. While this dataset is a valuable resource, its focus on investiture speeches restricts its scope, thus limiting opportunities to analyze a broader range of parliamentary speeches that are reflective of ordinary legislative interactions. Due to the challenges of obtaining machine-readable plenary speeches, researchers studying Italian parliamentary behavior and political practices have shifted their focus to other types of discourse recorded in parliament. To name a few examples, Cavalieri and Froio (Reference Cavalieri and Froio2022) examine the behavior of populist parties through parliamentary questions, while Salvati (Reference Salvati2021) analyzes prime ministers' speeches during votes of confidence from 1994 onward.

The dataset

The ItaParlCorpus dataset is a comprehensive, annotated, and machine-readable database of Italy's parliamentary speeches spanning from 1948 to 2022. It provides data on parliamentary plenary transcripts, which allows researchers to investigate a wide range of topics, including party positioning on various policy areas, elite rhetoric, and the salience of different issues within the Italian parliamentary context. The dataset as well as the codebook are open access and are freely available on the Harvard dataverse.Footnote ² In what follows, I briefly describe the construction of the database and the operationalization of the variables.

As discussed above, a significant hurdle for researchers studying historical parliamentary speeches is the poor quality of available material, which largely depends on existing digitization efforts and the availability of archival resources. In the Italian case, the Camera dei Deputati only offers images and scans for records prior to 1996. The further back one goes, the more unstructured do these files tend to be. The first necessary step is thus converting these scans and images into text files using optical character recognition (OCR) technology. This prior invaluable work was carried out by Frasnelli and Aprosio (Reference Frasnelli and Aprosio2024), who made the .txt files available on a GitHub repository. However, these text files are unstructured as illustrated by the example (Figure 1) and therefore are not readily utilizable by researchers interested in substantive political science questions. Moreover, a common challenge when using OCR technology for text files is ensuring the accurate conversion of scans into readable text. In the Appendix, I assess data quality by employing the Italian Hunspell spellchecker to identify improperly converted words. As illustrated in Figure A2, the proportion of misspelled words in parliamentary interventions remains relatively low, typically ranging between 1 and 3% across most years.

Figure 1. Original .txt files of parliamentary discussions on April 4, 1978.

The unstructured text files derived from the OCR scans present a challenge because they do not clearly separate the content of parliamentary speeches from the identities of the speakers. As illustrated in Figure 1, the speech content and the names of parliamentarians appear on the same line. Once the repository of these unstructured text files is downloaded, the task is thus to accurately differentiate between speakers' ID and the speech content that is associated to them. This distinction is essential for political science research, as correctly linking party affiliation to specific speeches is crucial for analyzing substantive policy discussions. As shown in Figure 1, parliamentarians are indicated in the text files using words which only contain capitalized letters. However, uncritically relying solely on words with capitalized letters to identify speakers is problematic because not all such words can be tied back to parliamentarians (e.g. acronyms or Roman numerals). Furthermore, inconsistent naming conventions add complexity: in some legislatures, only surnames are used, while in others, names may appear as “first name-surname” or “surname-first name”. Another complication is that not all parliamentary interventions are from members of the Camera dei Deputati. Invited speakers, such as ministers, technocrats without parliamentary seats, and members of the Senato, also intervene in plenary debates.

To determine the party affiliations of speakers in the text files, I utilized a list of Italian legislators from the Comparative Legislators Database (Göbel and Munzert, Reference Göbel and Munzert2022). This multinational dataset includes detailed information on national-level policymakers, such as their names, party affiliation as well as socio-demographic data. To ensure that I have accurately identified government members who do not serve in parliament, I supplemented this data by scraping the Italian government's website to collect the names of cabinet members across different governments.Footnote ³ Additionally, I incorporated data from prior data collections on Italian technocrats (Improta, Reference Improta2021). This comprehensive approach allowed me to compile a complete list of legislators, encompassing both parliamentarians and non-elected government officials, enabling precise separation of speakers from speech content. In order to address discrepancies between the text files and the compiled list of speakers, I conducted manual checks across legislative periods. Mismatches typically arose from differences in naming conventions, such as the use of middle names or maiden vs. married names in one source but not the other. The Appendix provides further information on how these manual checks were conducted.

The final output is a series of .csv files, in which every row corresponds to a parliamentary intervention/speech, recording the following information: the day in which the speech took place (date), the year (year), a document identifier (doc_id), a unique row identifier (row_id), the legislature (legislature), the speaker's name (speaker), the speaker's unique numerical identifier (pageid_wiki), which coincides with that used by the Comparative Legislators Database (Göbel and Munzert, Reference Göbel and Munzert2022), the party name (party_name), the party family (party_family) to which the party belongs to as recorded in the ParlGov Database (Döring and Manow, Reference Döring and Manow2024), the ParlGov unique numerical party identifier (party_id_parlgov), the ItaParlCorpus unique numerical party identifier (party_id_itaparl), a Boolean variable denoting whether the speaker is the chair (chair), another variable, which denotes whether the speaker is a cabinet member (cabinet) without being recorded as a member of the Camera dei Deputati (i.e. a member of the Senate or a technocrat), and finally the raw text (text). An example of the final output is shown in Figure 2.

Figure 2. Converted .csv files of parliamentary discussions on April 4, 1978.

Potential applications

The ItaParlCorpus database enables researchers to perform quantitative text analyses of Italy's parliamentary discourse, offering valuable insights for various subfields of political science and related disciplines. Beyond the raw text data, the database also provides contextual political information, such as party affiliation, which supports the development of more sophisticated research designs. Here, I present some brief examples to demonstrate potential applications of this new corpus of Italian parliamentary speeches. The analysis focuses on two particularly prominent and contentious topics in Italy's post-war republican history: abortion and the mafia. Specifically, it examines the salience of these topics in the parliamentary context, as well as the manner in which these issues have been debated by policymakers. To be clear, the purpose of this article is not to provide a substantive analysis of these well-studied topics, but rather to illustrate how this corpus can be used in ways that may be of interest to scholars of Italian politics.

In many European democracies, the struggle for women's reproductive rights, particularly access to safe and legal abortion, emerged as a pivotal civil rights issue in the latter half of the 20th century. In Italy, the introduction of abortion rights in 1978 through the Legge 194 marked a watershed moment, granting women the legal right to terminate a pregnancy within the first 12 weeks. At the time, this was a deeply contentious issue, influenced also significantly by the Catholic Church's prominent role in Italian political life. While left-wing parties and lay, liberal-centrist groups such as the Partito Liberale Italiano and the Partito Repubblicano Italiano supported the legislation, the Christian Democrats (DC) faced internal divisions, and the far-right Movimento Sociale Italiano opposed it.

A very different, but very politically salient issue which significantly shaped Italy's post-war history is that of the mafia. For much of this period, the mafia was a topic rarely addressed publicly by political elites, who often downplayed the infiltration of organized crime in the upper echelons of power. However, escalating public displays of violence, culminating in the high-profile assassinations of judges and politicians during the 1980s and 1990s, forced a shift in political discourse and policy action.

To explore the changing salience of these two key terms using the ItaParlCorpus database, I examine the share of parliamentary interventions discussing the terms “abortion” and “mafia” in the period 1948–1992 (Figure 3). I operationalize salience as the share of parliamentary interventions discussing these topics as a share of the total number of parliamentary interventions made by party groups. Abortion emerges as a highly salient topic, with nearly 10% of parliamentary interventions addressing this topic at its peak. Notably, the center-left and the PCI (Italian Communist Party) demonstrated significantly higher salience on this issue compared to the DC. In contrast, discussions of the mafia exhibited low salience during the early decades, gradually increasing in prominence throughout the 1980s and 1990s. While this analysis is indicative of the parliamentary salience of these terms and which political party emphasized these issues more than others, the next step is to examine the way in which parliamentary discourse has changed. In this article, I showcase two different analytical approaches.

Figure 3. Share of parliamentary interventions discussing abortion (left) or the mafia (right) as a share of all parliamentary interventions by party (1948–1992).

To analyze how discussions of the mafia evolved in parliamentary discourse, I conduct a descriptive analysis of the words most frequently employed in these debates. To do this, I extract sentences from parliamentary interventions that explicitly mention the mafia and conduct my analysis across four distinct historical periods: 1948–1959, 1960–1979, 1980–1999, and 2000–2022. As highlighted in Figures 4 and 5, one can observe notable temporal shifts in the most frequently used nouns and adjectives in sentences in which politicians discuss the mafia in parliament. During the initial period under analysis (1948–1959), parliamentary discussions frequently associated the mafia with banditry (banditismo) while also highlighting possible political connections. As the 1960s and 1970s unfolded, possibly mirroring a heightened awareness of the mafia, discussions across party lines increasingly incorporated terms such as “phenomenon,” “enquiry,” and “problem.” In the 1980s and 1990s, as the mafia emerged as a more prominent and widely discussed topic, there seems to have been greater alignment in the terminology employed by various political groups. This period also saw a growing emphasis on the term “public,” possibly reflecting increased awareness about the mafia's impact on society. Finally, in the 2000s, parliamentary language shifted toward a more legalistic framework, with terms like “criminality,” “laws,” and “criminal offense” becoming more prominent, signaling a possible focus on institutional and judicial responses to organized crime.

Figure 4. Most common nouns and adjectives used by political party families when discussing the mafia (1948–1979).

Figure 5. Most common nouns and adjectives used by political party families when discussing the mafia (1980–2022).

While the analysis presented above has relied on frequency-based methods, in recent years, political scientists have increasingly adopted more sophisticated text analysis techniques. Leveraging advances in natural language processing (NLP), such as word embeddings, researchers can in fact examine semantic relationships within large text corpora, capturing subtler connections and meanings that simple word counts or co-occurrence metrics might miss. As highlighted in the recent work of Rodriguez et al. (Reference Rodriguez, Spirling and Stewart2023), one specific application is the use of cosine similarity, which leverages word embeddings to identify the most distinctive words across different groups or categories.

In this context, a cosine similarity analysis of parliamentary interventions on abortion before and after the introduction of Law 194 in 1978 highlights the most distinctive terms of each period. Interestingly, as illustrated in Figure 6, parliamentary discourse prior to 1978 often emphasized topics and frames related to crime and morality. In contrast, post-1978 discussions shifted toward a medicalized vocabulary, focusing on terms such as contraceptives and sterilization, as well as broader bioethical issues like euthanasia.

Figure 6. Cosine similarity for speeches discussing abortion before and after the introduction of the Legge 194 (1978).