Further reading

1Challenges of processing parliamentary records, particularly regarding their acquisition, cleaning and annotating, are discussed in detail, e.g., for the Swiss Federal Assembly (Salamanca et al., 2024). The compilation of versions of the national corpora incorporated into the ParlaMint project is described for the Assembly of the Republic of Slovenia (Pančur & Erjavec, 2020), the Senate of the Italian Republic (Agnoloni et al., 2022), the Austrian Parliament (Wissik & Pirker, 2018), the Czech Chamber of Deputies (Kopp et al.,2021), the Parliament of Iceland, or Althingi (Steingrímsson et al., 2020), and the Vekhovna Rada of Ukraine (Kryvenko and Kopp, 2023).

Further reading

1Variation in theoretical perspectives on parliamentary records, key tensions in the making and editing of the official parliamentary transcripts as well as the process of producing the official parliamentary transcripts are discussed, e.g., with respect to Finnish (Voutilainen, 2023), British (Mollin, 2007; Slembrouck, 1992), or European and Japanese (Kawahara, 2024) parliaments.

Further reading

1Ilie (2017) outlines the differences between Westminster-model and European-model parliaments, describes the parliamentary discourse genre and its subgenres, and reviews key research topics on parliamentary debates. Taylor (2022) summarises recent applications of corpus linguistics in the study of political discourse, including parliamentary debates. Taylor & Del Fante (2020) offer a comparative perspective on corpus-assisted discourse studies and reflect on comparing the use of culturally loaded words, semantic groupings, discourse frames and rhetorical features across languages.

Corpus file download option

1ParlaMint files can be directly downloaded from the CLARIN.SI repository subject to Creative Commons - Attribution 4.0 International (CC BY 4.0) licence which allows free sharing and adaptation of data, but requires obligatory credit attribution. The files are available in different batches and several formats, namely as a TEI-encoded corpus, a linguistically marked-up corpus in several formats and a version machine-translated into English,⁵ all with added metadata files. The repository also offers a clear overview of different versions published throughout the years. In the case of direct download from the repository, the data can be processed by various desktop or online tools for data mining, like Orange ⁶ or Voyant.

Online tools with preloaded data – noSketch Engine and TEITOK

1For on-line analysis, the ParlaMint corpora are preloaded to noSketch Engine (Kilgarriff et al., 2014) and TEITOK ⁷ (Janssen, 2016), tools that enable open exploration of data.This digital textbook uses noSketch Engine in the showcases, so further details on how to use it are given in Part II. TEITOK, on the other hand, is not used in the showcases, so some general guidelines on how and when to use it are provided in this section.

Main TEITOK features

1The TEITOK platform allows for a reader-friendly visualisation of the speeches, which can also display text annotations – in the case of ParlaMint, e.g., marked-up named entities, like persons, organisations, etc., or transcriber’s notes, as shown in Figure 4. As both tools, noSketch Engine and TEITOK use the same data,⁹ the user can view this information in both. However, as mentioned, the presentation of extended context is more reader-friendly in TEITOK. Moreover, links to source documents and easy transition between different sessions and speeches inside a debate are only available in TEITOK.

2For each of the national ParlaMint corpora, TEITOK also provides information about the corpus compilation process, data sources used, characteristics of the national parliament, etc. which can be helpful for any users of this particular data in their study designs and interpretation of the results (see, for example, the description page of the Ukrainian parliamentary corpus ParlaMint-UA). This information is not available in noSketch Engine and it is advisable to check the TEITOK entry page for the selected ParlaMint corpus even if the rest of the analysis is planned in noSketch Engine.

3Although the TEITOK interface is less well adapted for computing and viewing advanced statistical overviews as in noSketch Engine, it provides some lists and statistics that can help the user navigate through extensive data. Once you select the desired national ParlaMint corpus and enter its description page, you can quickly display a list of all people and organisations included in the transcripts or browse through individual parliamentary proceedings one after the other, or by a particular metadata category, such as term, sitting date or body (see Figure 5).

4Such lists are available also for a subset of data obtained through a search query (e.g., for a particular word (at the bottom of the results page in TEITOK), a user can list all parties with information on how frequently they used this word; the search functionality will be presented further down). However, these lists are less informative compared to those produced in noSketch Engine which (1) allow to display also relative frequencies that are usually needed because the compared groups are rarely of the same size, and (2) the lists in noSketch Engine provide a direct link to display the actual speeches thus serving as a filtering mechanism to get the most relevant data for the analysis. So, given a more powerful and possibly more intuitive and transparent user interface, it seems preferable to use noSketch Engine for any statistical tasks, and TEITOK for clear visualisations of context and information that cannot be displayed in noSketch Engine, like details on corpus compilation/people/organisations or links to external sources which can help in the contextualisation of the results (see Figure 6 as an example).¹⁰

5TEITOK is also equipped with a relatively strong search interface¹¹ that allows building queries based on wordforms, linguistic annotation and/or metadata. The interface uses Corpus Query Language (CQL) to formulate queries, but for anyone not familiar with its syntax, it provides a query builder that facilitates query creation. As with statistical analyses, it seems that noSketch Engine has a somewhat easier and more powerful search interface than TEITOK. For this reason, we introduce the TEITOK interface only insofar as it helps locate the same section of a debate in both tools, thereby providing a more complete view of the context.

From noSketch Engine to TEITOK

1Interpretation of the results usually requires good understanding of the context which is dependent on the scope of the available context as well as its legibility. To exploit the visualisation advantages of TEITOK and analytical capabilities of noSketch Engine, we can search for the exact same debate section in TEITOK that we have found in noSketch Engine. Detailed instructions on using noSketch Engine are provided in Part II, so for illustrative purposes only, let us focus on occurrences of the word vriendschap in the Dutch parliament (ParlaMint-NL) and decide we want to get more context for the concordance line 5 (Ja, in alle vriendschap, dat begrijp ikare … by Martin van Rooijen). By clicking the light grey area on the left part of the line, we find, among other details, the date of the speech (speech.date = 2022-02-15). In TEITOK, we will use our query term (vriendschap) together with the speech date to quickly locate the debate in which the selected sentence was delivered. This is, of course, just one of the possible search approaches.

2To do this, we need to perform the following actions:

• Open the TEITOK ParlaMint landing page, select Subcorpora in the menu on the left and then click ParlaMint-NL.
• From there, we choose Corpus search in the menu on the left and expand the query builder option located to the right of the search box.
• On the left, there is a Text search box. One of the fields is Lemma. We input our search word in this field (see Figure 7).
• On the right, there are fields to define metadata. We locate the field Sitting date and input the date which we copied from noSketch Engine metadata (2022-02-15).
• We then click Create query after which the query is written down to the search field, and we can click Search.

• The tool returns two hits. Since we wanted to explore the context of the concordance line 5 in noSketch Engine (Ja, in alle vriendschap, dat begrijp ikare … by Martin van Rooijen), it is the second hit in TEITOK that we are looking for.
• To display context, we click context on the left side of the selected row (Figure 8), and we are presented with M. van Rooijen’s speech in the context of the entire debate. To navigate between pages, i.e., from the beginning to the end of the debate, scroll to the top and use the left and right arrows next to the page number.

Provenance information

1The layer of provenance information describes the methodology and process of corpus compilation, providing transparency about how and by whom the data was collected, organised, processed, as well as by whom and to what means the corpus can be used. For ParlaMint, this information is available:

• in the CLARIN.SI repository that serves as an archive and versioning manager for the ParlaMint corpora. The ParlaMint landing page in the repository gives useful provenance information. Figure 9 and Figure 10 highlight some of the key fields: a pre-formatted text for easy citation and licence information, description and link to additional documentation, overview and access to different versions, as well as links to corpus files (the relevant repository page is accessible also through the noSketch Engine corpus information page via the Corpus description & bibliography button on the left of the screen);

• in the TEITOK description page for ParlaMint and its individual parts as mentioned in Section Main TEITOK features
• in the scientific papers dedicated to ParlaMint compilation, namely Erjavec et al. 2023 and Erjavec et al. 2025

Structural markup

1The next layer of information consists of structural markup that captures the structural properties of the parliamentary proceedings, making it possible to analyse the documents at multiple levels of granularity. The structural markup is particularly important for corpus compilers in order to represent the internal structure of parliamentary proceedings in a consistent and machine-readable way. By encoding sessions, debates, speakers, and utterances, markup ensures that the corpus can be navigated, validated, and maintained across different parliaments and languages. For users, such structured markup is also highly valuable as it enables complex searches at different levels of textual granularity.

2In the case of ParlaMint, texts are encoded in TEI XML (Text Encoding Initiative XML), a widely used standard for representing textual data in a structured and machine-readable way. The encoding follows the ParlaMint schema which defines how the contents of the corpus, the actual speeches as well as all added details, are organised and named. Note that noSketch Engine uses data in the vertical format which is a derived, simplified representation of the TEI data, adapted for efficient loading and processing within the tool (see also Footnote¹²).

3The schema therefore not only defines the structural components into which the corpus texts are segmented (like sentences or paragraphs) but also provides a common predefined set of tags and their values that are filled with various information, like grammatical characteristics or metadata about speeches and speakers. Certain tags are constrained by closed typologies, e.g., speaker roles, political orientation, linguistic labels, while other tags accept open or unrestricted values, such as speaker names. A common schema like the one used in ParlaMint is a prerequisite for creating a comparable and interoperable set of corpora. To read more about the schema and the encoding process, see Erjavec & Pančur, 2022 , Erjavec et al., 2023 and Erjavec et al., 2025 .

Structural elements and tokenisation

1For each ParlaMint corpus, noSketch Engine provides a helpful information page (e.g., the information page for ParlaMint-PL which can be accessed from dashboard by clicking the Corpus info button or via the i-symbol at the top of the page) that outlines the main quantitative characteristics of the selected corpus, specifying its structural elements as well as listing the layers of linguistic annotation (see Section Linguistic annotation) and all metadata categories with its values (see Section Metadata). As an example, Figure 11 shows a part of this information page for ParlaMint-PL. We see frequency information that refer to all occurrences of a given structural element defined during structural markup. e.g., there are almost 230 thousand documents in ParlaMint-PL which refer to speeches and which are composed of more than 35 million words and some 10 million more tokens.

2In concordancers like noSketch Engine, which are intended for browsing and analysing the content of a linguistic corpus, the frequencies, when not specified otherwise, usually refer to tokens. A token represents a single occurrence of a running unit in the text between spaces, as defined by the concordancer’s tokenisation procedure. In the case of ParlaMint, tokens cover both words, which begin with a letter of the alphabet, and non-words, which include punctuation, numbers (see Figure 12) and tokens such as 3D or #SzczepimySię. This also explains the higher number of tokens in the corpus compared to words.

Linguistic annotation

1The next layer of information consists of linguistic annotations that outline linguistic properties of the parliamentary proceedings, making it possible to effectively analyse the language patterns.The linguistic annotation in ParlaMint includes lemmatisation , i.e., grouping all morphologically different forms of the same word to its base form, as well as annotation of word classes (also called parts-of-speech or PoS), morphosyntactic features and syntactic relations. The annotation of grammatical characteristics in ParlaMint follows commonly accepted tagsets based on the Universal Dependencies (UD) formalism . UD is an international framework for describing the structure of sentences in a consistent way across languages: it provides a standardised set of part-of-speech tags (verbs, nouns, adjectives, etc.), morphological features (number, case, tense, etc.), and syntactic relations (subject, object, modifier, etc.). This makes it possible to compare grammatical patterns not only within one corpus, but also across different languages and datasets.

2The annotation was performed automatically with language-specific models (see Erjavec et al., 2023 for detailed specifications). While these models were selected for their accuracy, users will still encounter errors in the annotations. This should not be seen as a discouragement from making use of the annotations, as even with some noise, they greatly enhance the possibilities for analysis and allow research questions that would be difficult or impossible to address by looking at raw text alone.

3Linguistic annotations are particularly important because they allow researchers to move beyond raw text and systematically explore patterns of language use in parliamentary debates, for example they can trace how political actors and social groups are framed grammatically by looking into the representation of the referents as subjects and grammatical objects, which in turn sheds light on how responsibility and power are distributed in discourse, or they can use grammatical information to study evaluative adjectives applied to different social and political groups revealing the attitudes towards those individuals., etc.

4Moreover, by exploiting linguistic annotation in queries, researchers can speed up the research process and avoid overlooking relevant material, e.g., searching by base forms instead of enumerating all possible word forms which can quickly lead to errors and missing entries, especially in morphologically rich languages. Annotations also make it possible to move beyond a predetermined set of keywords or phrases, opening the way for more explorative analyses, e.g., instead of only checking whether a specific term or phrase appears, researchers can examine how different grammatical constructions (mimicking the desired phrase) are distributed across the debates. In this way,linguistic annotations not only improve the precision and recall in corpus queries, but also expand the range of possible research questions beyond purely linguistic concerns, enabling scholars to investigate language use in order to reveal broader issues like political framing, identity construction, ideological positioning and the circulation of narratives.

Linguistic elements and lemmatisation

1In Figure 13, we see the list and frequencies of the elements ascribed during linguistic annotation.¹³ The numbers in this box refer to unique occurrences of a given element. This is illustrated well in the first line where we can observe that there are only around 320 thousand unique wordforms used in ParlaMint-PL, but all repetitions of the different word forms amount to the previously mentioned 35 million reported in the box represented in Figure 11 above. In fact, the full (non-repetitive) vocabulary used in ParlaMint-PL is actually smaller than the 320 thousand words just mentioned, since the lemma element tells us that there are only around 90 thousand unique units used in the corpus.

2A lemma refers to the canonical (dictionary) form of a word within a particular word class (part of speech) and, therefore, groups together all the different inflected forms of that word, disregarding their morphological variants used in different grammatical situations.¹⁴ Figure 14 shows the difference between wordforms and lemmas on the example of the verb głosować (to vote) from ParlaMint-PL.

Lowercase version

1To optimise corpus querying, corpora often include a lowercased version of the textual content. This means that all words which are otherwise written in uppercase letters, are given a lowercase equivalent. This facilitates more consistent searches, as users do not need to worry about differences in capitalisation (e.g., the Polish lower house Sejm vs. sejm), and ensures that case variation does not cause relevant results to be missed and at the same time avoid non-informative repetitions in word frequency lists. In Figure 13 above we can see two elements that refer to lowercased versions, namely lc which stands for lowercase word and lemma_lc which stands for lowercase lemma. Figure 15 provides an example sentence with annotations which show how the name Andrzej (the original word) is lemmatised as Andrzej, while the lowercased word and lemma equivalents are both andrzej. Note that the upper-case letters used at the beginning of sentences are often, although not necessarily, lowercased by default during lemmatisation, i.e. the process of ascribing lemmas to words, which is why both the lemma and the lowercased lemma of the original word Głos is głos. ¹⁵

Grammatical annotations

1Grammatical characteristics are encoded in ParlaMint on three levels. The box on the right in Figure 13 above encodes them as pos which stands for part-of-speech, feats which stands for morphosyntactic features and dep which stands for syntactic dependencies/relations. The frequencies given for a particular element refer to different forms of pos/feats/dep that appear in the corpus. As mentioned earlier, the grammatical information encoded in these elements can be useful in queries since they allow the researcher to search by a particular characteristic of a word/phrase rather than a list of actual words/phrases, and therefore to obtain more comprehensible results. However, apart from the pos categories which are only few and generally well known, e.g., verb, noun, adjective, adverb …, feats and dep values are much more numerous making it practically impossible to easily memorise all different characteristics that they encode. For this reason, it is best to always refer to the documentation of the tagset that was used in order to interpret or find a specific label.

2Corpus compilers may adopt different specifications/tagsets for linguistic annotation, and it is important for users to know which one was applied to correctly interpret the annotations and construct effective queries. The same linguistic feature can be referred to by different labels across frameworks, and these labels are exactly what queries operate on. In the case of ParlaMint, the compilers chose the Universal Dependencies (UD) formalism as the basis for annotation. The link to the tagset is provided also on the corpus information page in noSketch Engine. Figure 16 shows an example from ParlaMint-PL showing the three levels of annotations. In the first line, the subscripted labels indicate parts of speech (pos) such as NOUN and VERB, in the second line they mark morphosyntactic features (feats) such as Animacy=Inanimate, Case=Nominative, Gender=Masculine, Number=Singular, and in the third line they show syntactic relations (dep) such as nominal subject, object, root (referring to the main predicate here), and punctuation.

3➽ noSketch Engine tip: to display different levels of annotations, select the eye symbol (Figure 17) after you perform a search through the Concordance function (see Showcase II) or display the speeches/utterances through Text Type analysis (see Section Metadata).

Metadata

1The third layer of information can be considered metadata which adds another crucial level by describing the broader contextual properties of the texts. This information not only helps contextualise the results, but it can be used, similar to linguistic annotations, in queries to limit the context of the search task. The metadata in ParlaMint was drawn from diverse sources, e.g., official records, expert surveys and ParlaMint-specific expert annotations, and in some cases produced automatically with machine-learning algorithms which were trained and evaluated on manual expert annotations (the procedures are explained at relevant sections below). The potential errors in metadata, linked either to processing issues or source data inconsistencies, are regularly corrected, which is why it is important to always use the latest version of the corpus.¹⁶

Overview of metadata in noSketch Engine

1As with structural markup and linguistic annotations, noSketch Engine corpus information page provides a useful overview also in the case of metadata categories. Figure 18 shows the Text Type box of the ParlaMint-BE corpus information page in noSketch Engine which groups metadata categories according to the structural level at which they were assigned.

Sociodemographic and temporal information

1The speech level is the most numerous in terms of text type categories. It groups metadata on speakers and speeches as shown by a coarse division into red and blue boxes (for visualisation purposes only) in Figure 19. All speakers are assigned a unique ID and described with a set of sociodemographic characteristics, such as birth year, gender, and name,as well as certain professional characteristics. The latter include their status as MPs or ministers, their role in the debates (either as chairpersons or regular speakers), and their affiliation with a political party. Erjavec et al. (2025) note that official records and external sources like Wikipedia and alike were used to gather this type of information.

2Speeches are assigned unique IDs along with temporal information encoded in speech date, speech title, meeting title, session, sitting, and term (see Table 1 in The ParlaMint corpora for a temporal overview of the data included). Whether all of these categories are available depends on the organisation of parliamentary work and on national parliamentary traditions. For example, in ParlaMint-BE the proceedings are not structured according to sittings, which is why this information could not be included in the corpus. By contrast, in ParlaMint-PL the official records provide details about sittings, so this information could be added to the data. Moreover, speeches are also annotated with the languages identified in the transcripts. As expected, in most countries speeches are delivered in a single language, but in parliaments such as Belgium’s, proceedings may take place in more than one language. While some transcripts also contain agenda titles, these were not included in the noSketch Engine interface.

3➽ noSketch Engine tip: check the detailed statistics on sociodemographic information for ParlaMint-BE by going to this link and selecting the blue square with white stripes on the right side of the Text Types box beside the category of your choice (see the arrow in Figure 19).

4➽ noSketch Engine tip: understand and explore the detailed statistics. Figure 20 shows the distribution of MPs per gender in ParlaMint-BE according to the number of their speeches. By changing the option Show: Structure frequency right below the list of metadata categories, it is also possible to obtain the frequencies of tokens produced by gender. If you wish to get frequency information for groups defined by several parameters, e.g. by gender and party simultaneously, use the option analyse multiple text types written in blue on the left side of the screen. To check the speeches/utterances tagged with one of the metadata categories, e.g. only speeches by women (F), click the three dots next to frequency information at the bottom of the page.

5➽ noSketch Engine tip: once we display the concordance lines for a particular value (see the previous point), we can explore further the statistics for that set of data. To do this, the frequency button (blue square with white stripes, as in Figure 21) at the top of the page should be selected. From there, we have two options (Figure 22): we can either select the preset Text Types to show the statistics for all different metadata categories available for the selected subset of data, or we can define a list of one or multiple metadata categories at the Advanced tab which produces the statistics for a combination of metadata categories.

6➽ noSketch Engine tip: for advanced users, the noSketech Engine corpus information page is highly useful also as a cheat sheet in formulating complex queries with metadata parameters. Complex queries use a specific syntax, or Corpus Query Language (CQL), which requires the user wishing to include some metadata-based conditions into the query to specify the structure, the metadata category and the desired value in the format <structure category =”value”/>. To understand how to properly fill in the placeholders, the user can check the corpus information page to find out how the different levels are named. For example, to search for all speeches produced by women speakers on the topic of Culture (in ParlaMint-BE), the query would be: <speech speaker_gender="F" & topic="Cultuur"/>. On the corpus information page, the correct naming for the structure, metadata category and its values are provided in light grey and via the blue square box with white stripes. Note that the dot between the structure and the metadata category is replaced by a space in the query. See Showcase II for some examples of CQL queries.

Political organisation metadata

1In ParlaMint, the user can also get information about the parliament body/bodies included in the corpus, parliamentary status of MP’s party and party orientation. A given ParlaMint corpus includes transcripts from the main parliamentary bodies of a country. For unicameral parliaments, this means the debates of the single chamber. For bicameral parliaments, the corpus usually includes debates from the lower house and the upper house, or only one of them, depending on data availability and national practice. (The Belgian corpus also includes data from parliamentary committees). Erjavec et al. (2025) state the following:

2“This is important information for the comparability of the corpora, as it is sensible to compare the speeches of the same type of body, although most likely treating unicameral parliaments and lower house as the same type. Most corpora are either unicameral or contain lower house transcriptions only, while Great Britain, Netherlands, and Poland contain transcripts of both the upper and lower house. The Norwegian corpus contains labels for both unicameral, as well as for lower and upper houses because in 2009 Norway changed its parliamentary system from a (pseudo-)bicameral to a unicameral one. The only corpus containing only the transcripts of the upper house is the Italian one. The Belgian corpus is currently the only one in ParlaMint that also includes the sessions of various parliamentary committees.”

3➽ noSketch Engine tip: check the detailed statistics on parliamentary bodies for all corpora combined here

4A speech is also attributed information on the parliamentary status of the party that the MP delivering the speech belongs to. When using this information, it is important to understand that only the coalition status was consistently annotated with time-stamped tags across the ParlaMint corpora. However, for a large set of countries, the opposition status of parties was also tagged. There is also the nonvalue (-) which indicates that the speech either does not belong to an MP affiliated with a party being in the coalition or opposition at the time of speech delivery, or no time-stamped data is available.

5➽ noSketch Engine tip: check this link to observe the detailed statistics for party parliamentary status for all corpora combined

6Political orientation of parties was added to the corpus based on data from Chapel Hill Expert Survey Europe (CHESS; Jolly et al., 2022), Wikipedia and in some very infrequent cases, the knowledge of the corpus compilers. In the paper Meden et al. (2023), where you can find a more thorough description of the encoding process for political orientation, authors state that while the CHESS database was recognised as a highly valuable resource for this task, its main drawbacks for the use with ParlaMint data are (1) that it does not cover all ParlaMint countries, (2) the data is available only up to 2019, while ParlaMint data goes far beyond that point in time, and (3) not all parties included in ParlaMint are represented in the CHES database. For this reason, it was decided to add data from Wikipedia pages (and in very rare cases, also annotations produced by corpus compilers) as the second layer of the annotation of political orientation. Since the CHES database contains numerical representations of political orientation and there was no objective way to pair those with the categorial values used in Wikipedia, it was decided that only Wikipedia values will be represented in noSketch Engine.¹⁷ Their particularity lies in the fact that these are fixed values collected from Wikipedia in 2022 and therefore do not capture the potentially shifting political orientations of parties over time, as the CHES dataset does. Moreover, the fact that the values were collected in 2022 does not imply that they represent the parties’ orientations in that specific year (some parties had already been dissolved by then), but simply that the information was harvested at that time. For this kind of information, the user needs to consult the CHES values in the original corpus files retrievable from the CLARIN.SI repository page.

7➽ noSketch Engine tip: check this link to observe the detailed statistics for party orientation for all corpora combined

Integrated subcorpora

1ParlaMint comes with three prebuilt subcorpora which can be viewed at the level of text types in noSketch Engine. A subcorpus is a subset of data that was filtered out based on metadata and/or a search string containing particular words and linguistic tags. noSketch Engine deployment with log-in option, which is used in this digital textbook, allows to create user-defined subcorpora which is crucial in many analyses where the user is not interested in the entire corpus content but just a specific part of it. In addition to these user-defined subcorpora, ParlaMint offers three built-in subcorpora that divide the corpus into three periods and reflect the initial COVID-oriented focus of the ParlaMint project (see Section The ParlaMint project). The three subcorpora that can be found in all ParlaMint corpora are:

• Reference period covering all data before 30 January 2020, i.e. the pre-COVID period ending with the date of WHO announcing the Public Health Emergency of International Concern with regard to COVID-19 outbreak
• COVID period covering data from 31 January 2020 to 23 February 2022, ending on the day prior to Putin’s announcement of Russia’s full-scale invasion of Ukraine
• COVID/War covering data after 24 February 2024

Topic information

1The metadata category capturing the speech topic is based on an automated annotation using a schema that comprises 21 categories from the Comparative Agendas Project (CAP) ¹⁸, which classifies policy-making activities, along with two additional categories: Other and Mix (see Figure 23). The definitions used for the 21 topics are available in the Appendix of Ljubešić et al. (n.a.). The label Other covers speeches that do not address a policy issue, for example, when speakers discuss meeting logistics, share personal stories, or engage in arguments, while the label Mix is applied to cases for which the prediction confidence score of the model was below 60%.

2The annotation procedure involved several steps. First, samples were extracted from each ParlaMint dataset to enable pre-training of the XLM-RoBERTa-large model on parliamentary discourse, resulting in the BERT-like model XLM-R-Parla. Next, automatic annotation was performed with a Generative Pretrained Transformer (GPT) model using a prompt and adapted CAP topic definitions (see the Appendix in Ljubešić et al., n.a.). Manual annotation of a test dataset was conducted to evaluate the model’s output, confirming that its annotations reached human-level quality (Kuzman & Ljubešić, 2025). To improve topic coverage, infrequent instances (e.g., for the Public Lands topic) were added to the training dataset, and the model was re-trained, leading to the improved version XLM-R-Parla-impr, published as the ParlaCAP-Topic-Classifier.

3➽ noSketch Engine tip: check this link to observe the detailed statistics of topic distribution for all corpora combined

Sentiment information

1Moving from the speech to the sentence level, the corpus provides metadata on sentence ID and sentiment (senti label), ¹⁹ provided on three different scales but reflecting a single annotation task (Figure 24):

• senti_6 gives information on 6 levels and follows the original schema (see Table 2)
• senti_3 gives information on 3 levels and simplifies the 6-level schema
• senti_n gives the numerical output of the model which returns an estimated sentiment score on a continuous scale²⁰ which is then translated to the 6-level or 3-level scale

2The procedure involved the manual annotation of a training set using the six-level sentiment schema presented in Table 2. These annotations were then used to train the XLM-R-parla model, designed to predict sentiment scores on the same scale. XLM-R-parla is a domain-specific adaptation of XLM-RoBERTa-large, pretrained on parliamentary data. Building on this foundation, the ParlaSent model was developed, offering a deep learning approach to multilingual sentiment analysis within the parliamentary domain.

3To properly use and interpret the annotations, it is important to remember that in ParlaMint, sentiment is annotated at the sentence level, meaning that each sentence receives a label reflecting its overall polarity. The decision is based on the overall meaning of the sentence and the linguistic, rhetorical and typographical cues within the sentence itself, such as polarity markers (good, bad), intensifiers (enormous), diminutives, irony, metaphor, emoji, punctuation, etc. Importantly, context from surrounding sentences is not taken into account, so labels are assigned independently of the broader discourse.

4For this reason, the interpretation of sentiment labels should be approached cautiously. A sentence-level label does not necessarily indicate the attitude toward the sentence referents (since multiple entities may appear in the same sentence), nor does it capture the overall sentiment of an entire speech or debate (see also Showcase I and Showcase II). This is particularly relevant when formulating research questions about political attitudes, where more targeted methods looking only at the argumentation pertaining to a particular subject are necessary.

5Nevertheless, sentiment annotation remains useful for other kinds of analysis. It can help trace broad shifts in tone over time, compare variation in rhetorical style across speakers or parties, or examine whether certain policy areas are typically framed in more positive or negative terms. In this sense, sentiment tagging as realised in ParlaMint is often more informative about how things are said, e.g. emotionally charged vs. neutral style, than about what specific stance is taken toward an issue.

6➽ noSketch Engine tip: check the following links to observe the detailed statistics at 3-level scale / 6-level scale / n-level scale ²¹ for all corpora combined.

Transcriber notes

1The formal transcripts often include transcriber notes which include anything from voting results or timing and procedural information to information about different vocalisations in the room, like applause or interruptions, or remarks about inaudibility of the speaker, etc.

2Transcriber notes vary greatly between parliaments which is why there is, for the moment (ParlaMint 5.0, September 2025), no common categorisation proposed for this type of metadata. In ParlaMint (Figure 25), they are encoded as a separate structure that can be divided into up to two categories, namely type (the kind of action described) and content (a specification of that action), e.g., vocal:murmuring:Murmures sur les bancs du groupe LR. When displaying the values for type (by clicking the blue square with white stripes), we see that the type can be further specified as incident (non-communicative phenomena like action), kinesic (non-vocalised communicative phenomena like applause) and vocal (non-lexical vocalised phenomena like murmuring). Although the three specific types of notes were proposed as common labels in ParlaMint, there was no harmonisation of transcribers notes conducted between the corpora, which means that each corpora needs to be examined separately to understand what the notes contain and how they can be useful for the analysis.

3➽ noSketch Engine tip: check this link to observe the detailed statistics of note type in ParlaMint-FR. Switch to note content at the bottom of the list on the left side to display the statistics for that category.

4➽ noSketch Engine tip: to observe note type and content together, we can first display the speeches containing the notes by clicking the three dots beside a selected note type from the ParlaMint-FR Text Type analysis page. Once we are redirected to the list of utterances/concordances, there are two options. We can observe the notes in context, by selecting the View options button at the top of the page and then enable note (under Show structures) and the option Show more context (Figure 26). It is also possible to list all different combinations of note type and content by clicking the Frequency analysis button at the top of the page, going to the Advanced tab (Figure 27) and defining two metadata categories of interest, note type (as the first one) and note content (as the second one).

Named entities

1Named entity recognition / annotation (NER) enables researchers to identify references to various unique (individual or collective) entities in the data and analyse, for example, potential networks, centres of special interest or power, cross-references between debates, etc. As shown in Figure 28, named entities are encoded in ParlaMint as a separate structure retaining one metadata category, i.e. type, with the four standard values (Tjong Kim Sang & De Meulder, 2003) – organisation (ORG), person (PER), location (LOC) and miscellaneous (MISC ) where the latter refers to any named entity that does not belong to any of the previous three classes, such as Dispatch Box, Nationality and Borders Act or Shakespearean/Wagnerian (examples taken from ParlaMint-GB).

2➽ noSketch Engine tip: check this link to observe the detailed statistics of named entities distribution for all corpora combined

3➽ noSketch Engine tip: to list all different occurrences of a NER type, we need to produce a frequency list of these occurrences. One of the simple ways to do it, is to first go to the Text Types analysis page of the selected corpus (in this case, ParlaMint-AT) and click the three dots beside the selected NER value at the bottom of the page (for example, LOC). To obtain a non-repetitive list of all the locations in the data in focus, we click the Frequency analysis button and then, in the middle column under KWIC (Key Word In Context) select the option lemma (Figure 29). The same steps would be followed also if we want to observe only speeches with a more limited context and, therefore, we will first produce the concordance lines through the search option (Concordance) and then follow the steps just described. To display the annotations in context, use the View options button and check name under Show structures (see also Section Transcriber notes above).

Semantic annotation

1The machine-translated version of ParlaMint into English (Kuzman et al., 2025) is annotated with semantic tags, meaning that each word in the translated corpora is assigneda label describing its domain of meaning. In ParlaMint, these domains are drawn from the UCREL Semantic Analysis System – USAS (Rayson et al., 2004) which is composed of a closed set of 21 major top-level categories with several subcategories (Figure 30). The system covers both single words and multiword expressions and includes methods for contextual disambiguation. These methods combine part-of-speech tagging (to narrow down which semantic tags are relevant), general likelihood ranking, and rule-based heuristics, all of which increase the chances that the correct tag is assigned to a word in context.

2In noSketch Engine, the English machine-translated version of the corpus is called ParlaMint-xx-en and includes semantic annotations on a separate structural level under three metadata categories (Figure 31):

• USAS categories ²²: the numerical values that correspond to the entries in the USAS tagset²³ as specified in ParlaMint USAS taxonomy, e.g., Q4.3;
• USAS glosses: the numerical values together with their gloss/category name, e.g., Q4.3: The Media: TV, Radio and Cinema;
• USAS tags: the original output of the automatic annotation.

3USAS categories and glosses indicate the tag which was computationally determined as the most appropriate for a word or multi-word expression in a given context. USAS tag, on the other hand, shows all lexically possible tags for a given unit mapped directly from the USAS lexicon containing words and multi-word expressions with corresponding tags. When a unit was assigned several tags, e.g., parents: S4mf,A9+, the first tag (S4) was considered contextually the most suitable one (and encoded at the level of USAS categories/glosses).

4In interpreting semantic annotation in ParlaMint, it is important to keep in mind that the annotations are performed based on a lexicon of single words and multi-word expressions²⁴ that are labelled with appropriate semantic tags. The tagger matches the entries from the corpus with those in the lexicon and maps the corresponding semantic tags. The user might come across the following situations (Figure 32):

• when a unit is recognised as a multi-word expression (because it was defined as such in the lexicon), all its constituent parts are given the same, possibly compound tag, e.g., world war two where the constituent parts are interpreted as geographical names (Z2) or many years, tagged as P1 S2, where each part of the multi-word expression is given the tags of both parts (many: P1 + years: S2);
• when a unit is a single word or not recognised as a multi-word expression (even when effectively pointing to the same referent but not in a format recorded in the lexicon), each unit is given its own tag, e.g., the second world war which was tagged with the domains grammatical bin (Z5), linear order (N4), universe (W1) and warfare, defence and the army; weapons (G3);
• whenever there are several possibilities for a given unit, the first one is selected as the most appropriate one at the level of USAS categories/glosses, e.g., war which is tagged as G3,S7.3 in the initial output (USAS tag) keeps only tag G3 at the level of USAS categories/glosses.

5Please note that semantic tagging is a highly complex computational task, and even human annotators often differ in their judgments about the most appropriate semantic tag for a given word. For this reason, users are expected to familiarise themselves with the underlying taxonomy, i.e., set of possible categories, and adopt a flexible approach when interpreting the labels. Semantic tags are best understood as tools for gaining an initial overview of the data and identifying relevant samples for qualitative analysis, rather than as definitive annotations in their own right.

6Semantic tagging can be used to identify patterns in meaning across large corpora. However, as mentioned in an earlier section, ParlaMint also includes topic annotations (see Section Topic information), which can be considered another form of semantic annotation. To make best use of both layers, it is good to consider the following:

• Semantic annotations are only available for the machine-translated English version of ParlaMint, while topic annotations are provided for all ParlaMint corpora (both for corpora in the original language and for their translated versions).
• Semantic tagging is applied at the word level and could thus be considered actually part of the linguistic annotation layer, whereas topic annotation is applied at the speech level.
• The two annotation layers differ in their schemas. Semantic tagging relies on the USAS taxonomy, which comprises 21 broad domains of human experience and knowledge (Figure 30), while topic annotation is based on the Comparative Agendas Project (CAP) schema, which is specifically designed to classify policy agendas and parliamentary topics (Figure 23).
• Although both annotations are automatic and rely on manually created knowledge, they use different generations of computational models. Semantic tagging is performed with pymusas (Python Multilingual UCREL Semantic Analysis System), a rule-based tagger that combines lexicon lookups with spaCy components for tokenisation and handling of multiword expressions. Topic annotation, on the other hand, is performed with a large multilingual transformer model. Specifically, it uses the XLM-R-Parla model, a BERT-like model that has been additionally pre-trained on parliamentary proceedings. The speeches were then automatically classified into CAP categories using the GPT-4o model (see Section Topic information).

7Semantic tagging is sometimes used to extract the topics of speeches, but given the schema used in ParlaMint, this is not the most efficient application of this annotation layer – especially since a dedicated topic-annotation layer already exists for the corpus. Instead, the two layers should be seen as complementary. Semantic tagging provides a bottom-up perspective on how words are used (tags are assigned at the word level), making it particularly valuable for linguistic and discourse analysis. By contrast, topic annotation offers a top-down perspective on what debates are about (tags are assigned at the speech level), which is more suitable for political science research and related fields.

8Together, they enable researchers to study not only what parliaments talk about, but also how they talk about it. For instance, semantic annotations can be used to analyse the use of emotional language across debates or to track how economic vocabulary is distributed over time and across the topics identified by the topic-annotation layer. This, in turn, allows us to understand how much parliamentary attention is devoted to areas such as healthcare or infrastructure, for example. Semantic tags also make it possible to compare rhetorical styles, e.g., whether debates rely more on emotional or technical vocabulary, to measure shifts in framing, e.g., whether migration is discussed in terms of economy or security, and to conduct cross-national comparisons, e.g., whether different parliaments emphasise similar or different semantic domains when addressing the same policy issue.

9➽ noSketch Engine tip: check this link to observe the detailed statistics for semantic domains for the English machine-translated version of ParlaMint corpora