The following paper examines the application of graph technologies to parliamentary data. Using the debates that took place in the Croatian Parliament between 2003 and 2017 as an example, it demonstrates how natural language processing (NLP) tools, graph databases, and network algorithms can be used to conduct corpus statistical, stylometric, and semantic analysis. Special attention will be paid to the structure of morpho-syntactically tagged corpora, which are embedded in a property graph database that enables the exploration of corpus-specific semantic relations. As will be shown, such a structure allows parliamentary data to be empirically analyzed with regard to the communication, conceptualization, and framing of social identities, interactions, institutions, and cultural models.
This paper examines the application of graph technologies to parliamentary data by means of natural language processing (NLP) techniques, the graph database Neo4j, and a Python implementation of igraph algorithms. It describes the embedding of a dependency-tagged corpus in a graph database model, and presents an innovative approach to the empirical linguistic analysis of the social identities, interactions, institutions, and cultural practices recorded in parliamentary data. The goal of the project is to establish an extensive knowledge base with a clearly structured ontology, which in turn enables data integration, data enrichment, and quantitative-qualitative analysis.
In addition to the application of standard corpus statistical methods,
semantic and stylometric analysis of the syntactic dependency structures
was carried out using the Universal Dependencies NLP parser
By way of a case study, this paper considers records detailing Croatian parliamentary debates between 2003 and 2017. The following section (Section 2) discusses parliamentary data in general, while Section 3 elaborates on the genesis of the Croatian Parliamentary Corpus (CroParl). Section 4 presents a number of specific examples drawn from the said corpus, whereas the fifth and final section contains concluding remarks and some thoughts on the work that lies ahead. 2
Parliamentary data constitutes a rich and publicly available resource that is inherently endowed with politically and historically significant information concerning the discourse between the political representatives of democratic systems of government and the socio-cultural interactions in which they participate.
In recent years, the improved accessibility of parliamentary data has made
the democratic process increasingly transparent
Yet the proper archiving, structuring, synchronization, and visualization of the treasure trove of multimodal data that is generated over the course of the socially and institutionally highly complex interactions that characterize parliamentary debates are not without their pitfalls. For one, the diversity of the data management approaches of national and regional parliaments with respect to issues such as language, rhetorical strategies, actor properties, social features, and political context poses a significant obstacle to systematic scientific inquiry.
Such standards for the processing of textual data and the creation of
structured corpora of parliamentary and other governmental debates as currently
exist – along with a broad variety of metadata formalizations – are the result
of the eofrts of a small number of scholars based out of various European
research centers. At the European level, the Digital Corpus of the European
Parliament (DCEP)
Keeping these various factors in mind, the goal of this paper is to showcase a flexible information framework based on graph technologies that is capable of processing and integrating regional, national, and European parliamentary data. In so doing, I hope to demonstrate how current advances in natural language processing and data management can be harnessed to build a socio-linguistic parliamentary data network that is accessible not only to specialized scholars, but also to journalists, NGOs, and private citizens.
The working prototype for the web implementation of Croatian parliamentary data is available on the website hosted by the University of Rijeka’s EmoCNet project.3 1https://www.clarin.eu 2https://github.com/clarin-eric/parla-clarin 3http://emocnet.uniri.hr/croparl
Croatian Parliamentary Corpus A significant portion of parliamentary data is made up of texts that detail political debates among representatives. This corpus of speech acts allows researchers, particularly linguists and political scientists, to engage in diefrent kinds of semantic and pragmatic analysis.
Here I am concerned with the debates that took place in the Croatian
Parliament between 2003 and 2017, during the fifth to ninth parliamentary
assemblies. As can be seen in Figure 1, the process of corpus creation in this
instance consisted of data gathering, NLP parsing, data integration, and data
management in the property graph database
The data was gathered using a Selenium scraper4 from the Croatian Parliament web repository.5 Data pertaining to the debates that took place during the fifth to the ninth parliamentary assemblies is to be found in two separate datasets: one containing values that relate to the assemblies, sessions, and topics, and another containing values that concern persons, debate transcripts, topic IDs, and announcement metadata. 3.2
Extracted from 390,000 transcripts, the various speech acts of the
representatives in question were processed using the UDPipe Natural Language
Processing Toolkit
The NLP parsing created over 4 million sentences and 70 million word tokens in the CoNLL-U output format (Figure 2). The morpho-syntactic metadata appears in 10 tab-separated value fields: 1. ID: word index 2. FORM: word form or punctuation symbol 3. LEMMA: lemma or stem of word form 4. UPOS: universal part-of-speech tag 5. XPOS: language-specific part-of-speech tag 6. FEATS: list of morphological features from the universal feature inventory 4https://github.com/ropensci/RSelenium 5The data gathering process is published under the GitHub handle https://github.com/ rodik/Sabor. The Croation Parliament web repository can be accessed via http://edoc.sabor.hr/ 6https://lindat.mff.cuni.cz/repository/xmlui/handle/11234/1-2364
3.3
The data obtained from the Croatian Parliament was then integrated with
the CoNLL-U formatted results
Structurally complex entities are metonymically related to ontologically less complex entities: Assembly can be narrowed down to Session, which can be narrowed down to Topic, which can be narrowed down to Utterance, which can be narrowed down to Sentence, which can be narrowed down to Token. Utterances are related to a Representative, who is socially related to a Party and a Parliamentary Club. Some labels have inter-structural relationships: tokens, sentences, and utterances are sequential in nature, while tokens also have syntactic dependency relations.7
Arranging the data in question in a property graph has a number of
advantages. First, it enables the integration of various datasets into a single
framework. Second, labels, nodes, relations, and their properties can be
easily updated or remodeled according to newly adopted standards, and the
structure of the graph can be enriched with additional knowledge resources.
Last but not least, the user-friendly graph representation of the information
ontology enables digital humanities scholars to intuitively develop new
approaches to their material based on the interrelation of data structures.8
7For details concerning the data integration and NLP parsing process, see
8The tagged CroParl graph database’s Neo4j dump is a case in point, https://drive.google. com/file/d/1zRy3EmwPrb4r3vGM3J5N5bpeKwCLEaHF/view?usp=sharing 4.1
Statistical Summarization of Lexical Usage The CroParl corpus contains data from five parliamentary assemblies and covers 5,599 topics, which were broached by 895 members of parliament belonging to 42 political parties. In the process of NLP parsing, 390,078 utterances were related to 4 million sentences and 70 million tokens. 4.1.1
The graph-based lexical store thus enables standard lexical summarizations based on tokens, lemmas, or part-of-speech counts, and their respective proportions. For instance, Table 1 represents the lemma count for nouns in the corpus.
The corpus can also be used for the representation of a keyword in context (KWIC). The KWIC solution oefred on the oficial site (https://edoc.sabor.hr/) relies on a string-based query search. This type of search is relatively easy to implement from a technical standpoint, and is suitable for morphologically lean languages. However, this is not the case for Croatian, which has a complex declination and inflection system. Consequently, a lemma-based query yields much more accurate results in the KWIC concordance. Table 2 presents an example of sentence level results for the noun ljubav (‘love’) using a lemma-based KWIC query. The results are based on the lemma=‘ljubav’ and part-of-speech pos=‘noun’ properties assigned to the tokens. 4.1.2
Stylometric summarization based on the linguistic behaviour of
parliamentary representatives can provide us with valuable insights into their individual
linguistic profiles, stylistic idiosyncrasies, and similarities compared to other
speakers
The number of utterances of a speaker, for example, corresponds fairly directly to their prominence within the parliament. Table 3 shows a list of the 15 representatives with the highest number of utterances in the CroParl corpus, with the overwhelming majority belonging to Vladimir Šeks, who served as President of the Croatian Parliament from 2003 to 2008. A quick examination of Table 3 reveals that the two runners-up, Bebić and Leko, also served in this role, which, by its formal nature, involves numerous speech acts with low token counts per utterance.
Prominence can also be measured by other lexical features, such as lemmas per representative, part-of-speech count per representative, etc., and the data management structure enables various types of filtering (e.g. utterances/tokens/lemmas per date/assembly/session/topic).
Another application of corpus-based stylometric analysis is the frequency with which a certain lexeme is used by a specific representative. For instance, we might be interested in how often emotionally charged lexemes were used, and by whom. By way of example, Tables 4 and 5 show a portion of the data for two such terms: love and peace.
Here we see that the lexeme mir (‘peace’) was used almost five times more often than the lexeme ljubav (‘love’). Column “F” represents the frequency with which the respective lexeme occurred in the representative’s utterances, column “p in All” represents the overall proportion of the lemma in the corpus, column “Auth” the number of tokens per representative, and column “p in Auth” the proportion in which the lemma occurred in speeches by the representative. While the value “p in All” can be interpreted as an indicator for the representative’s conceptual influence on the debates, the value “p in Auth” represents the importance assigned to a particular concept by a given speaker.
This type of analysis opens up a whole range of possibilities when it comes to further studies concerning the significance of specific concepts to individual representatives. Stylometric linguistic analysis could also be used to interrogate the socio-, pragma-, and cross-linguistic profiles of a given political party or parliamentary club based on how often and in which contexts lexemes such as ‘love’ and ‘peace’ are used by its members. 4.2
Dependency-Based Semantic Network Analysis The NLP parsing of the texts in question created 74,968,809 syntactic dependency relations between tokens. The structure of the relations tagged by the Universal Dependencies parser is shown in Table 6.
The graph structure of these relations enables a dependency-based type of semantic analysis that is conducive to a number of NLP tasks, including the discovery of corpus-specific word senses. 4.2.1
Semantic Domain Induction Based on “Conj” Dependency
Word sense disambiguation and word sense identification have been
important tasks from the earliest days of natural language processing research
In the method being outlined here, the iterating graph algorithm is used to produce an undirected graph from all the nouns collocated with the construction dependency, operating as a kind of lexical embedding that represents the semantic structure of the concepts found in the Croatian Parliamentary corpus. The coordination dependency graph can be used to identify ontologically similar lexemes and related semantic domains for a given source lexeme: • choose a source lexeme and identify n number of the most frequently collocated target lexemes • construct a lexical network with associated lexemes as nodes and coordination-based weighted relations • detect lexically coherent communities with sub-graphs as a representation of the semantic domains, and use parametrizable granularity as a measure of the level of categorical consistency vs continuity and abstraction • analyze the distribution of prototypical association patterns as an indication of cross-cultural framing and cross-linguistic variations Graph analysis, meanwhile, is performed with the help of the Python implementation of igraph, which is used to identify, measure, and visualize the conceptual framing. The procedure starts with extracting collocations of an arbitrary source lexeme. For instance, the first 50 most common collocates for the lexeme mir (‘peace’) are represented in Figure 4.
For the source lexeme mir with n = 50 first order lexemes, the structural function of a friend coordination network is enhanced by identifying n=50 second order lexemes in a friend-of-a-friend (FoF) pattern. By analyzing the subgraphs created by the second order coordinated dependency noun collocates we can distinguish the semantic clusters that indicate the sense association of the source lexeme.
The FoF network with 945 nodes was pruned with regards to the node
degree measure degree>4 in order to filter out the less interconnected
lexemes from the semantic network. As can be seen in Figure 5, the resulting
FoF network for mir contained 120 nodes, which were then clustered using
the Leiden clustering algorithm
Finally, the prominence of the associative conceptual profiling of the source lexeme mir can be discerned from a speaker’s weighted contribution to the coordination dependency graph. The weighted graph was constructed using the collocates that are in a coordination dependency with the source lexeme mir used by each representative.
The resulting graph with 461 nodes was pruned to highlight the more prominent associations using weighted degree > 4 settings. The weighted degree represents the sum of weights assigned to the node’s connections. By ifltering out the nodes by weighted degree, we were able to focus on the lexemes and representatives that saliently contribute to the association matrix of the lexeme mir by their occurrence. As is shown in Figure 6, the pruned associative network with 97 elements was then clustered using the cpm resolution 0.34, which yielded 50 clusters. The 12 most prominent association clusters of the lexeme mir can be found in Table 8.
Clearly, then, graph measures and network representation of socio-lexical phenomenon can be used to glean an empirical, yet intuitive, understanding of the complex weighted structural relations between lexical content and the agents who generate this content. Of particular interest here is the question of which lexemes exhibit the highest number of associations, and who introduced these associations. Figure 6 shows that stabilnost (‘stability’), sigurnost (‘safety’), and red (‘order’) are central for understanding the associative conceptual content of the lexeme mir. The clusters, represented with diefrent colors in Figure 6 and transcribed in Table 8, reveal the agents who most frequently made use of the lexemes in question. Interestingly, most of the agents connected with the concepts ‘safety’ and ‘order’ have served as members of institutions responsible for the nation’s defense. For instance, both Ivan Šantek and Tomislav Čuljak were members of the committee on Internal Policy and National Security. Šantek was also a member of the Defense Committee from 2013, while Berislav Rončević served as its vice chairman.
Conversely, the concept ‘peace’ has a diefrent connotation when coupled with the lexemes ‘stability’ and ‘cooperation’ – the agents in this cluster seem to be more concerned with international relations. Jozo Radoš, for example, was vice-chairman of the Committee on European Aafirs and a member of the Committee on European Integration, while Davor Ivo Stier was a member of the Committee on Interparliamentary Cooperation, the Foreign Policy Committee, and the Committee on European Aafirs. The same holds true for Stier’s fellow part member, Andrej Plenković, who would go on to become prime minister. Similarly, the aspect of ‘prosperity’ in connection with ‘peace’ was promoted with particular vigor by the deputy chairman of the Committee on Croats outside the Republic of Croatia, Boro Grubišić, and by a member of the Committee on Regional Development and European Union Funds, Petar Baranović. The antonym ‘war,’ meanwhile, was prominently associated with ‘peace’ by the president of the Delegation of the Croatian Parliament to the NATO Parliamentary Assembly, Krešimir Ćosić.
Although these results were acquired from a complex set of queries and graph algorithms, they intuitively represent the conceptual associative dimension of the lexeme mir, and as such can help us to begin to understand how conceptual relations are influenced and motivated by the political views of the speakers in question and their respective institutional functions. The possibility of such an insight demonstrates the true value of graph-based socio-linguistically and morpho-syntactically tagged corpus analysis: it allows us not only to simply count lexical items or allocate frequencies of syntactical relations, but to explore the socio-cognitive aspects of conceptualization by revealing a dynamic pragmatic context that is determined by the graph structure of the nodes and the relationships that link them together. 5
In this essay, I have introduced a graph-based data management and analysis framework that seeks to integrate parliamentary data with a NLP dependency tagged corpus. The proposed framework can ingest multiple data formats with disparate internal structures, while producing flexible and intuitive information structures that are highly conducive to comparative data analysis. Making use of the readily available Neo4j database, NLP tools, and a Python implementation of graph algorithms, it can easily be adapted to a wide variety of discipline-specific research approaches and customized to meet the requirements of individual researchers.
In our continued work on parliamentary data analysis, we plan to: a) harvest recent parliamentary data, b) implement new NLP tools for the Croatian language provided by the CLASSLA initiative, and include namedentity recognition (NER) data in the data tagging, c) carry out further research on stylometric analysis, particularly with regard to its diachronic dimension, d) integrate data from external sources of parliamentary data (https://edoc.sabor.hr, wikipedia, etc.), and e) extend existing semantic graph research to other dependency relations.
Special attention will be paid to further enhancing the web infrastructure for Croatian parliamentary data,9 which has recently been established with the help of funding from the University of Rijeka and the University of Zagreb University Computing Centre (SRCE), as such resources are essential for establishing a platform for collaborative scholarship on parliamentary data that encourages the exchange of information at an international level.
This work has been jointly sponsored by the Croatian Science Foundation under the project number UIP-05-2017-9219 and the University of Rijeka under the project number UNIRI-human-18-243. O _ E ER TA E B :D T
A EM rom :oD M f- t
N I: R D e TR :TS I v i S e ta :e m t n m na e a e s re -n r
u p -s e r
T X E N N
A T
E e T I:N R L
N r T R O c :ID be t:S T O n I m xe :gS t:B
tcene ceuN ceneT irco em e n n t d e n
n en t e -se tsne -sen -re cun S o - nn a S A H
H S A S A H
S A H
S A H
Y B D E R E V I L E D S A H S A H S A H y t r a
P F O _ E ER TA E B :D T
A EM rom :oD M f- te v i t a t n e s e r p e R
N :I R D e TR :TS I v i S e ta :e m t n m na e a e s re -n r
u p -s e r
T N I: R D e TR :TS I v i S e ta :e m t n m na e a e s re -n r
u p -s e r
T X E
N e T
N R R O c I: E T T O na IcenD :TAD i:trpS i:gnS t:nB r e c d e e ra ta sn rco em tt tteu -d t-ra r-e cnu
U
N A E L o n n a
R
TN TR :TS c I i :
S g p ID : n c e i i m rd o T t-op -na co e r
T n
IN E o : T E i ID A T s on :D A
i D s s m :o
s ro te se fS y l e s s A b
D I y m l b T N I: m e s s a their key:value properties: (Assembly) -[HAS]-> (Session) -[HAS]-> (Topic) -[HAS]-> ((Utterance) -[HAS]-> (Sentence) -[HAS]-> (Token) -[SEQUENCE]-> (Token) -[DEPENDENCY]-> (Token)) -[DELIVERED BY]-> ((Representative) -[MEMBER OF]-> (Club) ,-[MEMBER OF]-> (Political party)).
z ic za it t i a c n N to lemm tsyna ENR P e L k END l:reS P p
ED -ed R T o t :IInDN IEUQ :TSR :aTSR :gaTSR :tgaTSR :egTSR I:TN l:eTSR :sTSR T :S e N rm m ts s ua ade rp ep tko nU f-o le p p n -h -ed -d
m o o g - ke - -u -x l-a
Htio bih odmah na početku kazati pošto govorimo o djeci, a ponekad i u nekim raspravama ima dosta politiziranosti, htio bih kazati da se djeca ne rađaju ni radi politike, ni radi države, ni radi klase, ni radi nacije, da se rađaju radi ljubavi kao što je netko već jutros rekao i da je to posljedica možda jednog od posljednjih prirodnih odnosa između muškarca i žene, a koji su opet posljedica ljubavi.
Dakle, ja mislim da je žena kao majka, žena kao dio obitelji i ne znam zašto su se ljudi smijali kada se govorilo o ljubavi pa ja onda ne moram ni ljubav spomenuti ali dakle, žena koja izabere da će sa svojim suprugom barem u prijateljstvu imati određeni broj djece a usput završi fakultet, usput doktorira, usput i radi nazadna ja mislim da je ona vrlo slobodna osoba, da je ona suvremena i da je ona ravnopravnija od kolega muškaraca jer emancipacija se živi a ne priča... ... Count 100,684 94,195 91,425 89,971 87,245 85,868 85,413 85,325 80,573 80,302
Name 1 Šeks, Vladimir 2 Tafra, Višnja 3 Kajin, Damir 4 Đakić, Josip 5 Brkić, Milijan 6 Jandroković, Gordan 7 Pernar, Ivan 8 Špoljar, Dunja 9 Ćosić, Krešimir 10 Zgrebec, Dragica p in All 7.13e-06 2.852e-06 1.617e-06 1.592e-06 1.401e-06 1.095e-06 1.082e-06 1.07e-06 1.057e-06 1.006e-06
Auth 1,247,153
95,950 2,171,645 501,417 141,999 404,767 461,554 60,456 242,607 699,250
Nodes mir, stabilnost, vrijednost, demokracija, napredak, tolerancija, oprost, proizvodnja, blagdan, život, tišina, prosperitet, blagostanje, rast, sustav, gospodarstvo, zaštita, kvaliteta, aktivnost, povećanje, uvjet, društvo, novac, imovina, načelo, uloga, rješenje red, sigurnost, sloboda, pravda, pozivanje, govor, sjednica, zastupnik, rasprava, tema, prijedlog, pitanje, izvješće, standard, zdravlje, pravo, zakonodavstvo, građanin, odbor, odgovornost, odluka materijal, država, akcija, desetljeće, općina, grad, ministarstvo, zakon, proračun, prihod, samouprava, županija, razlog, zajednica, broj, republika, sredstvo, uprava, jedinica, turizam, udruga suradnja, interes, fokus, razvoj, plan, dobar, cilj, područje, reforma, izmjena, politika, strategija, program, potpora, prioritet, rezultat, potreba, stvar, mjera, dokument žena, mjesto, ugovor, posao, plaća, doprinos, radnik, rad, način, uvođenje, odnos, čovjek, dio, prostor, tisuća, obveza, projekt rat, stanje, oružje, prijetnja, slučaj, kod, problem, vrijeme, situacija, godina, primjer, djelo, mogućnost, podatak material, state, action, decade, municipality, city, ministry, law, budget, revenue, self-government, county, reason, community, number, republic, means, administration, unit, tourism, association cooperation, interest, focus, development, plan, good, goal, area, reform, change, policy, strategy, program, support, priority, result, need, thing, measure, document woman, place, contract, job, salary, contribution, worker, work, manner, introduction, relationship, man, part, space, thousand, obligation, project war, condition, weapon, threat, case, code, problem, time, situation, year, example, work, possibility, data
C 1 2