Source-linked AI summary
Exploring the Political Agenda of the European Parliament Using a Dynamic Topic Modeling Approach
Derek Greene, James P. Cross
TL;DR
The study addresses the unavailable macro-level picture of MEP attention and difficult tracing of influence by proposing two-layer NMF for political speech corpora over time. It finds that NMF identifies semantically more coherent topics than LDA and is designed to capture niche topics.
Problem
A macro-level picture of where and on what topics MEP attention lies was unavailable, while tracing influence had been difficult.
Method
The study proposes a two-layer NMF methodology for identifying topics in large political speech corpora over time and extracting latent content patterns.
Results
NMF can identify topics that are semantically more coherent in political speech corpora than a probabilistic method such as LDA.
Takeaways & Limitations
The methodology is designed to identify niche topics and associated vocabularies in political speech corpora over time.
Takeaways & Limitations
Tracing influence and obtaining a macro-level picture of MEP attention had been difficult or unavailable.
Abstract
from arXiv · showhide
This study analyzes the political agenda of the European Parliament (EP) plenary, how it has evolved over time, and the manner in which Members of the European Parliament (MEPs) have reacted to external and internal stimuli when making plenary speeches. To unveil the plenary agenda and detect latent themes in legislative speeches over time, MEP speech content is analyzed using a new dynamic topic modeling method based on two layers of Non-negative Matrix Factorization (NMF). This method is applied to a new corpus of all English language legislative speeches in the EP plenary from the period 1999-2014. Our findings suggest that two-layer NMF is a valuable alternative to existing dynamic topic modeling approaches found in the literature, and can unveil niche topics and associated vocabularies not captured by existing methods. Substantively, our findings suggest that the political agenda of the EP evolves significantly over time and reacts to exogenous events such as EU Treaty referenda and the emergence of the Euro-crisis. MEP contributions to the plenary agenda are also found to be impacted upon by voting behaviour and the committee structure of the Parliament.
1 Introduction
The study addresses the lack of a holistic account of the European Parliament’s plenary agenda by analyzing latent themes in MEP speeches over three parliamentary terms. It develops a two-layer NMF dynamic topic model and finds that the agenda evolves over time in relation to external events.
- The study examines how the EP policy agenda evolves and how MEP speech-making responds to internal and external stimuli.
- The analysis applies a dynamic topic model based on two layers of Non-negative Matrix Factorization to legislative speeches.
- The corpus contains 210,247 speeches from 1,735 MEPs across the 28 EU member states during 1999–2014.
- The method reveals the breadth of the EP policy agenda and shows that it evolved significantly over time.
- Case studies connect topic evolution with exogenous events including the Euro-crisis and EU treaty changes.
- The study also examines determinants of MEP speech-making behavior on the detected topics using external data sources to assess topic validity.
2 Related Work
Related work shows that political agendas can change through punctuated attention dynamics, while MEP behavior and institutional rules shape speech-making. Topic models provide tools for tracing these patterns, but EP plenary agendas remain under-explored.
- The literature notes conceptual and measurement challenges in studying agenda dynamics.
- European Council agendas exhibit long periods of stability interrupted by sharp spikes in issue attention, shaped by institutional, contextual, and issue-specific factors.
- Research on EU policy agendas has largely focused on the European Council, while agendas in institutions including the EP remain difficult to capture and under-explored.
- MEP speech-making and voting behavior are influenced by plenary speaking-time rules, party-group allocation, committee membership, and committee roles.
- Topic models infer latent semantic structure from word co-occurrences and have been used to track political attention across documents and over time.
- NMF decomposes non-negative matrices into additive factors that can be interpreted as topics and can identify niche topics under-reported by traditional LDA approaches.
- Prior studies show that political speech agendas can respond to internal and external stimuli, including exogenous shocks that produce abrupt increases in issue attention.
3 Methods
The paper develops a two-layer NMF strategy that models topics within successive time windows and then links related window topics into dynamic topics. This approach uses TF-IDF-weighted speech matrices, coherence-based parameter selection, and top-term representations to track evolving political agendas.
- Dynamic topic modeling: Two-layer NMF first models speeches separately in successive time windows, then combines window-topic outputs to identify dynamic topics spanning multiple windows.The first layer produces successive window topic models; the second factorization operates on their condensed topic-term representation.
- Single-window modeling: TF-IDF weighting supports diverse, semantically coherent topics and helps the model identify both broad document groups and niche topics.The method is motivated by NMF’s ability to distinguish procedural parliamentary material from focused policy discussions.
- Single-window modeling: NMF factorizes a document-term matrix as A ≈ WH, where H represents topics through non-negative term weights and W captures speech-topic membership.The ranked terms in each row of H provide topic descriptors, while W can associate speeches and MEPs with topics using metadata.
- Parameter selection: TC-W2V selects the topic number by measuring semantic similarity among top-ranked topic terms and choosing the value with maximum mean coherence.Individual topic coherence is based on mean pairwise cosine similarity between word2vec term vectors; model coherence averages topic scores.
- Dynamic topic modeling: Disjoint time windows preserve both period-specific agendas and short-lived topics that could be obscured in a single full-corpus model or overlapping windows.The paper therefore divides the timestamped parliamentary corpus into equal-length windows before applying NMF.
- Dynamic topic modeling: The second NMF layer stacks top-ranked terms from window topics into a reduced matrix, retaining descriptive terms before extracting dynamic topics.Feature selection lowers the computational cost by excluding terms that never rank prominently in any window topic.
4 Data
The dataset contains English-language European Parliament plenary speeches retrieved from Europarl and organized into quarterly windows from 1999-Q3 through 2014-Q2. English speeches were used because multilingual modeling and automated translation raised accuracy and reliability concerns.
- Corpus construction: 210,247 English-language speeches were identified, representing 77.95% of the original collection of parliamentary speeches.The corpus focuses on English speeches, either native or translated, because multilingual and automated-translation strategies were considered insufficiently reliable.
- Temporal organization: The speeches were divided into 60 quarterly, non-overlapping windows from 1999-Q3 to 2014-Q2.Quarterly windows balance granular topic identification with sufficient speeches for modeling and avoid empty windows caused by the parliamentary summer recess.
- Temporal organization: Window sizes ranged from 679 speeches in 2004-Q3 to 9,151 speeches in 2011-Q4, averaging 4,811 terms per dataset.Initial experiments with shorter windows and fewer speeches often produced fewer coherent topics.
- Preprocessing: Speech preprocessing removed headers, footers, short tokens, generic and parliamentary stop words, politician names, and tokens occurring in fewer than five speeches.Remaining unigram tokens were case-converted, tokenized, lemmatized, TF-IDF weighted, and normalized by document length.
5 Assessing the Coherence of LDA and NMF Topic Models: A Baseline Comparison
The comparison evaluates NMF and LDA across EP time-window datasets using topic-coherence measures. NMF consistently produces more coherent, niche, and time-variant topic representations than LDA.
- For Eurocrisis debates, NMF produces richer and more varied vocabularies, yielding a more informative and time-variant picture of debate evolution.
- NMF achieves higher topic-coherence scores than LDA across all time-window datasets and tested topic numbers.
- NMF appears to uncover more niche and specific topics, whereas LDA extracts broader and less semantically coherent topics.The difference is attributed to the distinctiveness of NMF topic descriptors compared with LDA descriptors.
- NMF scores higher than LDA in 94.7% of the 300 experiments under the Cv coherence measure.
- TC-W2V is more sensitive to changes in topic terms and highlights windows with differing coherence for both algorithms.
6 Experimental Results
The dynamic modeling procedure identifies topic variation across 60 windows and links agenda changes to external events and MEP institutional behavior. The results support event-sensitive agenda dynamics and committee-structured speechmaking.
- 6.1 Experimental Setup: The number of detected topics varies across windows and is unrelated to quarterly speech volume, with a Pearson correlation of 0.006.This suggests variation reflects differences in discussed topics rather than the number of speeches.
- 6.1 Experimental Setup: The analysis yields 1,017 window topics represented by 2,710 distinct terms.
- 6.1 Experimental Setup: Dynamic-topic selection reaches a coherence maximum at k′ = 57, with similar models and nearby peaks between 62 and 80.The nearby models differ mainly through merges or splits of strongly related topics.
- Financial/Euro-crisis: Financial- and Euro-crisis topics show distinct attention peaks corresponding to major events, including the Lehman collapse, Greek debt revelations, the Irish bailout, and Draghi’s 2012 statement.Fewer speeches follow Draghi’s statement, which temporarily reassured markets.
- EU Treaty reforms: Treaty-related attention spikes after major external events, indicating that the EP is reactive rather than proactive in Treaty debates.The pattern is consistent with low baseline attention disrupted by events such as Treaty signatures and referenda.
- Fisheries Policy: Fisheries attention remains stable from 2000 to 2010 before increasing during the Commission’s 2010 public consultation on policy reform.
- 6.3 Explaining MEP Speech Counts: The committee system fundamentally shapes plenary speechmaking and the European Parliament’s policy agenda.
7 Conclusions
The study introduces two-layer NMF for tracing EP plenary agendas and applies it to English-language speeches from 1999–2014. It identifies coherent niche and broad topics, agenda changes associated with external events, and links between MEP attention, voting behavior, and institutional position.
- Method and contribution: Two-layer NMF is proposed for identifying dynamic topics in large political speech corpora over time.The method is designed to identify both niche topics with specialized vocabularies and broader topics with general vocabularies.
- Method and contribution: NMF topics can be semantically more coherent in political speeches than topics produced by probabilistic LDA.The method is applied to approximately 210,000 English-language EP plenary speeches from 1999–2014.
- Agenda evolution: The EP agenda distinguishes routine policy work from new discussion topics associated with economic crises and failed treaty referenda.Three case studies are used to examine these differences across the period studied.
- Agenda evolution: MEP voting behavior and institutional position affect whether MEPs contribute to particular agenda topics.The analysis examines determinants of MEP attention during the EP’s seventh sitting.
- Implications and scope: The approach provides insights into latent dynamics in MEP speech-making and the functioning of the EU as a political system.The authors state that the method can also be applied to legislative, media, and other digitally available text agendas beyond Europe.
- Implications and scope: Linking political attention to policy influence remains difficult because a macro-level picture of MEP attention had been unavailable.The conclusion identifies tracing attention to policy outcomes as an area for further exploration.
Appendix A: Baseline Comparison
The study compares NMF and LDA topic models across 60 time-window datasets and topic counts from 10 to 50. The comparison uses topic coherence scores to assess model quality.
- The comparison evaluates the coherence of models generated by both topic-modeling approaches.
- NMF and LDA are compared across 60 time-window datasets with k ranging from 10 to 50 topics.
- Figures 5 and 6 report median Cv and TC-W2V coherence scores across the time-window datasets.
Intra-Topic Validity
The authors assess whether individual dynamic topics are semantically coherent using TC-W2V scores calculated from term similarities in a word2vec space. The results show that coherence varies across policy and administrative topics.
- The dynamic topics are evaluated in a word2vec space constructed from the complete speech corpus.
- TC-W2V coherence values measure the mean pairwise cosine similarity among each topic’s top 10 terms in the word2vec space.
- The most coherent topics often correspond to core EU competencies, whereas broad administrative topics are least coherent.
- 0.36 is considerably higher than the lower bound for TC-W2V, suggesting a high level of semantic validity.
Inter-Topic Validity
The authors assess inter-topic semantic validity by clustering 57 dynamic topics according to their term-use patterns. The resulting groupings align with substantively related policy areas and institutional functions.
- Average-linkage hierarchical clustering groups the 57 dynamic topics using normalized Pearson correlation between second-layer NMF factor-H row vectors.
- Lower dendrogram connection heights indicate greater similarity in topics’ term-usage patterns.
- The clustering reveals related groups covering transport, energy, animal health, institutional interactions, education and research, trade, and EU enlargement.
- These hierarchical associations support semantic validity because expected topic relationships correspond to correlated terms in factor H.
External Validation
The authors externally validate the 57 dynamic topics against the Europarl taxonomy of legislative-procedure subjects. Matching examples and broader coverage indicate that the topics correspond to expected EU policy areas.
- The analysis compares 57 unsupervised dynamic topics with an existing Europarl taxonomy of legislative-procedure subjects.
- The taxonomy ranges from broad top-level subjects to highly specific low-level subjects, with the comparison using its second level.
- Dynamic topics are matched to taxonomy subject documents by cosine similarity between topic top-10 terms and subject descriptions with lower-level subjects.
- The topic labeled ‘Tax’ matches the Europarl subject ‘2.70 Taxation’ and its direct- and indirect-taxation subcategories.
- The topic labeled ‘Drugs’ matches ‘4.20 Public health’ and its pharmaceutical-products subcategory.
- The matches indicate good coverage of expected EP debate policy areas and increase confidence in the model’s construct validity.
Appendix C: Dynamic Comparison
The comparison finds that NMF and DTM produce broadly coherent dynamic topics but differ substantially in how topic terms evolve across time windows. NMF terms change more and better reflect window-specific parliamentary discussions, while DTM terms remain relatively stable.
- Methods: The comparison applied NMF and DTM to the same parliamentary-speech time-window datasets using k = 50 dynamic topics.DTM used its original C++ implementation with the authors’ recommended default parameters.
- Coherence: NMF achieved a marginally higher median TC-W2V coherence than DTM, 0.277 versus 0.276.The distributions for all 50 dynamic topics are shown in Fig. 9.
- Window interpretation: NMF window topics had more diverse top-10 terms for climate change, reflecting changing discussion items across five quarterly windows.The example is reported in Table 5 and concerns the same time-window datasets used by both methods.
- Term agreement: The overall mean Jaccard score was 0.166 for NMF and 0.921 for the probabilistic approach, indicating much more frequent term changes under NMF.The Jaccard coefficient measures the intersection-over-union of term sets; scores of 1 and 0 indicate identical and disjoint sets, respectively.
- Window interpretation: Because NMF models each window independently, its top terms more closely reflect parliamentary discussions during that window and support topic interpretation.DTM builds topics sequentially, so its reported top terms are relatively stable across windows.
- Limitations: The English-language speech corpus under-represents some MEPs, so the substantive results should be interpreted with this limitation in mind.The passage links this sample-selection issue to the variable availability of speeches in English.