Raadslicht^*

Extracting and recommending information nuggets from council papers.

* Formerly known as WooPush.

1. Motivation

• Publishing government information does not by itself produce informed citizens (Ruijer et al., 2017). Open government data is underused, and more is not necessarily better (Safarov et al., 2017; Kempeneer & Wolswinkel, 2023); information has to be more than just accessible, also comprehensible and actionable (Gurstein, 2011).
• Council papers (vergaderstukken) are the documents attached to the agenda of municipal council meetings. They are published under the Woo, but rarely find their way outside of municipal councils. They do, however, contain concrete and valuable local information for residents, e.g., plans for a school to be rebuilt, average waiting times for social housing, or raised costs of parking fines.
• Research on mining municipal council data typically centers on their meetings, written minutes and transcripts (Hu et al., 2023; van Wijk & Marx, 2025; Marques et al., 2026), and focuses on transparency for accountability: reconstructing what governments do, how decisions are being made, and who is accountable, serving watchdogs, journalists and residents already interested in local politics.
• The council papers themselves, however, contain rich knowledge and information that decision-making is based on, and that is directly relevant to residents who may not be looking for it.

2. Data

Motivate terminology: Council papers, not documents, attachments, etc.

• They include council proposals, reports and budgets of regional bodies, studies by consultancies, policy notes, letters from organizations and residents.
• In Diemen, most council papers are not written by the municipality, but for it, and none of them seem particularly accessible and comprehensible for the general public. They do contain concrete, actionable and relevant information for residents: changing rates, places, dates, plans, problems and requests.
• They are challenging to process due to their heterogeneous nature: many different authors and genres, long PDFs with tables, charts and scans. All are in Dutch, however, typically in bureaucratic, complex Dutch.
• Under the Woo (art. 3.3(2)(c)), meeting documents of municipal councils are a category that is under active disclosure.
• Their reuse is governed by the Wet hergebruik van overheidsinformatie (Who), which implements the EU Open Data Directive: public government information may be reused by anyone, free of charge as far as possible, under conditions that may not unnecessarily restrict reuse.
• However, information in which a third party holds copyright is excepted (art. 2(1)(b)); which may be the case for many council papers that are written by third parties, e.g., consultancy reports.
• The minutes of council meetings (which record what actually was said during council meetings), are not available everywhere. Many (smaller) municipalities only publish video recordings of their meetings, which require speech recognition and video analysis to use (cf. van Wijk & Marx, 2025). Council papers already exist as text (or OCR'd text), are structured by meeting and agenda item, and carry the information that municipal decision making rests on.
• They are open data: freely available through each municipality's council information portal. And in bulk through the API of OpenRaadsinformatie, which lists over 6.1M meeting documents from 326 organizations.

Diemen, January–June 2026	Count
Distinct council papers attached to agenda items	225
Pages (total)	2,290
Council papers with ≥ 20 pages	34

Meeting documents, January–June 2026	Count
Schiermonnikoog (fewer than 1,000 inhabitants)	322
Diemen (pilot)	446
Amsterdam, city council	5,404
Amsterdam, including district committees	7,471

3. Task 1: Nugget extraction

From long, heterogeneous council papers to short statements a resident understands without having seen the document. Connect to your background in entity linking and semantic search for e-discovery.

• Definition: we define information nugget in this context as a short, self-contained text for residents, written from and verifiable against verbatim fragments of a council paper.
• Fragment: the underlying evidence, anchored to page and position. Nugget: the rewritten, decontextualized statement, which includes context a resident needs to understand it.

Nugget	Evidence
Op een sociale huurwoning wacht je in Diemen ruim tien jaar. Dat is al jaren zo. Ondertussen steeg het aantal actieve woningzoekenden in de regio van ruim 13.800 in 2020 naar bijna 35.800 in 2024.	“In de periode 2020 tot 2024 is de inschrijftijd van woningzoekenden tussen 10,4 en de 11,3 jaar voordat zij een woning vinden.”

3.1 Related tasks

Describe related work and neighboring IR and NLP tasks of this nugget extraction task.

• Open IE extracts relations without a schema, originally from the web (Banko et al., 2007); Knowledge Base Population (KBP) (Ji & Grishman, 2011; AKBC) collects them into a canonical knowledge base. Here there is no knowledge base; the nugget is the unit.
• Decontextualization makes a given sentence stand alone (Choi et al., 2021), also for scientific snippets shown to readers (Newman et al., 2023), adding as little as possible (Gunjal & Durrett, 2024).
• Newsworthiness detection by Spangher et al. (2024) on policy proposals, where they aim to predict which proposals will be covered by a newspaper, learned from past coverage. This is a document-level task; newsworthiness is one (of multiple) dimensions of what we aim at.
• Text simplification makes complex bureaucratic Dutch government letters more comprehensible, at the risk of having meaning “slip” (Beks van Raaij et al., 2024).
• Nuggets are also known from IR evaluation: TREC question answering (Voorhees, 2003), automatic and pyramid-style nugget evaluation (Lin & Demner-Fushman, 2005, 2006; Nenkova & Passonneau, 2004; Dang & Lin, 2007), nugget-based test collections (Pavlu et al., 2012), and report and RAG evaluation (Mayfield et al., 2024; Pradeep et al., 2024). This differs in…
• Trivia mining ranks interesting “did you know” statements about Wikipedia entities (Prakash et al., 2015).
• Propositions (Chen et al., 2024) and “atomic facts” (Min et al., 2023) decompose text exhaustively, mostly Wikipedia-like text, for retrieval or for verifying generated text.

3.2 Challenges

1. Scoping/defining and evaluating nuggets. What counts as a nugget, and what makes a good one: correct, self-contained, supported, informative. Informativeness eventually has to be judged by real users in the down-stream task, not (only) by annotators.
2. Decontextualization across (heterogeneous) documents. The missing context can often be outside the sentence or even the document: author, place, municipality (regional reports cover several), meeting.
3. Text simplification. Decontextualized text should be ‘translated’ from bureaucratic Dutch (ambtelijk taalgebruik) to plain language, but remain faithful in meaning.
4. Timestamp/timeline extraction. Council papers can repeat old or redundant information. Being able to reliably position a nugget's information in time needs to be extracted, in some cases from the council paper itself, in others from the meeting's metadata.
5. Urgency and actionability. Deadlines and ways to respond are rarely stated explicitly; but should be extracted as additional attributes of the nugget: “can the citizen still act?” Information that affects someone, but offers no (more) way to respond may frustrate users, rather than empower them (Gurstein, 2011).
6. (Citation) Faithfulness and Factuality. Every nugget must be supported by verbatim fragments. Wrong information presented to residents, is the main risk. Nugget evaluation (does this text support this nugget?) also implies checking it for factuality.

4. Task 2: Nugget recommendation

From extraction to delivery: which nuggets does a resident see, when, and why? Informatie die je aangaat. This should introduce the main WooPush angle; using RecSys for better informed citizens; why (societal need) and how (“value-driven recsys”).

• Democratic processes. Ruijer et al. (2017) distinguish monitorial, deliberative and participatory democracy, each with different roles for citizens and government, and find open data platforms too simple for them. Raadslicht is closest to the monitorial process, aimed at what concerns residents rather than at overseeing government. We do not care about “exposing” the inner working of governments (transparency for transparency), but rather informing citizens of relevant information.
• Citizenship. Westheimer and Kahne (2004) distinguish personally responsible, participatory and justice-oriented citizens, with different information needs. We start with the personally responsible citizen; the participatory (sharing, networks) and justice-oriented (patterns across municipalities) come later.
• Monitorial citizens. Residents scan their environment and act when something concerns them (Schudson, 1998). News should work as a burglar alarm, not a police patrol (Zaller, 2003). We should design for reactive, not proactive, citizens.
• Effective use of open (government) data. Access is not use: information has to be accessible, comprehensible and actionable (Gurstein, 2011).

Dimension	Aspect	IR / RecSys concept
Relevance	Geographic, topical	Personalization, user modeling
Urgency	Time-bound	Freshness, temporal relevance
Urgency	Anomalous	Novelty, serendipity
Urgency	Controversial	Engagement prediction, diversity
Actionability	Comprehensible	Readability
Actionability	Concrete	Information extraction

4.1 What do residents want?

A small user study of a mock prototype with 26 residents from Amersfoort, in addition to interviews with 5 residents and 5 municipal experts (Summerfield-Son, 2026), for which the following set of “relevance dimensions” were defined (Table 4), found:

• Compared with raw text and a plain summary, an actionable summary (“What can you do?”) raised perceived relevance and urgency, without more perceived effort than the plain summary.
• Residents build trust in such a system through its track record, and through citation faithfulness, and sources they can check.
• Municipal experts warn that simplification loses legal nuance and needs verification (cf. Beks van Raaij et al., 2024), and that digital-only delivery excludes some residents.
• Experts suggest place and time (maps, timelines) rather than document categories as the way to reach residents: they want to know what is in it for them.

4.2 Challenges

1. What to optimize for. Clicks are a poor proxy for being informed. Recommenders can be optimized beyond clicks, for editorial values (Lu et al., 2020), and for normative goals such as those of democratic theory (Vrijenhoek et al., 2021). Here the target is informed citizenship: did a resident learn something that concerns them? This is a long-term signal with likely delayed feedback loops.
2. Stakeholders. Residents, the municipality and the council may want different things from the same system (Jannach & Adomavicius, 2016). One way to frame it is to say recommenders for citizens should serve society (Yang et al., 2017).
3. Proactive push delivery. No query, explicit information need, with users who may show little interest in and engagement with local politics. News push notifications face a similar challenge: which item is worth interrupting someone for, and whom (Loni et al., 2019). Newsrooms answer it mostly by newsworthiness for a general audience; here it depends on what concerns one resident; newsworthiness being but one potential signal.

References

1. Banko, M., Cafarella, M. J., Soderland, S., Broadhead, M., & Etzioni, O. (2007). Open Information Extraction from the Web. IJCAI 2007.
2. Beks van Raaij, N., Kolkman, D., & Podoynitsyna, K. (2024). Clearer Governmental Communication: Text Simplification with ChatGPT Evaluated by Quantitative and Qualitative Research. DeTermIt! Workshop, LREC-COLING 2024.
3. Chen, T., et al. (2024). Dense X Retrieval: What Retrieval Granularity Should We Use? EMNLP 2024. arXiv:2312.06648.
4. Choi, E., Palomaki, J., Lamm, M., Kwiatkowski, T., Das, D., & Collins, M. (2021). Decontextualization: Making Sentences Stand-Alone. TACL 9.
5. Dang, H. T., & Lin, J. (2007). Different Structures for Evaluating Answers to Complex Questions: Pyramids Won't Topple, and Neither Will Human Assessors. ACL 2007, 768–775.
6. Gunjal, A., & Durrett, G. (2024). Molecular Facts: Desiderata for Decontextualization in LLM Fact Verification. Findings of EMNLP 2024, 3751–3768.
7. Gurstein, M. (2011). Open Data: Empowering the Empowered or Effective Data Use for Everyone? First Monday 16(2).
8. Hu, Y., Ganter, T., Deilamsalehy, H., Dernoncourt, F., Foroosh, H., & Liu, F. (2023). MeetingBank: A Benchmark Dataset for Meeting Summarization. ACL 2023, 16409–16423.
9. Jannach, D., & Adomavicius, G. (2016). Recommendations with a Purpose. RecSys '16.
10. Ji, H., & Grishman, R. (2011). Knowledge Base Population: Successful Approaches and Challenges. ACL 2011.
11. Kempeneer, S., & Wolswinkel, J. (2023). Rethinking Open Government Data for Citizen Participation. Information Polity 28(2).
12. Lin, J., & Demner-Fushman, D. (2005). Automatically Evaluating Answers to Definition Questions. HLT-EMNLP 2005.
13. Lin, J., & Demner-Fushman, D. (2006). Will Pyramids Built of Nuggets Topple Over? HLT-NAACL 2006, 383–390.
14. Loni, B., Schuth, A., van de Haas, L., Jansze, J., Visser, V., & van der Wees, M. (2019). Personalized Push Notifications for News Recommendation. ORSUM @ RecSys 2019, PMLR 109, 36–45.
15. Lu, F., Dumitrache, A., & Graus, D. (2020). Beyond Optimizing for Clicks: Incorporating Editorial Values in News Recommendation. UMAP 2020, 145–153.
16. Marques, M., Fernandes, A. L., Pacheco, A. F., Rebouças, R., Cantante, I., Isidro, J., Cunha, L. F., Jorge, A., Guimarães, N., Nunes, S., Leal, A., Silvano, P., & Campos, R. (2026). CitiLink-Summ: A Dataset of Discussion Subjects Summaries in European Portuguese Municipal Meeting Minutes. Proceedings of the ACM Web Conference 2026 (WWW '26), 8705–8708. doi:10.1145/3774904.3792945.
17. Mayfield, J., et al. (2024). On the Evaluation of Machine-Generated Reports. SIGIR 2024 (perspective paper).
18. Min, S., et al. (2023). FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation. EMNLP 2023.
19. Nenkova, A., & Passonneau, R. (2004). Evaluating Content Selection in Summarization: The Pyramid Method. HLT-NAACL 2004.
20. Newman, B., Soldaini, L., Fok, R., Cohan, A., & Lo, K. (2023). A Question Answering Framework for Decontextualizing User-facing Snippets from Scientific Documents. EMNLP 2023, 3194–3212.
21. OpenRaadsinformatie (OpenBesluitvorming). zoek.openraadsinformatie.nl, accessed November 2026.
22. Pavlu, V., Rajput, S., Golbus, P. B., & Aslam, J. A. (2012). IR System Evaluation using Nugget-based Test Collections. WSDM 2012, 393–402.
23. Pradeep, R., Thakur, N., Upadhyay, S., Campos, D., Craswell, N., & Lin, J. (2024). Initial Nugget Evaluation Results for the TREC 2024 RAG Track with the AutoNuggetizer Framework. arXiv:2411.09607.
24. Prakash, A., Chinnakotla, M. K., Patel, D., & Garg, P. (2015). Did You Know? Mining Interesting Trivia for Entities from Wikipedia. IJCAI 2015, 3164–3170.
25. Ruijer, E., Grimmelikhuijsen, S., & Meijer, A. (2017). Open data for democracy: Developing a theoretical framework for open data use. Government Information Quarterly 34(1).
26. Safarov, I., Meijer, A., & Grimmelikhuijsen, S. (2017). Utilization of Open Government Data: A Systematic Literature Review. Information Polity 22(1).
27. Schudson, M. (1998). The Good Citizen: A History of American Civic Life. Free Press.
28. Spangher, A., et al. (2024). Tracking the Newsworthiness of Public Documents. ACL 2024.
29. Summerfield-Son, R. T. (2026). Towards Effective Use of Open Government Data: Investigating the Impact of Content and Presentation on Citizen Engagement in a Civic Recommender System (WooPush). BSc thesis, University of Amsterdam.
30. van Wijk, P., & Marx, M. (2025). Spoken Question Answering on Municipal Council Meetings. Advances in Information Retrieval (ECIR 2025), LNCS, 41–46.
31. Voorhees, E. M. (2003). Evaluating Answers to Definition Questions. HLT-NAACL 2003 (short papers), 109–111.
32. Vrijenhoek, S., Kaya, M., Metoui, N., Möller, J., Odijk, D., & Helberger, N. (2021). Recommenders with a Mission: Assessing Diversity in News Recommendations. CHIIR 2021, 173–183.
33. Westheimer, J., & Kahne, J. (2004). What Kind of Citizen? The Politics of Educating for Democracy. American Educational Research Journal 41(2).
34. Wet hergebruik van overheidsinformatie (Who), BWBR0036795, version of 21 November 2025. Implements Directive (EU) 2019/1024 on open data and the re-use of public sector information.
35. Wet open overheid (Woo), art. 3.3(2)(c). Working definition of the category vergaderstukken decentrale overheden on open-overheid.nl.
36. Yang, J., Cantador, I., Nurbakova, D., Cortés-Cediel, M. E., & Bozzon, A. (2017). Recommender systems for citizens: the CitRec'17 workshop manifesto. CitRec '17.
37. Zaller, J. (2003). A New Standard of News Quality: Burglar Alarms for the Monitorial Citizen. Political Communication 20(2).