PolicyStory: Chronological Policy Summaries
- PolicyStory is an information tool that aggregates and clusters Indian policy news into temporally ordered narratives.
- It employs an LLM-based classifier to group articles by topic and generate three distinct summary layers for varied reading needs.
- Empirical studies reveal high usability and trust, though users suggest enhancements like source citations and bias detection.
PolicyStory is an information tool designed to offer lucid, chronological, and summarized insights into Indian policy issues. It collects news articles from diverse sources, clusters them by topic, and generates three levels of summaries from longitudinal media discourse on policies, leveraging open source LLMs. Its stated purpose is to address volume and fragmentation of news, difficulty in building longitudinal understanding, and the complexity and jargon of policy coverage by giving users a bird’s-eye, chronological view of a policy issue rather than disconnected snapshots (Dar et al., 10 Sep 2025).
1. Definition and conceptual scope
PolicyStory, as introduced for Indian policy-news, is centered on two claims. First, existing formats—individual articles, news apps, feeds, and short-form videos—are designed for immediacy rather than for understanding how an issue started, how it evolved over time, what the key subtopics were, and how the government’s stance or numbers changed across years. Second, policy/news articles often use technical terms and complex sentences, so summarization must not merely shorten content but also explain jargon and preserve chronology (Dar et al., 10 Sep 2025).
The named system is most precisely a policy-news aggregation and summarization pipeline, not a document-grounded policy reasoning system. At the same time, the term also aligns with a broader research pattern in which policy is rendered as narratives, scenarios, or structured explanations. Answer Set Programming has been used to formalize privacy policies and generate temporally ordered stories about possible consequences (Dabral et al., 2022). The CONSENSUS project describes a “policy story” as a narrative connecting citizen objective priorities to publicly acceptable policy implementations and trade-offs (Tserpes, 2015). PolicyBot, in turn, explicitly treats its retrieval-augmented question-answering stack as a factual backbone for a future PolicyStory UX over laws, regulations, and guidelines (Nagarajan et al., 17 Nov 2025).
2. Corpus construction and longitudinal organization
The first pilot covers two major Indian policy events: Union Budget (2019–2024) and Farmers’ Protests (2020–2024, with 2022 included because public discussion continued, and early 2024 covering “Farmers’ Protest 2.0”) (Dar et al., 10 Sep 2025). Data acquisition is based on Media Cloud, with event-specific keyword queries. For Farmers’ Protests, the query is "farmers" AND ("protest" OR "agitation" OR "farm laws" OR "MSP" OR "march to Delhi"); for Union Budget, it is "budget" AND ("finance minister" OR "union budget" OR "fiscal policy" OR "tax reforms" OR "Nirmala" OR "budget speech" OR "budget allocation" OR "fiscal deficit") (Dar et al., 10 Sep 2025).
Because of computational limits, the system caps collection at 2,000 articles per year per policy issue and uses stratified sampling across months so coverage is temporally spread out rather than concentrated during peaks. After scraping and filtering with Newspaper3k, the usable article counts are reported as follows: Union Budget—2019: 157, 2020: 218, 2021: 1,467, 2022: 1,652, 2023: 1,453, 2024: 1,676; Farmers’ Protest—2020: 1,456, 2021: 1,508, 2022: 1,667, 2024: 1,669 (Dar et al., 10 Sep 2025).
Temporal structuring is year-wise. Within each policy issue, PolicyStory defines a fixed number of topic clusters: 5 for Farmers’ Protests and 8 for Union Budget. It does not use classical clustering such as k-means on TF–IDF or probabilistic topic models such as LDA. Instead, it uses an LLM-based classifier over title plus first paragraph, with a prompt describing the possible topics and asking the model to choose the most appropriate one. The paper gives the conceptual form
where is the predefined topic set for that policy (Dar et al., 10 Sep 2025).
Cluster quality is evaluated manually. Three annotators examined 100 articles per cluster and judged whether each article belonged to the cluster. The reported inter-annotator agreement is 0.89, which the paper interprets as strong consistency for the LLM-based topic assignment (Dar et al., 10 Sep 2025).
3. Hierarchical summarization architecture
The architecture is a pipeline: Data Acquisition; Organization by Date and Topic; Article-level Summarization; Cluster-level L2 Summaries; Cluster-level L1 Summaries; Numeric Summary Extraction; Storage and Frontend (Dar et al., 10 Sep 2025). The implementation uses Llama-3.2-1B through Hugging Face Transformers on Google Colab with an A100 GPU, and no fine-tuning is reported; the system is zero-shot and prompt-driven.
At article level, each article is summarized into 5–8 sentences. These article summaries are then grouped by . Because Llama 3.2 1B has an approximately 4,096-token context window, cluster-level processing batches up to 20 article summaries per call. L2 stories are then generated as detailed, story-like narratives that describe the evolution over that year within the topic and explain relevant jargon and technical terms (Dar et al., 10 Sep 2025).
The higher-level summary structure is explicitly layered:
| Layer | Input | Output |
|---|---|---|
| L2 | Article summaries in a topic-year cluster | Detailed, story-like narrative |
| L1 | L2 story | One-paragraph brief |
| Numeric | L2 story | Key–value pairs of numeric facts |
The L1 stage is defined conceptually as
where is policy, is year, and is topic (Dar et al., 10 Sep 2025). Numeric summaries are extracted from L2 stories with a specialized prompt to produce key–value pairs such as Defense Budget: [INR](https://www.emergentmind.com/topics/implicit-neural-representation-inr-decoder) 5.94 lakh crore. Storage is in MongoDB, and the frontend is a React + Next.js web application (Dar et al., 10 Sep 2025).
This hierarchical design is central to the system’s notion of comprehension. L1 is intended for approximately 30-second reading, L2 for deeper narrative understanding, and numeric summaries for quantitative inspection. The paper’s emphasis is therefore not only summarization quality in the generic sense, but multi-level access calibrated to different reading times and information needs (Dar et al., 10 Sep 2025).
4. Interface design and empirical findings
The interaction model is minimal and chronological. Users select a policy issue, choose a topic, and then browse year-wise summaries. L1 summaries are shown in chronological order, with L2 and numeric summaries available for specific year-topic pairs. The authors deliberately chose oldest-first ordering to emphasize narrative evolution, although some participants suggested reverse chronological order. Numeric summaries are presented on a separate page, and jargon is defined in simpler terms within the summaries (Dar et al., 10 Sep 2025).
The evaluation is an online survey with 22 participants recruited from two academic institutions. The majority were 25–34 years old; the sample included students, researchers, and working professionals; approximately 59.1% were male and 40.9% female. In news-consumption habits, 40.9% were daily readers and 27.3% occasional readers. The study focused on interface ease of use and navigation, clarity and helpfulness of L1 and L2 summaries, usefulness and clarity of numeric summaries, perceived trustworthiness and bias, whether participants learned something new, and whether they preferred using PolicyStory to reading multiple individual news articles (Dar et al., 10 Sep 2025).
The reported findings are specific. Eighteen of 22 users rated the interface 4 or 5 out of 5 for ease of use. For L1 summaries, 95.5% of participants reported that even from L1 summaries alone they learned something new about the policy issues. Users described L1 as clear and concise, and most felt the level of detail was “just right.” L2 summaries were praised for their narrative style, richer detail, and clearer overview of policy evolution. Numeric summaries were generally found informative, though some participants wanted more context and easier visibility (Dar et al., 10 Sep 2025).
Trust and adoption results were also favorable but not unqualified. The paper reports that 77.2% of users found the content trustworthy, and a majority found the system unbiased. At the same time, a few participants perceived tonal bias, with one participant describing some Budget summaries as somewhat “anti-government” and another describing some Farmers’ Protest summaries as more “pro-farmer.” Several participants requested explicit source citations next to summaries to improve trust. Adoption intent was reported as 81.8% for regular use and 86.4% for recommendation to others (Dar et al., 10 Sep 2025).
5. Relation to adjacent policy explanation paradigms
A key contrast is between PolicyStory’s policy-news focus and document-grounded policy question answering. PolicyBot is designed for reliable, document-grounded question answering over laws, regulations, circulars, and guidelines, using domain-specific semantic chunking, multilingual dense embeddings, multi-stage retrieval with reranking, source-aware generation, and citation tracing. The paper explicitly states that for a system like PolicyStory the same issues apply: precise retrieval of the right clauses, grounding in actual text, handling of cross-references, definitions, and exceptions, and balancing simplification with verifiability (Nagarajan et al., 17 Nov 2025). This suggests that PolicyStory’s current news-based abstraction and PolicyBot’s clause-level grounding are complementary rather than interchangeable.
A second relation is to formal narrative generation. “Exploring Consequences of Privacy Policies with Narrative Generation via Answer Set Programming” encodes privacy policies as constraints on a narrative planning space, where each answer set corresponds to a possible world and thus to a concrete story involving actors, roles, actions, and data flows. The system can answer possibility queries, identify which clauses are violated by a given sequence of events, and generate narratives explaining how PHI may be disclosed or why a disclosure is forbidden (Dabral et al., 2022). Relative to that framework, PolicyStory is not a planner over possible worlds; it is a summarizer over observed longitudinal discourse.
A third relation is to public-value storytelling. The CONSENSUS project models policy design as a multi-objective optimization problem, asks citizens to rank or prioritize objectives, and then identifies implementations that are Pareto-efficient and align with citizen-expressed priorities. In that setting, the “policy story” is the narrative that explains how publicly acceptable implementations emerge from objective priorities and what trade-offs they imply (Tserpes, 2015). PolicyStory’s own summaries of news discourse do not compute such trade-offs, but they occupy a nearby explanatory space: they organize a complex policy issue into temporally ordered thematic narratives rather than formal optimization narratives.
A fourth relation concerns controlled summarization at the entity level. EROS and PD-Sum treat privacy policy summarization as an entity-centric problem organized around data, source, medium, target, and reason, and use controlled abstractive summarization with PPO to enforce entity coverage (Singh et al., 2024). This suggests an additional control axis for PolicyStory: beyond chronology and topic structure, policy-news summaries could also be constrained to preserve critical entities, actors, and rationales whenever disclosure precision is important.
6. Limitations and prospective development
The current system has several explicit limitations. Media Cloud includes both mainstream and smaller outlets, raising potential bias in coverage and varying levels of reliability. The system currently lacks fact-checking and explicit bias detection. Its summaries depend fully on an open-source LLM used in zero-shot mode, so the paper notes risks of hallucinations, omissions, and tone that might lean toward certain framings. No automatic evaluation of information fidelity such as BERTScore, ROUGE, or BLEURT has yet been applied in the PolicyStory paper itself, and no formal readability metrics are used to quantify readability improvement. Scope is also narrow: the current system covers only two policy issues over a limited time range, with article counts lower for earlier years because of Media Cloud coverage (Dar et al., 10 Sep 2025).
The paper’s own future directions are concrete: integrate fact-checking modules and news bias detection techniques; add source citations within summaries; incorporate expert opinions; provide charts and trend analysis over numeric summaries; support a user-controlled time-to-read parameter; allow dynamic querying over new policy issues and keywords; expand to more policy issues and broader demographics in user studies; and extend beyond English to Indian regional languages (Dar et al., 10 Sep 2025). These are extensions of the existing architecture rather than a change in the system’s basic orientation.
Two further implications follow from neighboring work. A plausible implication is that source-grounded citation tracing, as implemented at chunk level in PolicyBot, would directly address the explicit user request for source citations and strengthen trust claims in governance settings (Nagarajan et al., 17 Nov 2025). A second plausible implication is that scenario-based explanations, of the kind generated by ASP narrative planning for privacy policies, could add a complementary mode to PolicyStory: not just what the news said year by year, but how a policy might apply in a structured fact pattern or “what-if” case (Dabral et al., 2022).
PolicyStory therefore occupies a specific position in the policy-technology landscape. It is not a legal authority, not a clause-grounded compliance engine, and not a formal planner of policy consequences. It is a longitudinal policy-news summarization system whose central contribution is to make complex policy developments browsable by year and topic, with layered summaries intended to reduce information overload while preserving chronology, subtopic structure, jargon explanation, and quantitative cues (Dar et al., 10 Sep 2025).