Recursively Expandable Abstracts Explained

• Recursively expandable abstracts are an interactive summarization paradigm that lets users reveal detailed content from short abstracts by expanding key phrases recursively.
• The system utilizes dense embeddings, few-shot prompting, and vector retrieval to dynamically extract and present contextual information.
• Empirical evaluations reveal reduced navigation effort and enhanced user satisfaction through verified, context-aware expansions.

Recursively expandable abstracts are an interaction paradigm designed to bridge the "cognitive gap" between the brevity of scientific abstracts and the detailed exposition in full-text research papers. Inspired by concepts like StretchText, recursively expandable abstracts allow users to interactively expand specific spans or entities within an abstract to reveal additional, contextually grounded information from the source document, with the expansion process itself being recursively applicable at arbitrary depth. This architecture supports low-effort, in-context information seeking and enables scholars to conduct just-in-time exploration of long-form scientific content without repeatedly switching to the full text (Fok et al., 2023).

1. Motivation and Conceptual Foundations

The primary motivation for recursively expandable abstracts lies in the observation that scientific abstracts distill lengthy papers into succinct summaries, often omitting operational definitions, experimental specifics, or nuanced motivations. This omission forces researchers into inefficient navigation: upon encountering an inadequately defined term or concept, readers must dive into full-text PDFs, incurring cognitive overhead and loss of navigation context.

Recursively expandable abstracts are inspired by Ted Nelson’s StretchText (1980s), which proposed in-situ toggling of detail levels. Unlike static abstracts, these interfaces allow users to dynamically "stretch" any identified entity or user-highlighted span to reveal synthesized expansions. Each expansion can itself be recursively expanded, forming a thread-like, hierarchical exploration of the source material while maintaining user orientation and promoting depth-first curiosity (Fok et al., 2023).

2. System Architecture and Algorithmic Pipeline

The Qlarify system operationalizes recursively expandable abstracts using a pipeline composed of document processing, entity extraction, retrieval, and generation components.

A. Document Processing (Offline)

• Convert PDF to plain text and extract metadata via GROBID.
• Segment the full text into overlapping "chunks" (windows of 3 sentences, 2-sentence overlap) to maintain contextual continuity.
• Compute dense sentence embeddings for each chunk (using all-mpnet-base-v2) and index in a vector database (e.g., Elasticsearch).

B. Expandable Entity Extraction (Offline & On-the-Fly)

• Use few-shot prompting (GPT-4) to identify and underline abstract spans that are likely candidates for expansion (technical terms, underspecified concepts).
• Dry-run expansion questions for each candidate and prune spans with no retrievable answer.
• The system generates clarifying questions for retained entities, visually marking them for user interaction.

C. Interaction, Retrieval, and Expansion (On-Demand)

• Upon user interaction (highlight or click on entity), present a question palette with static options (Define, Expand, Why) and a dynamically suggested LLM query.
• On selection, embed the question and retrieve the top-12 most similar text chunks (cosine similarity), forming the context $C$.
• Construct a generation prompt combining instructions, context examples, $C$, and the question $Q$.
• Call GPT-3.5-turbo (temperature=0) to synthesize a 1–3 sentence expansion.
• Attribute each expansion to the single most relevant paragraph by cosine similarity and display inline, with further drill-down to the original PDF passage supported.

The recursive nature is maintained by dynamically underlining new expandable entities within each LLM-generated expansion, allowing exploration trees of arbitrary depth.

3. Mathematical and Retrieval Formulation

Dense retrieval underpins the context selection process for generation: $$\text{sim}(q, c_i) = \frac{q \cdot c_i}{\|q\|\;\|c_i\|}$$ where $q$ is the question embedding and $c_i$ is a chunk embedding. The use of overlapping chunk windows addresses boundary artifacts common in naive segmentation strategies.

Few-shot prompting techniques are employed to steer LLM generations toward concise, precise, and locally grounded responses. Generation is constrained to short expansions for interface tractability and user efficiency.

4. Attribution, Verification, and Trust

Every synthesized expansion is attributed to an extractive evidence passage ($p^*$), identified as the most similar paragraph in the dense retrieval index: $p^* = \arg\max_i \cos(e_i, \text{embed}(A))$ This architecture enforces multi-level transparency: each expansion points to explicit documentary evidence, which can be viewed in situ (pop-up card) or in the original PDF source. Such traceability minimizes hallucination risk and allows manual validation.

Empirical studies indicate that user trust increases when supporting attribution is available. The system’s attribution layer, coupled with direct links to original context, provides just-in-time verification in the research workflow (Fok et al., 2023).

5. Empirical Evaluation and User Studies

Multiple user studies validate the effectiveness of recursively expandable abstracts:

• In interview studies (N=9), users initiated an average of 20.8 expansions, with 58% forming threaded (recursive) explorations. LLM-suggested questions accounted for 40% of user queries, with high perceived trust and satisfaction when attribution was present.
• Field deployment (VLDB’23, N=275) showed each user generated on average 3.3 expansions, and 27.7% were nested. The majority of expansions were triggered on pre-selected entities.
• Controlled, within-subject evaluation (N=12) comparing Manual (abstract+PDF), QA textbox, and Qlarify conditions demonstrated that Qlarify users submitted more questions (15.0 vs. 9.6), opened PDFs less frequently, and reduced overall PDF reading time (201s vs. 539s). Satisfaction (Likert 6.2 vs. 4.8), information gain, and depth of exploration were all significantly higher with recursively expandable abstracts.

These results demonstrate the paradigm’s ability to increase the number of clarification questions asked, reduce search effort, and raise user satisfaction and effective information gain (Fok et al., 2023).

6. Limitations and Open Challenges

Several limitations and avenues for improvement are documented:

• Hallucination: 12.5% of expansions exhibited minor factual inaccuracies; attribution decreases but does not eliminate this phenomenon.
• Visual overload: Frequent underlining may distract or bias attention, potentially derailing rapid triage. Configurable interface options (hiding expandable entities, sidebar view) are suggested as mitigations.
• Scope restriction: Limiting questions to predefined abstract entities may prevent broader context-seeking. Combining the palette with free-form QA boxes is proposed to overcome this.
• Beyond single documents: Integrating citation networks or expanding across multiple papers could further extend the utility of recursively expandable abstractions.
• Modal expansion: Incorporating figures, tables, video summaries, or personalizing expansions based on user history are identified as promising directions.
• Cross-domain generalization and accessibility: Adapting the paradigm to domains like legal, medical, or policy documents and enabling assistive reading is highlighted as future work.

7. Connections to GASP and Multihop Abstraction

A related direction is the GASP (Generating Abstracts from Abstracts of Cited Papers) task (Zanzotto et al., 2020), which models the creative process of synthesizing a target paper’s abstract from cited abstracts via sequence-to-sequence architectures. The GASP framework can be extended recursively, forming multi-hop abstractions by iteratively expanding upon previously generated abstracts:

• The iterative pipeline augments the source set at each hop, allowing simulation of layered abstraction (e.g., "abstracts-on-abstracts"), akin to multi-level recursively expandable abstracts but modeling content synthesis rather than interaction.
• Beam-search can be used for abstract generation, incorporating both likelihood and novelty metrics at each recursion level.

A plausible implication is that interactive, recursively expandable abstracts and generative, recursive abstraction tasks provide complementary approaches to the cognitive navigation and synthesis of scientific literature, both aiming to bridge the growing complexity of scientific reporting and user demands for explicability and efficient triage (Zanzotto et al., 2020).

---

Recursively expandable abstracts represent a verifiable, mixed-initiative framework that enhances scholarly navigation through just-in-time, context-aware expansion of academic content, empirically shown to support information needs with greater efficiency and user satisfaction compared to static summaries or unrestricted document search (Fok et al., 2023).