Unsupervised Context-Linking Retriever for Question Answering on Long Narrative Books
Bibliographic record
Abstract
Narrative Question Answering (QA) involves understanding the context, events, and relationships within narrative texts for accurate question answering. However, narrative books impose new challenges while utilizing recent pretrained large language models since such lengthy content requires additional computational costs and leads to performance degradation. Moreover, identifying the most relevant passages for a given question is particularly challenging due to the lack of labeled question-passage pairs for training the retriever. This paper introduces the Unsupervised Context Linking Retriever (UCLR), a novel approach that efficiently retrieves relevant passages from long narrative texts without requiring labeled (question, passage) pairs. UCLR uses an encoder-decoder model to generate synthetic (question, answer) pairs, measuring the relevance of passages by comparing the error between the generated pair and the reference pair, which serves as a synthetic training signal. This method optimizes the retriever to identify passages with sufficient context to accurately reconstruct both the question and the answer, improving retrieval accuracy. UCLR also identifies key events surrounding each passage in the retrieved set and constructs a new set of passages from these key events, enabling coverage of both broader narrative structures and finer details. Experimental results on the NarrativeQA benchmark show that UCLR achieves relative improvements of +8% on the validation set and +5% on the test set, outperforming state-of-the-art unsupervised retrievers. Additionally, the results demonstrate that combining UCLR with a simple reader model outperforms other state-of-the-art readers designed for processing lengthy documents, achieving a relative performance gain of 7.8% on the test set while being 5 times faster as UCLR allows the reader model to focus on a pertinent subset of tokens.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.001 |
| Open science | 0.001 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".