Comparing the Contribution of Syntactic and Semantic Features in Closed versus Open Domain Question Answering
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
In this paper we analyze the contribution of semantic, syntactic and word similarity of document features in closed and open domain question answering. Semantic similarity is computed as the similarity of the action in the candidate sentence to the action asked in the question, measured using WordNet::Similarity on main verbs. The syntactic similarity feature measures the unifiability of a candidate's parse tree with the question's parse tree. It uses syntactic restrictions as well as lexical measures to compute the unifiability of critical syntactic participants in the parse trees. Finally, the word similarity of the document containing a candidate sentence is computed as the cosine of the angle between the question keywords vector and the document vector. Since the semantic feature is more reliable on content verbs and syntactic similarity is suitable for questions with a subject- verb-object syntactic structure, we only consider questions with a main content verb in our analysis (non-copulative questions). This type comprise 70% of our closed domain and 33% of our open domain test questions. The combination of these three features achieves an MRR of 28% in our closed domain and 23% in open domain. Our analysis shows that the syntactic feature has a significant contribution in both open and closed domains. However, the path-based lch semantic similarity measure we used, only contributes in our closed domain probably because of less variation in the vocabulary and topic. Document IR score on the other hand, has more contribution in open domain, because query keywords are more discriminating in a large document set with a vast vocabulary range.
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.
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 | 0.002 | 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 | 0.000 |
| Open science | 0.001 | 0.001 |
| 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 it