--- title: "A General Framework for Distributional Similarity — Weeds 2003" description: "Weeds and Weir propose co-occurrence retrieval (CR), which treats word similarity as a retrieval problem. Cited by 72 papers here." authors: ["Julie Weeds", "David Weir"] year: 2003 venue: "Proceedings of the 2003 Conference on Empirical Methods in Natural Language Processing" source: "https://aclanthology.org/W03-1011/" licence: "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)" licence_url: "https://creativecommons.org/licenses/by-nc-sa/3.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/weeds-2003-general-framework-distributional-similarity/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/weeds-2003-general-framework-distributional-similarity/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2003](https://md.research.elamur.ai/free/en/papers/2003/) / A General Framework for Distributional Similarity # A General Framework for Distributional Similarity **Authors:** [Julie Weeds](https://md.research.elamur.ai/free/en/authors/julie-weeds/) and [David Weir](https://md.research.elamur.ai/free/en/authors/david-weir/) · **Year:** 2003 · **Venue:** Proceedings of the 2003 Conference on Empirical Methods in Natural Language Processing · **Licence:** [CC BY-NC-SA 3.0 (ACL Anthology, before 2016)](https://creativecommons.org/licenses/by-nc-sa/3.0/) **Links:** [Publisher's page](https://aclanthology.org/W03-1011/) · [PDF](https://aclanthology.org/W03-1011.pdf) · [OpenAlex](https://openalex.org/W2101941521) **Citations:** 126 citations (OpenAlex, 2026-10-14) · [72](https://md.research.elamur.ai/free/en/papers/weeds-2003-general-framework-distributional-similarity/cited-by/) in our library, of the 140,884 papers whose bibliographies we read ## Summary Distributional similarity measures are widely used to find words with related meanings, but the literature offers many competing formulas with little to say about why one should beat another. Weeds and Weir propose co-occurrence retrieval (CR), which treats word similarity as a retrieval problem. They ask how well the contextual features of one word "retrieve" the features of another, and they score this with precision and recall, borrowed from information retrieval. The two scores are combined through tunable parameters. Different ways of weighting features yield a family of concrete models, and several known measures turn out to be special cases of it. The authors test these models on nouns from the BNC in two ways: agreement with a WordNet-based similarity, and a pseudo-disambiguation task. Tuned CR models match or outperform standard measures. The best balance of precision and recall depends on the task and on word frequency. This matters because the choice of measure becomes a principled matter of parameter settings rather than trial and error. ## Key points - **Asymmetric framing.** Precision checks how many of the neighbour word's contexts are shared with the target word. Recall checks how many of the target's contexts the neighbour covers. Because the two can differ, similarity becomes asymmetric. - **Two combination parameters.** One parameter (γ) mixes a harmonic mean of precision and recall with a weighted arithmetic mean. A second parameter (β) sets the relative weight of precision versus recall. - **Model variants.** - Additive (type-based) models simply count shared features. - Difference-weighted (token-based) models compare how strongly each feature is associated with each word. - Feature weights can come from mutual information or from t-test scores. - **Existing measures recovered.** Measures such as Lin's similarity correspond to particular settings within the CR space. - **Evaluation.** Nouns from the BNC, split into high- and low-frequency groups, were assessed by closeness to WordNet-derived neighbours and by a pseudo-disambiguation task. ## Interesting ideas - **Precision/recall lens for any feature-overlap similarity.** Splitting similarity into precision and recall makes it easier to diagnose why a measure behaves as it does. This could transfer to comparing documents, user profiles or trade-partner profiles. - **One parameterised family instead of a zoo of formulas.** Placing known measures as points in a shared parameter space lets you tune the setting per task instead of picking a measure ad hoc. This is useful when building embeddings or similarity resources for Ukrainian. - **Deliberate asymmetry.** If one word's contexts largely cover another's, the first word may be the more general term. This hints at a cheap signal for detecting hypernym-like relations or more general entities. - **Evaluating separately by frequency band.** Reporting results for frequent and rare words separately shows that the best setting can shift with frequency. That is a worthwhile habit for low-resource or long-tail vocabularies. - **Pseudo-disambiguation as a resource-free benchmark.** Testing whether a model can pick the true co-occurring context over a distractor needs no manual annotation. This makes it practical for languages with few gold-standard resources. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* Fields: [Natural language processing](https://md.research.elamur.ai/free/en/shelves/nlp/) ## Institutions - [University of Sussex](https://md.research.elamur.ai/free/en/universities/university-of-sussex/) — University, Brighton, United Kingdom · 15 works · 3,064 citations ## Cite Cite as: Julie Weeds and David Weir. 2003. A General Framework for Distributional Similarity. Proceedings of the 2003 Conference on Empirical Methods in Natural Language Processing. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{weeds2003general, title = {{A General Framework for Distributional Similarity}}, author = {Julie Weeds and David Weir}, booktitle = {Proceedings of the 2003 Conference on Empirical Methods in Natural Language Processing}, year = {2003}, doi = {10.3115/1119355.1119366}, url = {https://doi.org/10.3115/1119355.1119366} } ``` - [← previous Exploiting Parallel Texts for Word Sense Disambiguation: An Empirical Study](https://md.research.elamur.ai/free/en/papers/ng-2003-exploiting-parallel-texts-word-sense/) - [next → Coreference Resolution Using Competition Learning Approach](https://md.research.elamur.ai/free/en/papers/yang-2003-coreference-resolution-competition-learning-approach/) - [↑ back to papers of 2003](https://md.research.elamur.ai/free/en/papers/2003/)