--- title: "Finding Parts in Very Large Corpora — Berland 1999" description: "Berland and Charniak tackle the automatic discovery of meronymy (part-of relations between words) from raw, unannotated text, a task Hearst had reported…" authors: ["Matthew Berland", "Eugene Charniak"] year: 1999 venue: "Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics" source: "https://aclanthology.org/P99-1008/" doi: "10.3115/1034678.1034697" 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/berland-1999-finding-parts-very-large-corpora/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/berland-1999-finding-parts-very-large-corpora/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [1999](https://md.research.elamur.ai/free/en/papers/1999/) / Finding Parts in Very Large Corpora # Finding Parts in Very Large Corpora **Authors:** Matthew Berland, [Eugene Charniak](https://md.research.elamur.ai/free/en/authors/eugene-charniak/) · **Year:** 1999 · **Venue:** Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics · **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/P99-1008/) · [PDF](https://aclanthology.org/P99-1008.pdf) · [DOI 10.3115/1034678.1034697](https://doi.org/10.3115/1034678.1034697) · [OpenAlex](https://openalex.org/W2160587453) **Citations:** 459 citations (openalex-snapshot · 2026-10-14) · [101](#cited-by) in our library, of the 140884 papers whose bibliographies we read ## Summary Berland and Charniak tackle the automatic discovery of meronymy (part-of relations between words) from raw, unannotated text, a task Hearst had reported failing at. Given a seed noun such as "car", their system collects candidate part words from the roughly 100-million-word North American News Corpus (NANC). It uses two lexico-syntactic patterns: the possessive ("building's basement") and an "of the/a" construction ("basement of a building"). Candidates ending in typical quality suffixes are discarded. The rest are ranked with a significance-based measure called sigdiff, applied to strongly conditioned probabilities. Five naive human judges decided by majority which proposed words were real parts. Across six seed words, about 55% of the top 50 candidates and about 70% of the top 20 were judged correct. A simple co-occurrence baseline reached only 8%. The authors argue that the output can be checked by a person in minutes and used to extend WordNet or to build a rough semantic lexicon. They attribute their improvement over Hearst to the very large corpus and more careful ranking statistics. ## Key points - Five candidate patterns were derived from corpus sentences containing the known pair basement/building. On the seed "car", only pattern A (possessive) and pattern B ("of" + the/a + optional modifiers) produced mostly genuine parts. The "in" patterns and the bare plural "of" pattern picked up passengers, contents and counting nouns such as "truckload". - Processing the whole NANC (~100M words) took about four hours, and about 90% of that time went to POS tagging. Parts and wholes are represented by head nouns only, which sometimes fails (e.g., "conditioner" instead of "air conditioner"). - Four ranking variants were compared: Dunning's log-likelihood versus Johnson's sigdiff test, each with loose or strong conditioning. Sigdiff with strong conditioning gave the most specific lists. The other variants pushed up generic words like "back", "cost" or "use". - Evaluation covered six seeds: book, building, car, hospital, plant, school. Each of the five judges rated 100-word lists in which 50 words were system output. Results were about 55% correct in the top 50 and about 70% in the top 20. The ambiguous seed "plant" yielded parts for both senses but gave the weakest list. - Against WordNet's parts of "car", "engine" and "door" were missing from the top 20 but appeared before rank 100. All WordNet car parts except "tailfin" showed up within the top 500. ## Interesting ideas - **Seed-pair pattern mining.** Take one known related pair, collect sentences where both words occur near each other, and turn the recurring constructions into extraction patterns. Then test each pattern separately on a single seed before committing to it. This is a cheap way to bootstrap extractors for any semantic relation in a language without labeled data. - **Penalizing evidence that is strong but thin.** A conditional probability like p(whole | part) equals 1 for any word seen only once. A significance-based score such as sigdiff weighs both the size of the difference and the amount of supporting data, which makes it a sensible choice for ranking any sparse co-occurrence statistics. - **Cheap surface filters for unwanted categories.** Removing candidates by suffix (-ing, -ness, -ity) strips out abstract qualities with almost no cost. Analogous morphological filters could clean up candidate lists in other languages. - **Cross-pattern consistency as a noise signal.** The authors note that idioms like "a jalopy of a car" occur in one pattern but never in the other. Requiring support from several independent patterns could therefore filter spurious matches once counts are large enough. - **Human-in-the-loop evaluation design.** Mixing system output with other words and taking the majority vote of judges who don't know the study's goal gives a workable reference when the target relation lacks a crisp definition. It also matches the intended use, where a person quickly vets a ranked list before adding it to a lexical resource. *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/) ## Cited by Cited by 101 papers of our library cite it (of the 140884 whose bibliographies we read); 11 of them have a page here, 90 not yet. - [Unsupervised Models for Named Entity Classification](https://md.research.elamur.ai/free/en/papers/collins-1999-unsupervised-models-named-entity-classification/) - [Bridging Resolution: A Survey of the State of the Art](https://md.research.elamur.ai/free/en/papers/kobayashi-2020-bridging-resolution-survey-state-art/) - [Semantic Class Learning from the Web with Hyponym Pattern Linkage Graphs](https://md.research.elamur.ai/free/en/papers/kozareva-2008-semantic-class-learning-web-hyponym/) - [Comparing Knowledge Sources for Nominal Anaphora Resolution](https://md.research.elamur.ai/free/en/papers/markert-2005-comparing-knowledge-sources-nominal-anaphora/) - [Using the Web in Machine Learning for Other-Anaphora Resolution](https://md.research.elamur.ai/free/en/papers/modjeska-2003-web-machine-learning-other-anaphora/) - [Espresso: Leveraging Generic Patterns for Automatically Harvesting Semantic Relations](https://md.research.elamur.ai/free/en/papers/pantel-2006-espresso-leveraging-generic-patterns-automatically/) - [Acquiring Lexical Knowledge for Anaphora Resolution](https://md.research.elamur.ai/free/en/papers/poesio-2002-acquiring-lexical-knowledge-anaphora-resolution/) - [Learning to Resolve Bridging References](https://md.research.elamur.ai/free/en/papers/poesio-2004-learning-resolve-bridging-references/) - [Extracting Product Features and Opinions from Reviews](https://md.research.elamur.ai/free/en/papers/popescu-2005-extracting-product-features-opinions-reviews/) - [Extracting and evaluating general world knowledge from the Brown Corpus](https://md.research.elamur.ai/free/en/papers/schubert-2003-extracting-evaluating-general-world-knowledge/) - [Similarity of Semantic Relations](https://md.research.elamur.ai/free/en/papers/turney-2006-similarity-semantic-relations/) ## Cite The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{berland1999finding, title = {{Finding Parts in Very Large Corpora}}, author = {Berland, Matthew and Charniak, Eugene}, booktitle = {Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics}, year = {1999}, doi = {10.3115/1034678.1034697}, url = {https://doi.org/10.3115/1034678.1034697} } ``` - [← previous Discourse Relations: A Structural and Presuppositional Account Using Lexicalised TAG](https://md.research.elamur.ai/free/en/papers/webber-1999-discourse-relations-structural-presuppositional-account/) - [next → Functional Centering – Grounding Referential Coherence on Information Structure](https://md.research.elamur.ai/free/en/papers/strube-1999-functional-centering-grounding-referential-coherence/) - [↑ back to papers of 1999](https://md.research.elamur.ai/free/en/papers/1999/)