Elamur Research⌕
☰
home / papers / 2002 / Named Entity Recognition using an HMM-based Chunk Tagger

Named Entity Recognition using an HMM-based Chunk Tagger

Summary

The paper addresses English named entity recognition: locating and classifying names of persons, organizations and locations, as well as dates, times, percentages and money amounts. At the time, learning-based systems still lagged handcrafted rule systems by about two points. Zhou and Su adapt their HMM-based chunk tagger, which was the best individual system at CoNLL'2000, by treating each entity as an "NE-Chunk". Unlike the classic HMM in BBN's IdentiFinder, their model factorizes the problem under a mutual information independence assumption. This lets each tag draw on the surrounding context, not just its own word. Each token pairs a word with a composite feature that combines four sources of evidence: surface shape (capitals, digits), semantic classes of trigger words, gazetteer matches, and links to entities already identified elsewhere in the same document. A two-level back-off scheme handles sparse data. On the MUC-6 and MUC-7 NE tasks, the system reaches F-measures of 96.6% and 94.1%. The authors report that these scores beat all published machine-learning and rule-based systems, which suggests that a trainable model can match expert-written rules.

Key points

  • Results: MUC-6 reaches F 96.6 (P 96.3, R 96.9) and MUC-7 reaches F 94.1 (P 93.7, R 94.5). In both cases the MUC dry-run data served for development and the formal test data for testing.
  • Training sizes were about 1330 KB for MUC-6 and 708 KB for MUC-7. A learning curve on MUC-7 shows roughly 90% F at 200 KB of training data and a clear drop at 100 KB.
  • Feature ablation on MUC-7:
    • f1 (shape features) alone gives 77.6.
    • Adding f2 (triggers) gives 87.4.
    • Adding f4 (document macro context) on top of f1+f2 gives 92.9.
    • All four sub-features give 94.1, so the gazetteer (f3) adds only about 1.2 points.
  • Each tag is structured into three parts: a boundary category (0 = whole entity, 1/2/3 = beginning/middle/end), an entity class, and the word feature. Viterbi decoding uses a table of allowed tag transitions to enforce well-formed chunks.
  • Sparseness is handled in two back-off stages. The first falls back through progressively smaller windows of features and words. The second falls back through subsets of the four sub-features, always keeping f1.

Interesting ideas

  • I1. Document-level memory as a feature. The model marks a token when it matches or abbreviates an entity already recognized in the same document. This exploits consistency within a single text and gave the second-largest gain in the ablation. It would transfer easily to alias-heavy domains such as company or place names in trade data.
  • I2. Trigger words grouped by meaning. Instead of raw cue words, important context and in-name words are collected semi-automatically from training data and grouped into semantic classes. This was the single largest gain, and it gives a cheap way to inject lexical knowledge into any sequence labeler.
  • I3. Prioritized, mutually exclusive shape classes. Orthographic and numeric features are checked in a fixed priority order, with a special class for capitalized sentence-initial words. This avoids overlapping features and keeps the sentence-start capitalization ambiguity explicit. Such a scheme is simple to re-implement for another language.
  • I4. Enriched tags plus hard transition constraints. Folding the word feature into the tag gives finer-grained statistics. A validity table blocks impossible boundary sequences during decoding. Any BIO-style tagger could adopt this to guarantee well-formed spans.
  • I5. Reading the ablation critically. The authors explain the gazetteer's small contribution by noting that it mainly helps well-known names that appear without helpful context. When building resources, this is a useful reminder to measure what a lexicon adds beyond contextual cues before investing heavily in it.

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

Topics: Named entity recognition · Part-of-speech tagging

Cited by

79 papers of our library cite it (of the 140884 whose bibliographies we read); 1 of them have a page here, 78 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.3115/1073083.1073163

@inproceedings{zhou2002named,
  title = {{Named Entity Recognition using an HMM-based Chunk Tagger}},
  author = {Zhou, GuoDong and Su, Jian},
  booktitle = {Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics},
  year = {2002},
  doi = {10.3115/1073083.1073163},
  url = {https://doi.org/10.3115/1073083.1073163}
}

References

27 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. J. Aberdeen, D. Day, L. Hirschman, P. Robinson and M. Vilain. MITRE: Description of the Alembic System Used for MUC-6. MUC-6 . Pages141-155. Columbia, Maryland. 1995. find
  2. C. Aone, L. Halverson, T. Hampton, M. Ramos-Santacruz. SRA: Description of the IE2 System Used for MUC-7. MUC-7 . Fairfax, Virginia. 1998. find
  3. S.W. Bennett, C. Aone and C. Lovell. Learning to Tag Multilingual Texts Through Observation. EMNLP'1996 . Pages109-116. Providence, Rhode Island. 1996. find
  4. Daniel M. Bikel, Richard Schwartz and Ralph M. Weischedel. An Algorithm that Learns What's in a Name. Machine Learning (Special Issue on NLP). 1999. find
  5. A. Borthwick, J. Sterling, E. Agichtein, R. Grishman. NYU: Description of the MENE Named Entity System as Used in MUC-7. MUC-7 . Fairfax, Virginia. 1998. find
  6. Andrew Borthwick. A Maximum Entropy Approach to Named Entity Recognition. Ph.D. Thesis. New York University. September, 1999. find
  7. Eric Brill. Transform-based Error-Driven Learning and Natural Language Processing: A Case Study in Part-of-speech Tagging. Computational Linguistics 21(4). Pages543-565. 1995. find
  8. Nancy Chinchor. MUC-6 Named Entity Task Definition (Version 2.1). MUC-6 . Columbia, Maryland. 1995. find
  9. Nancy Chinchor. Statistical Significance of MUC-6 Results. MUC-6 . Columbia, Maryland. 1995. find
  10. Nancy Chinchor. MUC-7 Named Entity Task Definition (Version 3.5). MUC-7 . Fairfax, Virginia. 1998. find
  11. Nancy Chinchor. Statistical Significance of MUC-7 Results. MUC-7 . Fairfax, Virginia. 1998. find
  12. K. Humphreys, R. Gaizauskas, S. Azzam, C. Huyck, B. Mitchell, H. Cunningham, Y. Wilks. Univ. of Sheffield: Description of the LaSIE-II System as Used for MUC-7. MUC-7 . Fairfax, Virginia. 1998. find
  13. G. R. Krupka, K. Hausman. IsoQuest Inc.: Description of the NetOwl TM Extractor System as Used for MUC-7. MUC-7 . Fairfax, Virginia. 1998. find
  14. D. McDonald. Internal and External Evidence in the Identification and Semantic Categorization of Proper Names. In B. Boguraev and J. Pustejovsky editors: Corpus Processing for Lexical Acquisition . Pages21-39. MIT Press. Cambridge, MA. 1996. find
  15. S. Miller, M. Crystal, H. Fox, L. Ramshaw, R. Schwartz, R. Stone, R. Weischedel, and the Annotation Group. BBN: Description of the SIFT System as Used for MUC-7. MUC-7 . Fairfax, Virginia. 1998. find
  16. A. Mikheev, C. Grover, M. Moens. Description of the LTG System Used for MUC-7. MUC-7 . Fairfax, Virginia. 1998. find
  17. A. Mikheev, M. Moens, and C. Grover. Named entity recognition without gazeteers. EACL'1999 . Pages1-8. Bergen, Norway. 1999. find
  18. Morgan Kaufmann Publishers, Inc. Proceedings of the Sixth Message Understanding Conference (MUC-6) . Columbia, Maryland. 1995. find
  19. Morgan Kaufmann Publishers, Inc. Proceedings of the Seventh Message Understanding Conference (MUC-7) . Fairfax, Virginia. 1998. find
  20. L. Rabiner. A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition”. IEEE 77( 2). Pages257-285. 1989. find
  21. Satoshi Sekine. Description of the Japanese NE System Used for MET-2. MUC-7 . Fairfax, Virginia. 1998. find
  22. Erik F. Tjong Kim Sang and Sabine Buchholz. Introduction to the CoNLL-2000 Shared Task: Chunking. CoNLL'2000 . Pages127-132. Lisbon, Portugal. 11-14 Sept 2000. find our card
  23. A. J. Viterbi. Error Bounds for Convolutional Codes and an Asymptotically Optimum Decoding Algorithm. IEEE Transactions on Information Theory. IT(13). Pages260-269, April 1967. find
  24. B. Watson and Tsoi A Chunk. Second Order Hidden Markov Models for Speech Recognition”. Proceeding of 4 th Australian International Conference on Speech Science and Technology . Pages146-151. 1992. find
  25. Yu Shihong, Bai Shuanhu and Wu Paul. Description of the Kent Ridge Digital Labs System Used for MUC-7. MUC-7 . Fairfax, Virginia. 1998. find
  26. Zhou GuoDong, Su Jian and Tey TongGuan. Hybrid Text Chunking. CoNLL'2000 . Pages163-166. Lisbon, Portugal, 11-14 Sept 2000. find
  27. Zhou GuoDong and Su Jian, Error-driven HMM-based Chunk Tagger with Context-dependent Lexicon. EMNLP/ VLC'2000 . Hong Kong, 7-8 Oct 2000. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhou-2002-named-entity-recognition-hmm-based/