Simple Coreference Resolution with Rich Syntactic and Semantic Features
Summary
The paper shows that a deliberately simple, deterministic coreference system can compete with much more complex learned models if it uses good syntactic and semantic information. Each mention is linked to an antecedent chosen with a small set of modular components rather than a trained pairwise classifier. On the syntactic side, the system uses constraints read off the parse tree, such as appositives and predicate nominatives, and ranks candidate antecedents by their position in the tree instead of by flat surface distance. On the semantic side, it filters candidate pairs for compatibility using head-word relations mined in an unsupervised way from a large unlabeled corpus. The authors report that this simple system was competitive with, and in places better than, the state-of-the-art supervised and unsupervised systems of the time on standard coreference benchmarks. The paper was influential in showing that much of the work in coreference is done by syntax and lexical semantics rather than by elaborate learning.
How Mova uses it
- the development notes of
corefandcoref/src/sieve.rs: Mova's coreference module is a deterministic multi-pass sieve (architecture after Raghunathan et al. 2010 and Lee et al. 2013). From this paper it takes the principle that candidates higher in the parse tree come first. Candidate antecedents are sorted by the depth of their head (root = 0), then by position, in the Hobbs-style traversal order. - Tree depth is precomputed per sentence (
coref/src/doc.rs), so the ordering is an integer sort, which keeps every linking decision explainable.
Effectiveness in Mova
One ordering principle inside the coreference sieve. Not measured separately; there is no ablation of the depth ordering on its own.
👨🔬💥
Library summary
Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)" does not allow redistributing the text itself. Original — aclanthology.org/D09-1120.pdf
Summary
Haghighi and Klein ask how far coreference resolution can go without learned models or explicit discourse modelling. Their system is fully deterministic and splits the task into independent parts. A syntactic module parses documents with the Stanford parser and maps each mention to a parse-tree node. Its conservative rules then force or forbid certain links, for example appositive configurations. A semantic module filters candidate antecedents using agreement features and compatibility lists for headword pairs, which are mined from large unlabeled parsed corpora (BLIPP newswire and WIKI abstracts). Among the remaining candidates, the system picks the one closest in parse-tree distance and merges links by transitive closure. It needs no coreference-annotated training data and has no tunable parameters. On ACE 2004 and MUC-6 test sets, it beats all unsupervised systems and approaches supervised state-of-the-art results. An error analysis shows that most remaining mistakes come from missing semantic knowledge or flawed syntactic structure, not from discourse effects. This makes the system a strong, easily reproducible baseline and suggests where future work should focus.
Key points
- Development progression (pairwise F1 on ACE2004-ROTH-DEV):
- A flat-distance, head-match baseline (BASE-FLAT) scores 48.9.
- Switching to shortest parse-tree path distance (BASE-TREE) raises this to 51.7.
- The same switch reduces pronoun antecedent errors from 51.3% to 42.5%.
- Unlabeled resources for mining semantic compatibility:
- BLIPP: 1.8 million newswire sentences parsed with the Charniak (2000) parser.
- WIKI: 25k English Wikipedia abstracts parsed with the Klein and Manning (2003) parser.
- Evaluation setup: Results are reported with pairwise F1, MUC, B³ and CEAF on ACE2004-CULOTTA-TEST, ACE2004-NWIRE and MUC-6-TEST.
- Main comparisons:
- The best B³ score of 79.0 is broadly comparable to the supervised approaches of Culotta et al. (2007) and of Bengston and Roth (2008), and it is reached without gold NER tags or gold mention types.
- On MUC-6-TEST, the system outperforms Poon and Domingos (2008) and Finkel and Manning (2008) on all comparable measures.
- On ACE2004-NWIRE, it also beats Poon and Domingos (2008).
- Error analysis:
- Nominal mentions cause the largest number of errors, often because the system leaves a nominal unlinked for lack of semantic knowledge.
- Overall, semantic and syntactic compatibility problems clearly outweigh pragmatic or discourse-related errors.
Cite
@inproceedings{haghighi2009simple,
title = {{Simple Coreference Resolution with Rich Syntactic and Semantic Features}},
author = {Haghighi, Aria and Klein, Dan},
booktitle = {Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing},
year = {2009},
doi = {10.3115/1699648.1699661},
url = {https://doi.org/10.3115/1699648.1699661}
}References
- B. Amit and B. Baldwin. 1998. Algorithms for scoring coreference chains. In MUC7. find our card
- Eric Bengston and Dan Roth. 2008. Understanding the value of features for corefernce resolution. In Empirical Methods in Natural Language Processing. find
- E. Charniak. 2000. Maximum entropy inspired parser. In North American Chapter of the Association of Computational Linguistics (NAACL). find
- Mike Collins. 1999. Head-driven statistical models for natural language parsing. find our card
- A Culotta, M Wick, R Hall, and A McCallum. 2007. First-order probabilistic models for coreference resolution. In NAACL-HLT. find
- Pascal Denis and Jason Baldridge. 2007. Global, Joint Determination of Anaphoricity and Coreference Resolution using Integer Programming. In HLT-NAACL. find
- Jenny Finkel and Christopher Manning. 2008. Enforcing transitivity in coreference resolution. In Association of Computational Linguists (ACL). find
- Jenny Finkel, Trond Grenager, and Christopher Manning. 2005. Incorporating non-local information into information extraction systems by gibbs sampling. In ACL. find
- Barbara J. Grosz, Aravind K. Joshi, and Scott Weinstein. 1995. Centering: A framework for modelling the local coherence of discourse. find
- Aria Haghighi and Dan Klein. 2007. Unsupervised coreference resolution in a nonparametric bayesian model. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics. Association for Computational Linguistics. Comment 2 general appos. patterns 2 cataphora find