An Entity-Centric Coreference Resolution System for Person Entities with Rich Linguistic Information

Marcos Garcia, Pablo GamalloCOLING 2014

TL;DR

This paper introduces LinkPeople, an entity-centric system for resolving coreferences of person entities in text. It leverages a multi-pass architecture and rich linguistic information, including syntactic constraints and cataphoric analysis, to achieve high accuracy across multiple languages, outperforming several baselines.

What Problem Does This Paper Address?

Coreference Resolution (CR) is a fundamental task in Natural Language Processing (NLP) that involves identifying all mentions in a text that refer to the same real-world entity. While CR is crucial for applications like text summarization and information extraction, resolving coreferences for person entities presents specific challenges. Existing approaches often fall into two categories: mention-pair systems that analyze pairs of mentions, and entity-centric systems that consider all mentions of an entity. The latter have shown better performance, but effectively incorporating linguistic knowledge, especially for person entities and handling complex linguistic phenomena like cataphora (where a pronoun or noun phrase refers to something that appears later in the text), remains an open area. This paper aims to improve person entity coreference resolution by developing an entity-centric system that integrates a wider range of linguistic information.

Methods

LinkPeople employs an entity-centric, multi-pass architecture. This means it considers all mentions belonging to a potential entity when making decisions, and it processes the text in multiple passes, with each pass building upon the results of the previous ones. The system starts by identifying mentions of person entities using NLP tools for Part-of-Speech (PoS) tagging, Named Entity Recognition (NER), and dependency analysis.

The core of LinkPeople is a battery of deterministic modules that are applied sequentially. These modules are sorted from high-precision to high-recall, meaning they prioritize making correct links early on and then work to increase the number of links found. Key novelties include:

  • Cataphoric Analysis: Modules specifically designed to handle cataphoric mentions, particularly definite noun phrases (NPs) and pronouns that refer to entities appearing later in the text. This is based on the observation that definite NPs are not always anaphoric.
  • Syntactic Constraints for Pronoun Resolution: The system incorporates a set of syntactic and morpho-syntactic constraints to improve the precision of pronominal coreference resolution. These constraints help to block incorrect links, for example, by ensuring object pronouns don't corefer with subjects or that possessives don't corefer with the NP they belong to.
  • Multi-Pass Strategy: The system iteratively refines entity clusters. Later passes can leverage the information gathered by earlier modules, allowing for more informed decisions.

The system also includes modules for:

  • String Matching: Exact matching of mention strings.
  • PN Inclusion: Checking if a proper noun phrase is included within another.
  • PN Tokens: Matching based on the order of tokens within proper noun phrases.
  • Head Matching: Linking mentions that share the same head word.
  • Orphan NP: Linking definite NPs that don't contain personal names to previous nominal mentions.
  • Pronominal Resolution: Standard pronoun resolution with syntactic constraints.
  • Pivot Entity: A final module that links orphan pronouns to a prominent entity in the text.

Datasets

The paper evaluates LinkPeople on three corpora: Portuguese, Galician, and Spanish. These corpora are annotated for person entity coreference and follow the SemEval-2010 guidelines. The authors mention that the annotation includes PoS-tagging and dependency information, which were added by tools like FreeLing and DepPattern. The corpora are freely distributed.

Key Insights and Ideas

  • Entity-centricity is key: The paper reinforces the idea that considering an entire entity's mentions, rather than just pairs, leads to better coreference resolution.
  • Linguistic knowledge pays off: Integrating specific linguistic rules, especially for cataphora and syntactic constraints on pronouns, significantly boosts precision and recall.
  • Multi-pass architecture is effective: The sequential application of modules, ordered by precision, allows the system to build upon previous decisions and progressively refine coreference chains.
  • Cataphora is an important, often overlooked, phenomenon: The dedicated modules for cataphora highlight its significance in achieving comprehensive coreference resolution.
  • Syntactic constraints are powerful for pronouns: The paper demonstrates how well-defined syntactic rules can effectively prune incorrect pronoun coreference links.

Key Results

LinkPeople achieved promising results across the three languages, with macro-average F1 scores of approximately 83% (BLANC score) and 81% (CoNLL score). The system consistently outperformed several baseline methods, including "Singletons" (where each mention is its own entity) and "All_in_One" (where all mentions belong to a single entity). The paper also highlights that modules like PN_Inclusion and the combination of Pronominal and Pivot_Ent modules contributed significantly to the overall performance.

Limitations and Future Work

The authors acknowledge several limitations and suggest future work:

  • Error Analysis: The error analysis revealed that a significant portion of errors (46%) are due to missing links between Noun Phrases and/or Proper Nouns. This includes issues with synonym heads, reliance on external real-world knowledge, semantic knowledge (like hyponymy), head modifiers, and spelling differences.
  • Syntactic and Morpho-syntactic Errors: Errors in predicted PoS-tagging and dependency analysis (15.3%) also lead to incorrect coreference links.
  • Long-Distance Pronominal Anaphora: Missing links due to pronouns being too far from their antecedents (11.3%) remain a challenge.
  • Quoted Speech: Coreference within quoted speech (10%) is another area where errors occur.
  • Spurious Links: Incorrect links due to plural mentions and gender agreement issues were also identified.

Future work includes:

  • Improving nominal and pronominal constraints.
  • Developing dedicated handling for plural mentions.
  • Implementing an "inheritance constraint" to prevent merging partial entities if their mentions were blocked by previous modules.
  • Extending the system to resolve coreferences for entity types other than persons.

Summary updated

BibTeX
@inproceedings{garcia2014entity,
  author = {Marcos Garcia and Pablo Gamallo},
  title = {{An Entity-Centric Coreference Resolution System for Person Entities with Rich Linguistic Information}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2014},
}
Figure