Anaphora for Everyone: Pronominal Anaphora Resolution without a Parser

Christopher Kennedy, Branimir BoguraevCOLING 1996

TL;DR

This paper presents a novel algorithm for resolving pronominal anaphora (like "he," "she," "it") that significantly reduces the reliance on full syntactic parsing. By leveraging shallower linguistic analysis, it makes accurate anaphora resolution accessible to a wider range of NLP applications that can't afford or perform robust parsing.

What Problem Does This Paper Address?

Resolving pronominal anaphora – figuring out which noun phrase a pronoun refers to – is a crucial step for understanding text. Existing high-accuracy methods, like the Lappin and Leass algorithm, heavily depend on detailed syntactic parse trees. However, generating these trees for real-world text is computationally expensive and often unreliable, limiting the practical application of these powerful anaphora resolution techniques. This paper aims to bridge this gap by developing an anaphora resolution system that works with less demanding linguistic input, specifically the output of a part-of-speech tagger augmented with grammatical function information.

Methods

The core of this work is a modification and extension of the Lappin and Leass (1994) algorithm. Instead of relying on full syntactic parses, the authors use a "shallower" linguistic analysis derived from a part-of-speech tagger (LINGSOFT) that also provides grammatical function annotations for each token.

Key methodological adaptations include:

  • Reduced Input: The algorithm operates on sequences of tokens tagged with their part-of-speech and grammatical function, along with their position (offset) in the text.
  • Inferred Constituent Structure: While not explicitly parsed, phrasal units (like noun phrases) are identified using regular expressions and filters applied to the tagged output.
  • Discourse Referents: Entities mentioned in the text are represented as "discourse referents," each carrying information about its text form, referential type, agreement features, grammatical function, and position. Crucially, the explicit configurational relations used by Lappin and Leass are replaced by inferred hierarchical relations based on grammatical function and precedence.
  • Salience Factors: The algorithm uses a set of "salience factors" (e.g., being in the current sentence, grammatical function like subject) to determine the prominence of discourse referents. These factors are assigned numerical weights, and the overall salience of a discourse referent is a sum of these weights.
  • Variable Salience and Decay: The salience of a discourse referent can increase when it becomes an antecedent for an anaphor and decreases over time (decay function) to reflect local prominence.
  • Disjoint Reference Filters: To handle cases where a pronoun cannot refer to certain entities, the algorithm employs filters based on grammatical function and precedence, mimicking the role of syntactic constraints without explicit parsing. These filters address conditions like a pronoun not coreferring with a coargument or a constituent it commands.

Datasets

The paper does not explicitly mention specific datasets used for training. However, the evaluation was performed on a dataset of 27 texts randomly selected from various genres, including press releases, product announcements, news stories, and magazine articles, available as World Wide Web pages. This dataset contained 306 third-person anaphoric pronouns.

Key Insights and Ideas

  • Parsing is Not Always Necessary: The most significant insight is that high-accuracy anaphora resolution can be achieved without the heavy machinery of full syntactic parsing, opening doors for broader application.
  • Grammatical Function and Precedence as Proxies: The paper demonstrates that grammatical function and positional information (precedence) can effectively approximate the role of syntactic structure in determining anaphora resolution and disjoint reference.
  • Salience as a Unifying Principle: The concept of salience, dynamically updated and decayed, serves as a robust mechanism for ranking potential antecedents, even with reduced linguistic input.
  • Adaptability to Diverse Text: By avoiding reliance on specific text genres or deep parsing, the algorithm is designed to be more robust and applicable to a wider variety of real-world texts.

Key Results

The anaphora resolution algorithm achieved an accuracy rate of 75% on a diverse set of 27 texts, resolving 231 out of 306 anaphoric pronouns correctly. This rate is considered comparable to the 85% accuracy reported by Lappin and Leass (1994), especially given the significantly less demanding input analysis. The authors note that the absence of explicit configurational information resulted in only a small number of errors that could be directly attributed to this lack.

Limitations and Future Work

  • Gender Mismatch: A significant source of error (35%) is gender mismatch, largely due to the lack of consistent gender information in the LINGSOFT tagger output. Future work includes augmenting the system with a lexical database for more detailed gender information.
  • Quoted Text: Handling anaphora within quoted passages presents challenges, requiring methods to evaluate quoted speech separately. The authors believe their text segmentation approach could help address this.
  • "Expletive" It: The evaluation excluded certain instances of "expletive" it, and refinement of the patterns for identifying these is ongoing.
  • Connecting Discourse Referents: The paper acknowledges the need for further heuristics to connect discourse referents that refer to the same underlying entity but are not directly linked by the current algorithm (e.g., "company" and "Apple").
  • Improved Salience Factors: The authors suggest that incorporating more explicit morphological and contextual information could further improve accuracy, potentially bringing it closer to Lappin and Leass' results.

Summary updated

BibTeX
@inproceedings{kennedy1996anaphora,
  author = {Christopher Kennedy and Branimir Boguraev},
  title = {{Anaphora for Everyone: Pronominal Anaphora Resolution without a Parser}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1996},
}
Figure