Antelogue: Pronoun Resolution for Text and Dialogue

Eleni MiltsakakiCOLING 2010

TL;DR

Antelogue is a novel, modular pronoun resolution system designed for both text and dialogue. It employs a hybrid approach, combining statistical/ML techniques for feature acquisition with a symbolic algorithm for resolution, and offers an off-the-shelf API for easy integration.

What Problem Does This Paper Address?

Pronoun resolution, the task of identifying the antecedent of a pronoun (e.g., linking "he" to "John"), is a fundamental challenge in Natural Language Processing (NLP). While significant progress has been made for text, dialogue presents unique difficulties. These include the fragmented nature of spoken language affecting POS taggers, disfluencies, the prevalence of first and second-person pronouns, and the need to handle references to entities that are visually present but not explicitly mentioned in the text. Existing systems often struggle with these dialogue-specific nuances.

Methods

Antelogue tackles pronoun resolution in a two-step process:

  1. Feature Acquisition: This phase gathers information about words and the entities they represent. It utilizes a flexible, modular architecture that can incorporate various feature acquisition sub-modules. The current implementation includes:
  2. Sentence and Word Tokenization: Standard techniques adapted to respect speaker turns in dialogues.
  3. Named Entity Recognition (NER): Integration of Stanford's NER tagger.
  4. Gender and Animacy Processing: Uses a corpus to identify gender and animacy, with a confidence metric to handle unreliable data.
  5. Part-of-Speech (POS) Tagging: A POS tagger trained on dialogue data from the English CTS Treebank.
  6. Word Processor: Creates indexed entries for words in the Antelogue repository.
  7. NER Processor: Associates NER tags with words.
  8. Resolution Algorithm: A symbolic algorithm uses the acquired features to make resolution decisions. It communicates with the repository to retrieve feature information and outputs annotated text or an "e-grid" representation. The system prioritizes finding non-pronominal antecedents.

A key aspect of Antelogue is its hybrid approach, combining statistical and machine learning for feature extraction with a symbolic resolution mechanism informed by prior anaphora resolution research and entity salience models.

Datasets

  • English CTS Treebank: Used for training the dialogue POS tagger.
  • Gender Corpus: Used for gender and animacy information.
  • Google n-gram corpus: Used for identifying non-referential instances of "it".
  • Lost TV series dialogues: Used for pre-processing steps to extract dialogues and text.
  • Human corpus: Used for gender information.

The paper does not explicitly mention a newly introduced benchmark dataset for evaluation, but rather discusses evaluation on annotated pronouns.

Key Insights and Ideas

  • Dialogue-Specific Challenges: The paper clearly articulates the unique difficulties of pronoun resolution in dialogue, such as fragmented speech and the need to handle visually salient but textually absent referents.
  • Modular Architecture: The flexible, modular design of Antelogue is a significant advantage, allowing for easy integration of new feature acquisition modules and adaptation to different domains.
  • Hybrid Approach: The combination of statistical/ML for feature extraction and a symbolic algorithm for resolution offers a robust approach, leveraging the strengths of both paradigms.
  • Handling Non-Referential "it": The method for identifying non-referential instances of "it" using Google n-gram data is an interesting technique to distinguish between pleonastic "it" and referential uses.
  • Importance of Unresolvable Declarations: Correctly declaring pronouns as unresolvable is crucial in dialogue, especially for identifying potential cross-modal references (e.g., to visual elements in a movie).

Key Results

  • Antelogue achieves performance comparable to some of the highest reported results, correctly identifying an antecedent or correctly declaring a pronoun unresolvable or non-referential in 85% of 600 annotated pronouns.
  • The text module, specifically for stage directions, exceeds 90% performance due to the relative simplicity of stage directions.
  • The system can process reasonably sized files (approx. 2,000-3,000 words) in seconds, and larger files in a couple of minutes, demonstrating efficient processing.

Limitations and Future Work

  • Current Limitations:
    • The current version does not handle plural pronouns.
  • It does not recognize impersonal uses of singular "you".
  • The gender corpus data is not always reliable, and a confidence metric is used to mitigate this.
  • The text module currently ignores "I" and "you" pronouns.
  • Future Work:
    • Implementing resolution for plural pronouns and impersonal "you".
  • Extending the text module to handle "I" and "you" and a dialogue-style resolution for embedded speech.
  • Further exploring the use of correctly declared unresolvable pronouns as cues for visual search in cross-modal scenarios.
  • Investigating the use of the "e-grid representation file" with models like Barzilay and Lapata's (2008) for evaluating text coherence.
  • The authors plan to release Antelogue as an off-the-shelf software component.

Summary updated

BibTeX
@inproceedings{miltsakaki2010antelogue,
  author = {Eleni Miltsakaki},
  title = {{Antelogue: Pronoun Resolution for Text and Dialogue}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2010},
}
Figure