A Generic Anaphora Resolution Engine for Indian Languages

Sobha Lalitha Devi, Vijay Sundar Ram, Pattabhi Rk RaoCOLING 2014

TL;DR

This paper introduces a novel, language-independent anaphora resolution engine specifically designed for the morphologically rich and resource-poor Indian languages. By leveraging in-depth morphological analysis and a Conditional Random Fields (CRFs) machine learning approach, the engine aims to overcome the challenges posed by variations in pronoun agreement across languages like Hindi, Bengali, and Tamil.

What Problem Does This Paper Address?

Anaphora resolution, the task of identifying what a pronoun or other referring expression refers to (its antecedent), is crucial for natural language processing (NLP) applications like question answering and summarization. While many anaphora resolution systems exist, they are often developed for specific languages, particularly English. Indian languages, however, present unique challenges: they are morphologically rich, often lack sophisticated pre-processing tools, and exhibit significant variations in how pronouns agree with their antecedents in terms of Person, Number, and Gender (PNG). Building a single, generic engine that can handle these diverse languages without language-specific modules is a significant hurdle. This paper tackles this challenge by proposing a unified approach that accounts for these linguistic variations.

Methods

The core of the proposed solution lies in a two-pronged approach:

  1. In-depth Morphological Analysis: Instead of relying on full parsing, the system utilizes shallow parsing combined with a detailed morphological analyzer. This analyzer extracts crucial information like root words, lexical categories, and, most importantly, PNG characteristics and case markers. This detailed morphological information is key to handling the agreement variations.
  2. Conditional Random Fields (CRFs): A machine learning approach, CRFs are employed for the actual anaphora resolution. The system models this as a binary classification task: for a given pronoun and a potential antecedent (a noun phrase), it predicts whether the noun phrase is the correct antecedent.

The engine operates in two phases:

  • Training: Annotated data with correct antecedent-pronoun pairs is used to train the CRF model. Features are extracted from this data, including:
  • Positional Features: The relative position of the antecedent to the pronoun (e.g., in the same sentence or prior sentences).
  • Syntactic Features: The syntactic role of the noun phrase (subject, object, etc.) and linguistic characteristics like POS tags and suffixes.
  • Constraint Features: Information derived from clause boundaries and Named Entities to constrain possible antecedents.
  • Verb Suffixes: Information about gender agreement attached to verbs.
  • Testing: Unseen text is processed, candidate antecedents are selected based on heuristic rules (which dynamically capture PNG agreement variations), and the trained CRF model predicts the most likely antecedent.

Datasets

The paper evaluates its approach using a dataset provided for the "Anaphora Resolution for Indian Language" tool contest, which was part of ICON 2011. This dataset includes data for three languages: Tamil, Hindi, and Bengali. The dataset was enriched with in-depth morphological analysis by the authors.

Key Insights and Ideas

  • Morphological Richness as an Asset: The paper demonstrates how the inherent morphological richness of Indian languages, often seen as a challenge, can be leveraged to build a language-independent engine. Detailed morphological analysis provides the necessary features to handle cross-linguistic variations.
  • Heuristic Rules for Dynamic Agreement: The use of heuristic rules to dynamically capture PNG agreement variations based on morphological analysis is a clever way to avoid language-specific modules. These rules adapt to the diverse ways gender and number are expressed.
  • Shallow Parsing Sufficiency: The success with shallow parsing highlights that full syntactic parsing might not always be necessary for anaphora resolution, especially in resource-constrained scenarios. This makes the approach more practical for many Indian languages.
  • CRFs for Robust Classification: CRFs prove to be a powerful tool for this task, effectively learning complex patterns from the extracted features to distinguish correct antecedents from incorrect ones.

Key Results

The paper reports evaluation results using Precision, Recall, and F-measure. While specific quantitative results are presented in Table 4 (not fully detailed in the provided text), the authors state that the results are "encouraging" and comparable to other reported works. They note that Bengali showed lower scores, attributing this to the lack of gender distinction in Bengali third-person pronouns and verb agreement, making animacy and syntactic roles more critical.

Limitations and Future Work

  • Bengali Performance: The authors acknowledge lower performance on Bengali, suggesting that further work on animacy features and incorporating more robust syntactic role and Named Entity information might improve results.
  • Resource-Poor Languages: While the approach uses minimal resources (shallow parsing), further exploration into leveraging existing linguistic resources or developing new ones for other Indian languages is a natural next step.
  • Scalability: The engine is designed to be scalable and allow for plug-n-play architectures, implying that extending it to more languages or refining its performance for existing ones is a planned direction.
  • Ambiguity Reduction: The paper suggests that pre-processing data to reduce ambiguity before it reaches the resolution engine can significantly improve performance.

Summary updated

BibTeX
@inproceedings{devi2014generic,
  author = {Sobha Lalitha Devi and Vijay Sundar Ram and Pattabhi RK Rao},
  title = {{A Generic Anaphora Resolution Engine for Indian Languages}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2014},
}

Loading related content...

Figure