Automatic Construction of Nominal Case Frames and its Application to Indirect Anaphora Resolution

Ryohei Sasano, Daisuke Kawahara, Sadao KurohashiCOLING 2004

TL;DR

This paper introduces a novel method for automatically constructing Japanese nominal case frames by combining dictionary definitions with large-scale corpus analysis. These constructed frames are then successfully applied to the challenging task of indirect anaphora resolution, demonstrating their practical utility.

What Problem Does This Paper Address?

Understanding the relationships between words in text is fundamental to natural language processing. While verbs have well-established "case frames" that define their arguments (e.g., who did what to whom), nouns also have implicit relationships with entities they inherently relate to. For instance, "price" is inherently linked to "ticket" or "goods," and "roof" to "house." These relationships, termed "obligatory cases" for nouns, are crucial for resolving indirect anaphora, where a noun refers to an entity not explicitly mentioned but implied by context (e.g., "I bought a ticket. The price was 20 dollars." Here, "price" refers to the "price of the ticket"). Existing methods for identifying these noun-based relationships were limited, often relying on hand-crafted rules or focusing only on specific relations like "part-of." This paper aims to automate the creation of comprehensive nominal case frames to improve indirect anaphora resolution.

Methods

The core of the proposed method lies in a two-pronged approach to analyzing Japanese noun phrases of the form "Nm no Nh" (equivalent to "Nh of Nm" in English).

  1. Dictionary-Based Analysis (DBA): This module leverages existing Japanese dictionaries (specifically, the Reikai Shougaku Kokugojiten) and the NTT Semantic Feature Dictionary. It attempts to match the noun "Nm" with obligatory cases described in the definition of the head noun "Nh." For example, if the definition of "coach" mentions an obligatory case "sport," and "rugby" appears as "Nm," DBA tries to link "rugby" to the "sport" case. Similarity scores are calculated using the semantic feature dictionary to quantify these matches.
  2. Semantic Feature-Based Analysis (SBA): This module uses a set of hand-crafted rules based on semantic features to identify relationships not easily captured by DBA. These rules are designed to detect specific obligatory cases (e.g., HUMAN, TIME, PLACE) and other modifications.

These two analyses are integrated, with DBA taking precedence if its confidence score is high. The results of this analysis are then used to construct preliminary case frames for each noun.

Datasets

  • NTT Semantic Feature Dictionary: A dictionary containing semantic features for approximately 300,000 nouns.
  • Japanese Corpora: Large corpora were used for collecting "Nm no Nh" noun phrases. Specific details on the size and composition are provided in the evaluation section.
  • Relevance-tagged corpus: Used for evaluating the indirect anaphora resolution system.

Key Insights and Ideas

  • Bridging the Gap: The paper successfully bridges the gap between traditional verb-centric case frame analysis and the less-explored nominal case frames, recognizing their parallel importance in text understanding.
  • Integrated Approach: The combination of dictionary definitions (providing explicit semantic roles) and corpus data (revealing actual usage patterns and frequencies) is key to building robust nominal case frames.
  • "Nm no Nh" as a Goldmine: The paper highlights the "Nm no Nh" construction as a rich source of information for identifying noun-based obligatory cases.
  • Practical Application: Demonstrating the effectiveness of the constructed case frames in an indirect anaphora resolution system validates the practical utility of the proposed method.

Key Results

  • Case Frame Construction: The system automatically constructed nominal case frames from large corpora. For a sample of 100 nouns, 68 case frames were created, with 35 nouns having no case frames.
  • Evaluation of Case Frames: Against a hand-created gold standard, the system achieved a precision of 0.829, recall of 0.853, and an F-score of 0.841.
  • Indirect Anaphora Resolution: The preliminary anaphora resolution system, using the constructed case frames, achieved an overall F-score of 0.660. The upper bound for recall was estimated at 84.3%, indicating significant room for improvement.

Limitations and Future Work

  • Coverage: The current system's recall was affected by highly restricted example collection. Future work should include compound noun phrases ("PRODUCTS maker") and other phrasal constructions.
  • Generic vs. Specific Usage: The system struggles with nouns that have both specific and generic usages, leading to incorrect antecedent assignments. Improving this requires better handling of world knowledge and context.
  • Selectional Restrictions: Distinguishing between candidates belonging to HUMAN or ORGANIZATION remains a challenge, suggesting a need for more sophisticated world knowledge integration.
  • Future Directions: The authors plan to expand the coverage of case frames and improve the anaphora resolution system by considering larger contexts and more diverse factors.

Summary updated

BibTeX
@inproceedings{sasano2004automatic,
  author = {Ryohei Sasano and Daisuke Kawahara and Sadao Kurohashi},
  title = {{Automatic Construction of Nominal Case Frames and its Application to Indirect Anaphora Resolution}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2004},
}
Figure