Translating HPSG-Style Outputs of a Robust Parser into Typed Dynamic Logic

Manabu Sato, Daisuke Bekki, Yusuke Miyao, Jun'ichi TsujiiACL 2006 · OpenReview

Abstract: Manabu Sato, Daisuke Bekki, Yusuke Miyao, Jun’ichi Tsujii. Proceedings of the COLING/ACL 2006 Main Conference Poster Sessions.

2006.

Introduction

The field of Natural Language Processing (NLP) has long grappled with a fundamental tension between robustness and descriptive adequacy. Statistical parsers, while capable of processing vast amounts of real-world text, often produce shallow semantic representations. Conversely, sophisticated semantic theories like Typed Dynamic Logic (TDL) offer the expressive power to model complex discourse phenomena—such as anaphora and quantification—but have historically lacked the scalability required for large-scale applications. This paper presents a novel framework to bridge this gap by translating the outputs of a high-coverage Head-driven Phrase Structure Grammar (HPSG) parser into TDL.

The Robustness-Adequacy Trade-off

Traditional semantic parsing often relies on First-Order Predicate Logic (FOPL). While effective for sentence-level facts, FOPL struggles with discourse-level dependencies, such as anaphora resolution (e.g., determining what "he" refers to across multiple sentences) and the flow of information through a conversation. Dynamic semantic theories, such as TDL, treat the meaning of a sentence as its "context-change potential"—a function that updates the current state of information. However, the complexity of these theories has made them difficult to pair with robust, wide-coverage syntactic parsers that work on messy, real-world data.

Bridging the Gap: HPSG to TDL

The authors leverage a robust HPSG parser (Miyao et al., 2005) which provides detailed syntactic information through feature structures. To map these structures to TDL, they introduce TDL Extended Structures (TDLESs). A TDLES consists of a TDL expression paired with information that facilitates its composition within the HPSG framework.

Mapping HPSG syntactic structures to TDL semantic composition rules.

As shown in Figure 5, the translation process involves aligning HPSG schemas (like the Subject-Head Schema) with TDL composition rules. This ensures that the semantic derivation mirrors the syntactic structure of the sentence, making the interface between syntax and logic transparent.

The Translation Framework

The core of the system lies in four specialized translation rules designed to handle the nuances of natural language:

  1. Function Application Rule: The standard mechanism for combining a functor with its argument in lambda calculus.
  2. Word Formation Rule: Designed to handle compound nouns where multiple syntactic words correspond to a single semantic unit in TDL.
  3. Non-local Application Rule: Utilizes HPSG's "SLASH" features to handle long-distance dependencies, such as wh-movement in questions.
  4. Unary Derivation Rules: Manages "type-shifting," allowing lexical items to adapt their semantic roles based on the syntactic context without changing their form.
Compositional derivation of the TDL representation for 'a boy ran'.

Figure 3 illustrates the compositional derivation for the sentence "a boy ran." The system builds the logical form incrementally, capturing the existential quantification and the relationship between the entities and events within the dynamic context.

Experimental Results

The system was evaluated on the Penn Treebank (Wall Street Journal corpus), using Section 22 for development and Section 23 for testing. The results demonstrate that deep semantic parsing can indeed be robust:

  • Total Coverage: 88.3% of test sentences were successfully assigned a TDL representation.
  • Word Coverage: 99.6% of words were covered by the lexical mapping.

Error analysis revealed that failures were nearly equally split between missing lexical rules (6.2%) and logic mismatches during composition (5.5%). These figures are competitive with systems using simpler logics, proving that moving to a more complex dynamic logic does not necessarily sacrifice performance.

Discussion and Future Work

While the system successfully generates the logical structures necessary for discourse processing, it does not yet perform automated anaphora resolution. It identifies that a pronoun like "he" requires an antecedent but leaves the selection of that antecedent to a subsequent layer. Future work aims to integrate statistical resolution models into this framework. Additionally, the authors highlight the challenge of evaluating the "correctness" of deep semantic outputs, as gold-standard logical forms for large corpora are currently scarce.

By providing a transparent interface between robust syntax and dynamic logic, this work paves the way for NLP systems that can understand not just individual sentences, but the complex threads of information that weave through entire discourses.

Summary updated

BibTeX
@inproceedings{sato2006translating,
  author = {Manabu Sato and Daisuke Bekki and Yusuke Miyao and Jun'ichi Tsujii},
  title = {{Translating HPSG-Style Outputs of a Robust Parser into Typed Dynamic Logic}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2006},
  url = {https://openreview.net/forum?id=qIX-JJnInTr},
}
Figure