Introduction
In the field of Natural Language Processing (NLP), a persistent challenge known as the "semantic dilemma" exists: while syntactic parsing tools have become highly robust, the automatic acquisition of semantic knowledge remains difficult. This paper presents a novel method for bridging this gap by automatically acquiring thematic knowledge—semantic roles such as Agent, Goal, or Instrument—using existing syntactic parsers and a set of linguistic heuristics. By leveraging syntactic clues as a primary filter and resolving ambiguities through minimal human interaction or corpus analysis, the proposed system achieves an 86% accuracy rate in building thematic lexicons.
The Semantic Dilemma
Syntactic structures describe the arrangement of words in a sentence, but they do not inherently convey meaning. For instance, in the sentence "John inherited a million dollars," a standard syntactic parser identifies "John" as the subject. However, semantically, John is not an "Agent" performing an action; he is a "Goal" or recipient. Traditionally, mapping these syntactic positions to semantic roles has required extensive manual labor or domain-specific rules that lack scalability. This study explores whether syntactic resources can be used to bootstrap a domain-independent semantic dictionary.
Methodology
The authors implement a three-step pipeline designed to extract thematic structures with high precision and minimal manual intervention.
1. Syntactic Filtering
The system first applies a set of "Syntactic Clues" to narrow down the potential thematic roles for a given constituent. These clues are based on properties such as constituent type (e.g., Noun Phrase), animacy, and sentence position (Subject vs. Object). For example, an inanimate Noun Phrase in the object position is statistically unlikely to be an "Agent" but highly likely to be a "Theme" or "Result."
2. Heuristic-Guided Resolution
To resolve remaining ambiguities after the initial filter, the system applies six linguistic heuristics derived from established theories by Jackendoff and Gruber:
- Volition/Imperative Heuristics: These confirm if a subject can act intentionally, helping to identify an Agent.
- Thematic Hierarchy: This rule dictates that certain roles (like Agent) must occupy "higher" positions in a sentence structure than others (like Theme).
- Uniqueness Heuristic: This ensures that no two arguments within a single clause are assigned the same thematic role.
3. Validation
When heuristics alone cannot resolve an ambiguity, the system employs two strategies: it can generate a test sentence for a human "trainer" to validate with a simple yes/no response, or it can check for the existence of similar structures within a large corpus to find statistical support for a specific role assignment.
Experimental Setup and Datasets
The methodology was evaluated using the PENN Treebank, specifically focusing on argument packets extracted from four distinct sub-corpora. These datasets represent a variety of linguistic styles, from the imperative-heavy commands of ATIS to the narrative reports found in MUC datasets.
Key Results and Discussion
The empirical evaluation yielded several significant insights:
- High Accuracy: The system achieved an average accuracy of 86% for acquired argument structures across the tested corpora.
- Efficiency of Interaction: The system is remarkably efficient; for every 100 successfully acquired structures, it required only 113 simple queries to a human trainer. This indicates that most roles are identified automatically, with roughly one question needed per verb.
- Corpus-Specific Performance: The ATIS corpus showed the lowest levels of ambiguity. This is attributed to its high volume of imperative sentences, which allowed the "Imperative Heuristic" to function with near-perfect efficacy in identifying Agents.
Conclusion and Future Work
This research demonstrates that thematic structures are deeply linked to syntactic structures. By treating semantic acquisition as an engineering task informed by universal linguistic constraints rather than domain-specific biases, the authors have created a scalable path toward automated semantic understanding.
However, limitations remain. The system occasionally struggled with "accompaniment" roles and subtle distinctions between "Experiencer" and "Agent" roles, as both often appear as animate subjects. Future work may involve integrating these domain-independent heuristics with domain-specific semantic biases to further refine accuracy in specialized applications like machine translation.
Summary updated