Analysing Dictionary Definitions of Motion Verbs

Antonietta AlongeCOLING 1992

TL;DR

This paper explores how to automatically extract rich semantic and syntactic information about motion verbs from machine-readable dictionaries (MRDs). By analyzing dictionary definitions, the authors demonstrate a method for identifying verb properties like transitivity, aspect (Aktionsart), and thematic roles, paving the way for more robust computational lexicons.

What Problem Does This Paper Address?

Building comprehensive computational lexicons for Natural Language Processing (NLP) is a significant challenge. Manually creating these lexicons is time-consuming and resource-intensive. While machine-readable dictionaries (MRDs) offer a vast source of lexical information, extracting this knowledge in a structured, usable format for NLP tasks is non-trivial. This paper focuses on motion verbs, a semantically rich and syntactically diverse class, to investigate how to semi-automatically extract crucial lexical information from dictionary definitions, specifically aiming to capture nuances that influence syntactic behavior.

Methods

The core methodology involves analyzing dictionary definitions, particularly the "genus term" (hypernym) and "differentia" (distinguishing part) structure. The authors employ a pattern-matching procedure to identify recurring structures within the "differentia" that correspond to specific semantic categories. This allows for the semi-automatic extraction of:

  • Transitivity/Intransitivity/Reflexivity: Often explicitly coded.
  • Aktionsart (Lexical Aspect): Exploiting the IS-A (taxonomical) hierarchy of dictionaries, where hyponyms inherit aspect from their superordinates unless specific patterns in their definitions indicate otherwise.
  • Unaccusativity/Unergativity: For intransitive verbs, this is identified by the auxiliary verb used (e.g., "essere" in Italian for unaccusatives).
  • Components of Meaning, Typical Subjects, Thematic Roles: Identified by recognizing patterns in the "differentia" that map to semantic categories.
  • Proto-Agent/Proto-Patient Roles: Derived by analyzing verb meanings for contributing properties like volition, sentience, change of state, etc., based on Dowty's (1988) proposal.

The process involves both automated analysis and manual intervention for disambiguation and choosing the correct sense of a genus term.

Datasets

The primary datasets used are:

  • Italian Mono-lingual Dictionaries:
    • Nuovo Dizionario Garzanti (GRZ)
    • Dizionario-Macchina dell'Italiano (DMI)
  • Collins Bi-lingual Italian-English Dictionary

The paper specifically mentions an analysis of verbs within the "muoversi" (to move) taxonomy from GRZ, comprising 244 entries. No new benchmarks are introduced; the focus is on extracting information from existing MRDs.

Key Insights and Ideas

  • Dictionary Structure as a Knowledge Source: The inherent hierarchical and definitional structure of MRDs (genus/differentia) is a powerful, albeit implicit, source of semantic and syntactic information.
  • Semi-Automatic Extraction is Feasible: Combining pattern matching with linguistic knowledge allows for significant automation in extracting complex lexical features, reducing manual effort.
  • Aktionsart Inheritance: The IS-A hierarchy in dictionaries naturally supports the inheritance of lexical aspect, simplifying the classification of many verbs.
  • "Differentia" is Key for Nuance: The distinguishing part of a definition is crucial for capturing specific semantic components (like manner, goal, source) that differentiate verbs and influence their syntactic behavior.
  • Prepositional Phrases Encode Meaning: Specific prepositional phrases (e.g., "a NP," "da NP") are strong indicators of semantic roles like GOAL or SOURCE.
  • Ambiguity Resolution is Necessary: The same lexical patterns (e.g., preposition "a") can encode different semantic categories, requiring explicit lists of idiomatic expressions for disambiguation.
  • Proto-Roles as a Unified Framework: Using Dowty's proto-role framework provides a principled way to analyze verb semantics and connect them to argument structure.

Key Results

The paper demonstrates the successful semi-automatic extraction of:

  • Transitivity, Aktionsart, unaccusativity/unergativity.
  • Components of meaning, typical subjects, and thematic roles for verbs within the "muoversi" taxonomy.
  • Information that allows for the classification of verbs according to Levin & Rappaport's (1991) three classes of intransitive motion verbs (arrive, run, roll).

While specific quantitative results (like accuracy percentages) are not the primary focus, the authors emphasize that this approach represents a "significant saving both in time and resources" compared to manual lexicon building.

Limitations and Future Work

  • Manual Intervention: Some manual work is still necessary, particularly in selecting the correct sense of a genus term when its hyponyms are analyzed.
  • Dictionary Coverage and Consistency: Not all definitions are structured uniformly, and terminology can differ between sources. The authors acknowledge the need to combine information from multiple dictionaries to overcome the limitations of any single source.
  • Protagonist Control: The paper notes a potential contradiction with Levin and Rappaport's hypothesis regarding protagonist control in unaccusative verbs, suggesting further study is needed.
  • Future Work: The authors are working on automating more stages of the analysis and encoding the extracted information into a multi-lingual Lexical Knowledge Base (LKB) using a typed unification system. They also aim to relate their data to data from other dictionaries used within the project.

Summary updated

BibTeX
@inproceedings{alonge1992analysing,
  author = {Antonietta Alonge},
  title = {{Analysing Dictionary Definitions of Motion Verbs}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1992},
}
Figure