An Application of Computer Techniques to Analysis of the Verb Phrase in Hindi and English: A Preliminary Report

L. M. Khubohandani, W. W. GloverCOLING 1969

TL;DR

This paper explores using computer techniques to automatically analyze verb phrases in Hindi and English. By applying a classification criterion based on word context, the authors demonstrate a method for identifying patterns and classes within verb phrases, laying groundwork for more sophisticated linguistic analysis.

What Problem Does This Paper Address?

Understanding the structure and patterns of verb phrases is fundamental to natural language processing and linguistic analysis. Manually identifying and classifying these phrases across different languages is a time-consuming and complex task. This paper addresses the challenge of automating this process by developing computer-based methods to analyze verb phrases, specifically in Hindi and English, with the goal of discovering underlying grammatical patterns and classes. The authors aim to move beyond simple word occurrences to understand how words function together in phrases.

Methods

The core of the methodology involves using a computer program to analyze a corpus of verbal phrases. The program's task is to group words that appear in similar contexts into "classes." Once words are assigned to classes, the program substitutes the class reference for the actual word in the data. This substitution process helps to reveal underlying phrase patterns, where the "filler class" for each part of the pattern is defined by the class of word that occupies that slot.

Two main criteria for defining "context" were explored:

  1. "Identical context one-deep on both sides": This criterion considers the immediate neighbors of a word to determine its context.
  2. "Identical context two-deep on both sides": This more stringent criterion expands the context to include words two positions away from the target word.

Datasets

The paper utilizes two primary sets of data:

  • Hindi Corpus: A corpus of 65 verbal phrases was initially used, later extended to 248 phrases of three words or more.
  • English Paradigm: A "complete paradigm" of 112 sentences was used, described as highly restricted data.

No new benchmarks were introduced; rather, existing linguistic data was used for analysis.

Key Insights and Ideas

  • Context is Key: The effectiveness of the classification heavily relies on how "context" is defined. A broader definition of context can lead to more nuanced classifications but may also introduce more complexity.
  • Language Differences: The "one-deep" context criterion proved "quite satisfactory" for Hindi, suggesting that immediate word neighbors are significant in its verb phrase structure. However, the same criterion was less successful for English, hinting at potentially different structural properties or the need for more context in English.
  • Hierarchical Classification: The process of grouping words into classes and then using these classes to define patterns suggests a hierarchical approach to linguistic analysis, moving from individual words to phrase structures.
  • The Power of Substitution: Replacing words with their class references is a clever way to abstract away from specific vocabulary and reveal underlying grammatical structures.

Key Results

  • Hindi (Criterion 1: "one-deep"): 25 classes were formed from 65 phrases. At least one word was classified in 62% of phrases, and all words were classified in 15% of phrases.
  • English (Criterion 1: "one-deep"): 24 patterns and 15 classes were obtained from 112 sentences.
  • Hindi (Criterion 2: "two-deep"): This criterion was less promising. With an extended sample of 248 phrases, 223 patterns and 13 classes were discovered, but only 13% of patterns involved class substitution.
  • English (Criterion 2: "two-deep"): This criterion showed some success with the restricted data, yielding 30 patterns and 19 classes. Of these, 18 patterns were found to be "quite homogeneous" and generated "legitimate" sentences.

Limitations and Future Work

  • Limitations: The paper acknowledges that some classes contained "grammatically dissimilar members" when using the "one-deep" criterion. The "two-deep" criterion, while more precise, yielded fewer instances of class substitution, suggesting it might be too restrictive or that the data sample was not sufficiently large for it to be fully effective in Hindi. The English data is described as "highly restricted."
  • Future Work: The authors suggest that a useful next step would be to "refine the broad classes formed with a 'one-deep' criterion" in subsequent runs, possibly using the same or new data. This implies a desire to build upon the initial findings and create more precise and robust classifications.

Summary updated

BibTeX
@inproceedings{khubohandani1969application,
  author = {L. M. Khubohandani and W. W. Glover},
  title = {{An Application of Computer Techniques to Analysis of the Verb Phrase in Hindi and English: A Preliminary Report}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1969},
}
Figure