Beyond the Feasibility Study: Lexicographic Progress

Annette StachowitzCOLING 1973

TL;DR

This paper details the ongoing lexicographic work at the Linguistics Research Center aimed at building a dictionary for machine translation. It moves beyond initial feasibility studies by focusing on capturing detailed syntactic and semantic features of words to enable high-quality translation, addressing the need for richer lexical information than traditional dictionaries provide.

What Problem Does This Paper Address?

The paper addresses the limitations of existing machine translation (MT) systems, which often struggle with producing high-quality, nuanced translations. A 1971 feasibility study highlighted the critical need for "lexical research to determine the syntactic and semantic patterns of linguistic entities." This research would form the basis of a dictionary that goes beyond simple word-to-word mappings, incorporating features necessary for understanding sentence well-formedness, avoiding awkward phrasing (forced reading), and correctly interpreting semantic meaning for accurate translation choices. The challenge is to create a comprehensive yet manageable system that captures these linguistic intricacies.

Methods

The core of the paper describes a three-step method for compiling MT dictionaries:

  1. Inclusion of information from existing dictionaries: The process begins by leveraging established German-English and English dictionaries, particularly those that already distinguish between different readings of lexical items based on syntactic or semantic criteria.
  2. Coding of carefully selected semantic and syntactic features: This is the most novel aspect. The researchers are developing a classification system to assign specific features to lexical entries. These features cover:
  3. Properties of the element itself: Syntactic category, discontinuous elements (like separable prefixes), and semantic sub-classification.
  4. Selection restrictions: Syntactic and semantic constraints on complements (e.g., what kind of subject a verb requires, or what type of object it can take). These are represented using a subscript notation (e.g., TC(HU, AB) for "requires a human or abstract complement").
  5. Additional restrictions: Tags for subject areas (e.g., zoology, law) and special usage.
  6. Interpretation of information: The final stage involves producing concordances and sub-lists to identify regularities, semantic sub-classes, and potential redundancies, aiming to streamline the dictionary and develop lexical redundancy rules.

Datasets

The paper does not explicitly mention specific datasets used for training or evaluation in the modern sense. Instead, it details the compilation of extensive bilingual word lists derived from existing dictionaries and linguistic research. These lists include:

  • German verb list with English equivalents (approx. 17,500 entries).
  • English verb list (approx. 6,500 entries).
  • German noun list with English equivalents, gender, and inflectional information (approx. 73,000 entries).
  • Bilingual adjective list (approx. 27,000 entries).
  • English-German adverb list (in progress, approx. 3,500 entries).
  • Bilingual lists of phrasal verbs.

These lists serve as the foundational data for the feature coding and classification process.

Key Insights and Ideas

  • Beyond Surface-Level Translation: The paper strongly advocates for moving beyond simple surface-level analysis of words. True quality translation requires understanding the underlying syntactic and semantic patterns that govern how words function in sentences.
  • The Power of Features: The core innovation lies in the detailed feature system designed to capture nuanced linguistic information. This includes not just word meaning but also how words interact with their environment (selection restrictions) and their specific roles in language.
  • Bridging Surface and Deep Structure: The system aims to capture functional relationships between surface and deep structures, particularly for translation equivalents where these relationships differ between languages (e.g., the German verb mißglücken vs. its English equivalent fail).
  • Language-Specific vs. Universal Features: While the classification system is largely the same for German and English, the paper acknowledges that some syntactic structures are language-specific, requiring careful handling.
  • The Importance of "Real Distinctions": The researchers emphasize capturing distinctions made in the language rather than purely abstract, real-world distinctions. This is exemplified by the noun "novel" needing to be classified as both abstract and an inanimate physical object.
  • Managing Complexity: A significant challenge is balancing comprehensiveness with manageability. The classification system and the process of identifying redundancies are key to achieving this balance.

Key Results

The paper primarily describes the process and methodology for building a richer lexicon, rather than presenting final quantitative results of an MT system. The key "results" are the development and ongoing compilation of these detailed lexical lists and the feature system itself. The paper outlines the substantial progress made in creating these extensive bilingual lists and the initial coding of syntactic and semo-syntactic features for a significant number of entries.

Limitations and Future Work

  • Tentative Classification System: The classification system for adverbs, prepositions, and conjunctions is explicitly stated as "very tentative."
  • Need for Derivational Analysis: The paper notes the lack of a derivational analysis system, which hinders the mechanical recognition of underlying verbs from deverbative nouns.
  • Extending Classification: The most immediate task is extending the classification system, especially for the semantic classification of verbs and adjectives.
  • Restrictive Glosses: There's a recognized need to establish "restrictive glosses" where syntactic and semantic features are insufficient, drawing on information rarely found in conventional dictionaries.
  • Derivational and Compounding Rules: Future work includes studying how syntactic and semo-syntactic features of compounds and derived words can be derived from their roots, aiming to reduce dictionary size through productive rule systems.
  • Integration of Information: The goal is to incorporate information that helps choose correct translations and distinguish between different readings of lexical items.

Summary updated

BibTeX
@inproceedings{stachowitz1973beyond,
  author = {Annette Stachowitz},
  title = {{Beyond the Feasibility Study: Lexicographic Progress}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1973},
}
Figure