An Active Bilingual Lexicon for Machine Translation

Igal Golan, Shalom Lappin, Mori RimonCOLING 1988

TL;DR

This paper introduces an "active" bilingual lexicon that guides machine translation (MT) by using executable descriptions of word senses. This approach aims to improve the transfer phase of MT by allowing linguists to control translation effects through a flexible, user-modifiable lexicon, rather than relying solely on rigid structural transformations.

What Problem Does This Paper Address?

Traditional machine translation systems often struggle with the "transfer" phase, where linguistic structures and lexical items from a source language (SL) are mapped to a target language (TL). While many systems separate lexical and structural transfer, the role and implementation of the bilingual lexicon in this process remain challenging. Specifically, how to represent and utilize word senses in a way that is both linguistically accurate and computationally efficient for translation is a key problem. Existing approaches often rely on static dictionaries or complex, intertwined rules, making them difficult to maintain and adapt. This paper addresses the need for a more dynamic and linguistically informed bilingual lexicon that can actively guide the translation process, particularly for languages with significant structural differences like English and Hebrew.

Methods

The core of this paper is the concept of an active bilingual lexicon. Unlike passive dictionaries, this lexicon contains executable descriptions of word senses. These descriptions are essentially mini-programs that can be triggered by the SL parse tree. The lexicon guides the transfer phase by selecting appropriate TL equivalents based on syntactic context and limited semantic features.

Key methodological aspects include:

  • Executable Lexical Definitions: Word senses are defined by rules that can be directly executed. These rules specify correspondences between SL syntactic environments and TL translations.
  • Separation of Concerns: The system strictly separates lexicon-driven lexical transfer from global syntactic structure mapping. Lexical selection happens before structural transformations.
  • Flexible Rule Specification: A custom specification language is introduced, allowing lexicographers to define rules using a combination of general functions and rule-specific search procedures on parse trees. This language supports checking for syntactic relations, semantic features, and even custom "list markers" for grouping words with similar semantic properties.
  • User Modifiability: The lexicon is designed to be accessible to users, allowing them to refine sense specifications, tailor definitions to specific text domains, or even create private addenda lexicons. This control is constrained to prevent disruption of the core MT architecture.
  • Focus on Transfer: The authors argue that handling sense distinctions and mapping complex lexical correspondences is best done in the transfer phase, rather than pushing it to analysis or generation, especially for linguistically distant language pairs.

Datasets

Not discussed in the paper. The paper focuses on the formalism and methodology, using examples from English to Hebrew translation.

Key Insights and Ideas

  • The Lexicon as an Active Agent: The most significant idea is treating the bilingual lexicon not as a passive lookup table, but as an active component that guides the translation process through executable rules.
  • Sense Distinction for Translation vs. Monolingual Sense: The paper highlights that the sense distinctions required for translation are often different from those used in monolingual dictionaries. A single SL word sense might map to multiple TL senses, and vice-versa, necessitating a translation-specific approach.
  • Modularity and User Control: The proposed system achieves modularity by isolating bilingual lexical considerations. It also empowers users (linguists) to control translation nuances without needing to modify the core MT engine, fostering adaptability and domain-specific tuning.
  • Balancing Generality and Specificity: The specification language aims to strike a balance between general grammatical functions and rule-specific search procedures, recognizing that neither extreme is sufficient for all lexical disambiguation tasks.
  • Critique of Postponing Decisions: The authors argue against postponing lexical disambiguation decisions to the generation phase, especially for distant language pairs, as it can lead to loss of crucial information and make generation overly complex.

Key Results

The paper does not present quantitative results from a fully implemented system. Instead, it focuses on demonstrating the expressive power and flexibility of its proposed lexicon specification language through detailed examples. The "key results" are conceptual:

  • A formal framework for an active bilingual lexicon.
  • A specification language capable of handling complex lexical correspondences and sense distinctions.
  • A methodology that emphasizes user control and modularity in the transfer phase of MT.
  • Demonstration of how this approach can handle challenging cases, such as verbs with multiple senses and different subcategorization frames across languages.

Limitations and Future Work

  • Complexity of Implementation: The authors acknowledge that creating a large-scale, full-fledged bilingual lexicon is a "very heavy task."
  • Need for Generalization: While the current formalism is powerful, they note that further generalization of relations (e.g., a more general "postbrother" concept) could simplify rule writing.
  • Interactive Disambiguation: For cases where sense distinction is difficult (e.g., homonymous nouns like "bank"), they suggest marking such ambiguities to be handled by an interactive disambiguation module in the future.
  • Multi-Target Generality: While the formalism supports multi-target translations with tags, they suggest that constructing distinct variants of the lexicon for different target languages might be more practical due to rule redundancy.
  • Further Study: The paper calls for more research into the nuances of lexical transfer and its impact on overall translation quality.

Summary updated

BibTeX
@inproceedings{golan1988active,
  author = {Igal Golan and Shalom Lappin and Mori Rimon},
  title = {{An Active Bilingual Lexicon for Machine Translation}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1988},
}
Figure