A Solution for the Problem of Interactive Disambiguation

Herve BlanchonCOLING 1992

TL;DR

This paper presents a method for interactive disambiguation in machine translation (MT) systems, specifically for monolingual users. It tackles the challenge of resolving ambiguities in the syntactic and semantic structure of sentences by engaging the user in a dialogue, aiming to produce higher quality translations with less linguistic expertise required from the user.

What Problem Does This Paper Address?

Building high-quality machine translation systems often requires a deep understanding of the source language, which can be a barrier for monolingual users who want to use MT without being fluent in the source language. A significant hurdle in MT is disambiguation: determining the correct meaning and structure of a sentence when multiple interpretations are possible. This paper focuses on the problem of interactive disambiguation, where the MT system, after analyzing a sentence, identifies potential ambiguities and engages the user in a dialogue to clarify the intended meaning. The goal is to create a system that can produce accurate translations by leveraging user input to resolve these ambiguities, rather than relying solely on complex linguistic knowledge within the system.

Methods

The core of the proposed solution lies in a structured approach to identifying and resolving ambiguities within the Multisolution, Multilevel and Concrete (MMC) analysis structure. This structure represents multiple possible interpretations at different linguistic levels (syntactic classes, syntactic functions, logic/semantic relations). The authors categorize ambiguities into three main types:

  1. Ambiguity of class: When a word or phrase can be assigned different morpho-syntactic categories in different analyses.
  2. Ambiguity of geometry: When different analyses result in distinct structural graphs, indicating different relationships between sentence components.
  3. Ambiguity of decoration (syntactic, logic, and semantic): When the labeling of nodes in the analysis tree differs across interpretations, affecting the syntactic role or the logical/semantic relations.

To address these, they define a strategy based on three principles: first, identify correct segmentation into simple sentences; second, find subjects, objects, and adjuncts for each predicate; and third, determine the internal structure of these elements. The system then generates a dialogue, using projection (presenting parts of the analysis) and paraphrase (rephrasing potential interpretations) to elicit user clarification. For each ambiguity type, specific methods are employed, such as replacing ambiguous prepositions or permuting sentence elements to highlight different meanings.

Datasets

Not discussed in the paper. The paper focuses on the methodology and provides examples rather than reporting on experiments with specific datasets.

Key Insights and Ideas

  • User-centric disambiguation: The paper champions the idea that involving the user, even a monolingual one, is a practical way to achieve high-quality MT, shifting the burden of linguistic expertise from the system to the interaction.
  • Structured ambiguity classification: Categorizing ambiguities into class, geometry, and decoration provides a systematic way to identify and address different types of structural and semantic uncertainty.
  • Principle-driven strategy: The three guiding principles for disambiguation (segmentation, predicate arguments, internal structure) offer a logical and hierarchical approach to resolving complex linguistic issues.
  • Dialogue as a tool: The use of projection and paraphrase in the dialogue generation process is a key mechanism for making abstract linguistic ambiguities understandable to a non-expert user.

Key Results

The paper does not present quantitative results from a large-scale evaluation. Instead, it demonstrates the effectiveness of its approach through examples of dialogues generated for specific ambiguous sentences. These examples illustrate how the system can present different interpretations and guide the user to select the correct one, leading to a more accurate understanding of the sentence's intended meaning.

Limitations and Future Work

The authors acknowledge that the current method is "fixed once and for all", implying a lack of flexibility. Their future work aims to:

  • Reduce analysis complexity: By getting more information from the user, they hope to reduce the number of analyses generated, thereby speeding up the disambiguation process.
  • Develop tools for linguists: They plan to create tools that allow linguists to define their own disambiguation methods and dialogue generation processes, moving towards a more generic and customizable system.

Summary updated

BibTeX
@inproceedings{blanchon1992solution,
  author = {Herve Blanchon},
  title = {{A Solution for the Problem of Interactive Disambiguation}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1992},
}
Figure