A Hybrid Approach to Interactive Machine

IJCAI 1997

TL;DR

This paper proposes a novel hybrid interactive machine translation (MT) system that seamlessly integrates rule-based, corpus-based, and example-based translation methods. By leveraging an interactive user interface, the system allows users to guide the translation process, leading to more accurate and contextually appropriate translations, especially for writing support.

What Problem Does This Paper Address?

The paper addresses the limitations of traditional automatic machine translation (MT) systems, particularly for writing support tasks like composing emails. While existing MT systems are useful for quick reading, their output often requires significant post-editing for writing tasks where high quality is paramount. Purely rule-based systems struggle with expressions not covered by explicit rules, and purely example-based systems falter when encountering novel expressions. The authors aim to bridge this gap by creating a more flexible and user-controllable MT system that combines the strengths of different translation paradigms.

Methods

The core of the proposed method is a stepwise bottom-up interactive translation scheme. This approach breaks down the translation of a sentence into smaller components, allowing the user to interact at each step. The system presents intermediate translation results and offers choices for translation equivalents and syntactic structures.

The novelty lies in the hybrid translation module that integrates multiple translation paradigms:

  • Rule-based Module: Serves as the backbone, providing default translations and handling well-defined linguistic structures like simple sentence case frame transfers.
  • Corpus-based Module: Primarily used for lexical translation, employing the DMAX method to select the most plausible translation equivalents by analyzing word co-occurrence frequencies in independent source and target language corpora. This method doesn't require parallel corpora.
  • Example-based Module: Applied to syntactically uniform structures like compound nouns or noun phrases. It leverages stored source-target expression pairs to capture semantic and idiosyncratic properties that rule-based systems might miss.

Crucially, the interactive operations are designed to control both rule-based and example-based translation. Users can:

  • Correct the translation area, resolving dependency ambiguities.
  • Specify part-of-speech for translation equivalents.
  • Select translation equivalents from alternatives, influencing the choice of examples.

An online learning scheme is also introduced, where user interactions and corrections are used to learn new translation pairs, allowing the system to adapt to user preferences over time.

Datasets

Not explicitly discussed in the paper. The paper mentions the system dictionary contains about 100,000 Japanese entries and more than 15,000 idiomatic expressions. The idiomatic expressions were built using a method described in [Tamura et al., 1996].

Key Insights and Ideas

  • Hybridization is Key: No single MT paradigm is universally superior. Combining rule-based, corpus-based, and example-based methods offers a more robust solution.
  • Interactivity as a Unifying Force: The interactive interface is not just for correction but is fundamental to controlling the decisions within different translation paradigms, especially for example-based translation.
  • Bottom-up, Incremental Processing: Breaking down translation into smaller, manageable steps with user feedback at each stage makes the process more transparent and controllable.
  • Adaptability through Online Learning: The system can learn from user interactions, becoming more personalized and effective over time.
  • Writing Support Focus: The system is specifically designed to aid users in producing foreign language documents, where translation quality is critical.

Key Results

The paper describes the methodology and architecture of the hybrid system rather than presenting extensive quantitative results. The key qualitative finding is that the interactive operations, originally useful for rule-based disambiguation, are equally effective for controlling example selection and coverage in example-based translation. The system aims to provide results comparable to fully manual translation through user interaction, bridging the gap between automatic MT and manual translation.

Limitations and Future Work

  • Current Implementation Status: The paper notes that the statistics-based lexical transfer module and example-based transfer modules are "currently under way," along with an example learning module. The current implementation is an English writing support tool with a rule-based transfer module and an idiom processing function.
  • Future Work: The authors plan to measure the effectiveness of the method once the expansion (addition of statistics-based and example-based modules) is completed. They also mention the ongoing development of the example learning module.

Summary updated

BibTeX
@inproceedings{1997hybrid,
  title = {{A Hybrid Approach to Interactive Machine}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1997},
}
Figure