Unit Completion for a Computer-aided Translation Typing System

Philippe Langlais, George F. Foster, Guy LapalmeMach. Transl 2000 · DOI · OpenReview

TL;DR

This paper introduces TRANSTYPE, an interactive machine translation system that goes beyond single-word suggestions to predict and offer multi-word completions for the target translation. By integrating language and translation models, TRANSTYPE aims to significantly reduce translator effort and improve productivity.

What Problem Does This Paper Address?

Traditional Interactive Machine Translation (IMT) systems often require translators to actively assist the machine in analyzing the source text. TRANSTYPE tackles the challenge of making the interaction more direct and efficient by focusing on the establishment of the target text. The core problem addressed is how to provide intelligent, multi-word suggestions to a translator as they type, thereby speeding up the translation process and reducing the number of keystrokes required. The authors highlight the potential for significant gains by moving from single-word completion to "unit" completion, which can encompass phrases or even short sentences.

Methods

The TRANSTYPE system's core is a completion engine with two main components: an evaluator and a generator.

  • Evaluator: This component assigns probabilistic scores to completion hypotheses. It uses a linear combination of a language model $p(t|t')$ (predicting the next unit given the preceding ones) and a translation model $p(t|s)$ (predicting a unit given the source text). This differs from some IBM models by not using a "noisy channel" decomposition, prioritizing speed. The interpolation between these models is context-dependent.
  • Generator: This component identifies units that match the current prefix typed by the user and selects the best candidate based on the evaluator's scores. To manage computational complexity, it divides the target vocabulary into a small "active" component (dynamically computed for likely candidates) and a large "passive" component.

A key novelty is the modeling of "unit associations". This involves:

  1. Finding Monolingual Units: Identifying groups of words that function as units within a language. This is done using statistical measures like likelihood and entropy on n-grams extracted from a training corpus.
  2. Mapping Units: Learning the translation equivalents between source and target language units. This is achieved by training a new translation model (IBM 2) using the EM algorithm on a corpus where tokens are grouped into units based on the identified lexicons.

The paper also discusses filtering strategies (F1, F2, and FNP for noun phrases) applied to reduce noise and improve the precision of unit identification and mapping.

Datasets

  • Training Data: A segment of the Hansard corpus was used, consisting of 15,377 sentence pairs, totaling 278,127 English tokens and 292,865 French tokens.
  • Evaluation Data: A test corpus of 747 sentences (13,386 English tokens and 14,506 French tokens) was used, randomly selected from sentences not used for training.

Key Insights and Ideas

  • Multi-word completion is significantly more impactful than single-word completion: Table 1 illustrates that unit completion can save over 85% of keystrokes compared to word completion.
  • Decoupling unit identification from bilingual mapping introduces noise: The authors found that identifying units monolingually and then mapping them can lead to partially correct but less probable associations, and sometimes erratic ones.
  • Filtering is crucial for improving performance: Applying filters (especially for noun phrases) significantly reduces the number of candidate units and improves the accuracy of the models.
  • Compositionality is not always required for good translations: The paper notes that many good unit associations are not necessarily compositional (e.g., "people of Canada" -> "les Canadiens").
  • A balance between coverage and precision is key: The best performing model had relatively poor coverage (did not propose units for many sentences) but still outperformed others, suggesting that filtering out noise is more important than maximizing the number of proposed units.

Key Results

  • The TRANSTYPE prototype, with unit completion, showed modest improvements in prediction performance, serving as a baseline for future work.
  • The best performing model (line 15 in Table 4: E3 + β = 0.4 + F1(2,2,5,0.2) + FNP) outperformed the baseline models by around 1.5% in terms of saved characters.
  • This best model achieved a "spared" character percentage of 53.22%, compared to the baseline models which were around 49-51%.
  • Filtering strategies, particularly noun-phrase filtering (FNP), significantly improved results, removing 92% of candidate sequences.
  • The best model, despite its limitations, demonstrated that reducing noise through filtering can lead to better overall performance.

Limitations and Future Work

  • Decoupling of Unit Identification and Mapping: The authors acknowledge that revising the strategy of decoupling the selection of units from their bilingual context is a clear direction for future research.
  • Coverage: The best model had limited coverage, failing to propose units for a significant portion of sentences. Future work could explore ways to improve this without sacrificing precision.
  • Computational Cost: While efforts were made to optimize speed (e.g., avoiding noisy channel decomposition), the computational demands of evaluating and generating completions remain a consideration.
  • Contextual Adaptation: The paper mentions the potential for context-dependent interpolation coefficients but notes that in their work, the language model's contribution was set to zero. Exploring more dynamic weighting could be beneficial.

Summary updated

Abstract

This work is in the context of, a system thatwatches over the users as they type a translation andrepeatedly suggests completions for the text already entered.The users may either accept, modify, or ignore these suggestions. Wedescribe the design, implementation, and performance of aprototype which suggests completions of units of texts that arelonger than one word.
BibTeX
@article{langlais2000unit,
  author = {Philippe Langlais and George F. Foster and Guy Lapalme},
  title = {{Unit Completion for a Computer-aided Translation Typing System}},
  journal = {Mach. Transl.},
  year = {2000},
  doi = {10.1023/a:1012262211784},
}
Figure