Improving Performance of Transfer-Driven Machine Translation with Extra-Linguistic Information from Context, Situation and Environment

IJCAI 1997

TL;DR

This paper tackles the challenge of making spoken-language machine translation more natural by incorporating extra-linguistic information (context, situation, environment) into a Transfer-Driven Machine Translation (TDMT) system. By doing so, they demonstrate improved translation accuracy and naturalness, particularly in handling politeness and situational nuances.

What Problem Does This Paper Address?

Spoken language translation, especially for dialogs, presents unique challenges compared to written text. It often involves ungrammatical or incomplete sentences, requires real-time processing to avoid interrupting the flow of conversation, and necessitates contextually appropriate expressions. Traditional machine translation systems, even those using transfer-based approaches like TDMT, often struggle with these aspects because they primarily rely on linguistic information within individual sentences, neglecting the broader context, the speaker's social role, gender, or the overall situation. This leads to translations that might be grammatically correct but unnatural or even inappropriate for the given communicative context. The paper aims to bridge this gap by integrating "extra-linguistic" information to enhance the naturalness and accuracy of spoken dialog translation.

Methods

The core of the paper's contribution lies in extending the existing Transfer-Driven Machine Translation (TDMT) system with a mechanism to incorporate extra-linguistic information.

  • Transfer-Driven Machine Translation (TDMT): The paper builds upon TDMT, which uses constituent boundary patterns for parsing and a best-first mechanism for selecting translation candidates. This approach is noted for its efficiency and robustness in handling ungrammatical input.
  • Extra-Linguistic Information Integration: The proposed method attaches extra-linguistic information (e.g., speaker's role, social rank, gender) to transfer patterns. This information is then used in a "parameter-based preference" module to select the most suitable word usage for a given communicative context.
  • Unit-Based Processing: The system processes sentences in units, allowing for different contextual preferences to be applied to different parts of a sentence. For example, the politeness level required for "I" and "you" might differ within the same sentence.
  • Best-Only Phrase Structure Selection: This mechanism, combined with examples and the extra-linguistic parameter model, aims to achieve deterministic translation for each pattern.

Datasets

The paper does not explicitly mention specific large-scale datasets used for training the core TDMT system. However, it describes an experimental evaluation using:

  • "Clerk's" Sentences: A subset of 316 sentences previously graded (B, C, D) from an evaluation of the TDMT EJ system. These sentences were used to test the effectiveness of the proposed extra-linguistic information integration.
  • Unseen Dialogs: Translations for 69-87 unseen dialogs (approximately 1,000 unseen sentences) were manually evaluated.

The paper also refers to the availability of linguistic resources on the web, such as for "donatory verbs," but these are not presented as formal datasets for this study.

Key Insights and Ideas

  • Context is King (Even for Machines): The paper strongly emphasizes that natural communication relies heavily on understanding the context, situation, and environment, not just the words themselves. This is a crucial insight for improving MT beyond literal translation.
  • Politeness is a Complex Linguistic Feature: The study highlights how social rank and role significantly influence politeness in languages like Japanese, and that a robust MT system must account for these nuances to avoid awkward or incorrect translations.
  • Extra-linguistic Information as a Preference Mechanism: The authors propose a practical way to integrate abstract concepts like social role and gender into a concrete preference mechanism within the MT transfer process.
  • Unit-Based Application of Context: Recognizing that different parts of a sentence might have different contextual requirements (e.g., different politeness levels for speaker vs. listener), the paper's approach of applying preferences to smaller units is a clever way to handle this complexity.

Key Results

The preliminary experiments showed promising results:

  • Improved Acceptability: By upgrading sentences that were previously graded as "Nonsense, wrong sense" (D grade), the total acceptability of the system could be improved by approximately 8%.
  • Sentence Transformation: 115 out of 316 tested sentences were changed into different expressions, indicating that the system was actively applying the extra-linguistic information to alter translations.
  • Grade Improvement: 84 sentences were improved by at least one grade, demonstrating a tangible positive impact on translation quality.
  • Efficiency Maintained: The proposed scheme was designed to preserve the efficiency of the TDMT system, which is crucial for real-time spoken dialog translation.

Limitations and Future Work

  • Overly Polite Expressions: A noted limitation is the potential for constructing overly polite expressions in Japanese due to the current generation system lacking sufficient knowledge of collocational constraints of politeness. The authors suggest statistical methods (n-grams, co-occurrence) as a way to overcome this.
  • Manual Assignment of Information: Currently, some extra-linguistic information, like a speaker's social role, needs to be assigned manually.
  • Automatic Extraction of Information: A significant area for future research is the efficient and automatic extraction of extra-linguistic information from dialogs. The authors suggest using statistical and stochastic approaches to find information like preferences for social position, gender, and conversational domain from large corpora.

Summary updated

BibTeX
@inproceedings{1997improving,
  title = {{Improving Performance of Transfer-Driven Machine Translation with Extra-Linguistic Information from Context, Situation and Environment}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1997},
}
Figure