A Limited-Domain English to Japanese Medical Speech Translator Built Using REGULUS 2

Manny Rayner, Pierrette Bouillon, Vol Van Dalsem Iii, Hitoshi Isahara, Kyoko Kanzaki, Beth Ann HockeyACL 2003

Sure, here's a blog-style summary of the paper "A Limited-Domain English to Japanese Medical Speech Translator Built Using REGULUS 2":

TL;DR

This paper presents a novel architecture for a limited-domain English-to-Japanese medical speech translator, designed to bridge the communication gap between physicians and non-native English-speaking patients in emergency settings. By combining linguistically motivated grammar with efficient phrasal translation, the system achieves a balance of accuracy and speed, demonstrating the feasibility of spoken machine translation for critical medical dialogues.

What Problem Does This Paper Address?

Effective communication is paramount in medical diagnosis, especially in emergency departments where a patient's history is as crucial as a physical examination. However, language barriers pose a significant challenge, particularly for immigrant populations seeking medical care. While a full-fledged, open-domain medical speech translator is currently infeasible, this paper addresses the need for a practical solution in a constrained, yet critical, scenario: eliciting basic headache symptom information from English-speaking physicians to non-native Japanese-speaking patients. The authors aim to create a system that facilitates this crucial initial interaction, overcoming the limitations of current translation technologies in this specific context.

Methods

The core of the proposed architecture is a compromise between principled linguistic processing and efficient phrasal translation. At runtime, the system functions as a phrasal translator, allowing for some variation in input language, similar to a phrasebook with "slots." However, the underlying compile-time architecture is linguistically motivated, deriving phrasal patterns from a single, large, unification grammar. This approach aims to provide speed and reliability during runtime while making the system easier to extend and modify.

The system comprises three main modules:

  1. Speech Recognition and Source Language Analysis: This module uses the Open Source REGULUS 2 platform to build a constrained Context-Free Grammar (CFG) language model. This model is compiled from a linguistically motivated unification grammar using a small corpus of examples.
  2. Transfer and Generation: This component, written in SICStus Prolog, handles the translation of semantic representations. Attribute-value structures are used for semantic representations, keeping them as simple as possible. Transfer rules map these structures, and a small Definite Clause Grammar (DCG) generates target language surface strings.
  3. Target Language Speech Synthesis: This module is handled by either the Nuance Vocalizer TTS engine or by concatenating recorded wave files.

A key aspect of ensuring reliability is the use of "all-solutions" versions of processing algorithms to automatically check for deficiencies or ambiguities in the rule sets.

Datasets

The paper mentions a "small corpus of examples" used to drive the grammar specialization process. Additionally, a "training corpus" of approximately 200 examples was hand-constructed from an initial corpus provided by a medical professional. The content of the questions was preserved, but the phrasing was revised for spoken dialogue. The system is designed to translate a vocabulary of about 200 words.

Key Insights and Ideas

  • The Power of Limited Domains: The paper strongly argues that for specific, constrained applications like medical symptom elicitation, spoken machine translation is not only feasible but can be highly effective, even with current technology.
  • Bridging Linguistics and Efficiency: The hybrid approach of using a linguistically motivated grammar at compile time to generate efficient runtime phrasal translation rules is a clever way to balance expressiveness, extensibility, and performance.
  • Attribute-Value Structures for Simplicity: The use of flat attribute-value structures for semantic representations, while potentially problematic in broad domains, proves advantageous in this narrow context by simplifying transfer operations and avoiding ambiguity.
  • Reliability through Generation and Testing: The method of generating random utterances within system coverage and automatically checking results is a robust way to identify and eliminate rule-set deficiencies and ambiguities.
  • Handling Elliptical Utterances: The paper highlights the importance of handling elliptical utterances, which are natural in the one-way physician-to-patient interaction, and demonstrates a practical approach to managing these in the English-Japanese translation.

Key Results

The prototype system successfully translates spoken yes-no questions about headache symptoms from English to Japanese, using a vocabulary of about 200 words. The system is capable of handling variations in phrasing and elliptical utterances, which are common in natural dialogue. Table 1 provides examples of utterances covered by the prototype and their Japanese translations, demonstrating the system's ability to map semantic representations to grammatically correct and contextually appropriate Japanese sentences. The paper emphasizes that for the given limited domain, the target-language grammar is constrained enough to generate only one Japanese sentence from a given representation, ensuring a high degree of accuracy.

Limitations and Future Work

The authors acknowledge that designing a spoken translation system for obtaining a detailed medical history would be difficult or impossible with the current state of the art. Their system is specifically designed for a limited-domain application. While the system handles elliptical utterances, it notes that English and Japanese have different conventions, requiring specific transformations to ensure grammatical correctness. Future work could involve expanding the vocabulary and the range of symptoms covered, as well as exploring more complex dialogue structures. The paper also implicitly suggests that the REGULUS 2 platform and the grammar specialization techniques could be applied to other limited-domain speech translation tasks.

Summary updated

BibTeX
@inproceedings{rayner2003limited,
  author = {Manny Rayner and Pierrette Bouillon and Vol Van Dalsem III and Hitoshi Isahara and Kyoko Kanzaki and Beth Ann Hockey},
  title = {{A Limited-Domain English to Japanese Medical Speech Translator Built Using REGULUS 2}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2003},
}
Figure