[Paper Review] A Lexicalist Approach to the Translation of Colloquial Text

Fred Popowich, Davide Turcato+7 more|ArXiv.org|Jun 18, 1997

Natural Language Processing Techniques11 references20 citations

TL;DR

This paper presents a lexicalist machine translation system based on the Shake and Bake (S&B) paradigm for translating colloquial English from closed-captioned television broadcasts into Spanish and Brazilian Portuguese. The system uses shallow parsing and lexical transfer rules to handle short, ungrammatical, and idiomatic utterances robustly, achieving 66.27% acceptable or correct translations in evaluation, with visual context enhancing comprehension despite semantic inaccuracies.

ABSTRACT

Colloquial English (CE) as found in television programs or typical conversations is different than text found in technical manuals, newspapers and books. Phrases tend to be shorter and less sophisticated. In this paper, we look at some of the theoretical and implementational issues involved in translating CE. We present a fully automatic large-scale multilingual natural language processing system for translation of CE input text, as found in the commercially transmitted closed-caption television signal, into simple target sentences. Our approach is based on the Whitelock's Shake and Bake machine translation paradigm, which relies heavily on lexical resources. The system currently translates from English to Spanish with the translation modules for Brazilian Portuguese under development.

Motivation & Objective

• To address the challenge of translating colloquial English from closed-captioned television, which features short, ungrammatical, and idiomatic utterances not found in formal written text.
• To develop a fully automatic, large-scale multilingual NLP system capable of translating colloquial input with minimal syntactic and semantic analysis.
• To prioritize robustness over perfect accuracy, leveraging visual and contextual cues from broadcasts to compensate for translation imperfections.
• To enable rapid extension to new target languages through modular language-specific grammars and lexicons.

Proposed method

• The system employs the Shake and Bake (S&B) machine translation framework, which relies on lexical resources and avoids deep syntactic or semantic analysis.
• Input sentences are shallow-parsed to assign structural information to ungrammatical or fragmented utterances, enabling mapping to target language structures.
• Lexical transfer rules map source language bags of words and syntactic features to corresponding target language bags, preserving word order and grammatical function.
• Language-specific grammars and lexicons are used to generate correct target language constructions, with Spanish modules already implemented and Brazilian Portuguese under development.
• The system treats parentheticals, vocatives, and interjections separately, placing their translations according to monolingual target language conventions.
• Evaluation uses human judgment on translation acceptability, categorizing results as correct (yes), acceptable but not ideal (ok), or unacceptable (no).

Experimental results

Research questions

• RQ1 How can a machine translation system effectively handle the syntactic and semantic irregularities of colloquial spoken language found in closed-captioned television?
• RQ2 To what extent can a lexicalist approach based on shallow parsing and lexical transfer achieve acceptable translation performance for informal, fragmented utterances?
• RQ3 How does visual and contextual information from the broadcast compensate for translation inaccuracies in a real-time, multimodal environment?
• RQ4 Can the Shake and Bake framework be effectively scaled and modularized for multilingual deployment in a low-resource, high-robustness setting?
• RQ5 What is the trade-off between translation accuracy and robustness in a domain where input is inherently ambiguous and context-dependent?

Key findings

• The system achieved a combined 66.27% rate of correct or acceptable translations, with 38.46% classified as correct and 27.81% as acceptable but not ideal.
• Approximately 84.5% of sentences could be translated using the S&B approach, indicating strong coverage of the input domain.
• The system successfully translated ungrammatical and idiomatic expressions such as 'how you doing?' into natural Spanish ('qué tal está?'), demonstrating robust lexical mapping.
• Despite semantic inaccuracies—such as translating the passive participle 'strangled' as the active 'estranguló'—translation acceptability was preserved due to visual context.
• The system struggled with certain fragments like 'bag all the stuff on the bedside table?' and 'that’s the only one i need.', indicating limitations in handling complex or ambiguous phrasal constructions.
• The modular design enabled efficient development of Spanish modules and ongoing work on Brazilian Portuguese, supporting rapid extension to additional languages.