An English-to-Korean Machine Translator: MATES/EK

Key-Sun Choi, Seungmi Lee, Hiongun Kim, Deok-bong Kim, Cheoljung Kweon, Gilchang KimCOLING 1994

TL;DR

This paper introduces MATES/EK, a research prototype for English-to-Korean machine translation. It employs a transfer-based architecture with novel subsystems for grammar development and analysis, aiming to overcome the significant syntactic differences between the two languages.

What Problem Does This Paper Address?

Machine translation (MT) between languages with vastly different structures, like English and Korean, presents a significant challenge. English, an SVO (Subject-Verb-Object) language with relatively fixed word order and prepositions, contrasts sharply with Korean, a SOV (Subject-Object-Verb) language that relies heavily on postpositions and exhibits more flexible word order. Developing an MT system that can accurately capture these structural nuances and produce natural-sounding translations is a complex undertaking. This paper addresses the need for a robust system capable of handling these linguistic disparities.

Methods

MATES/EK is built upon a transfer-based MT architecture. This approach involves:

  • English Analysis:
    • Morphological Analysis: Utilizes N-gram models combined with rules to resolve category ambiguities and identify basic word units, affixes, and idiomatic phrases.
  • Syntactic Analysis: Employs an augmented context-free grammar (CFG) parsed using the generalized Tomita LR parsing algorithm. The grammar is inductively constructed from a carefully selected corpus.
  • Semantic Analysis: Transforms the syntactic structure into a Lexical Semantic Structure (LSS). LSS is a dependency structure that aims to be language-independent, representing constituents in a head-dependent relation without strict ordering. This is achieved through over 300 tree-transforming rules.
  • Transfer:
    • English-to-Korean Lexical Transfer: Utilizes an English-Korean bilingual dictionary and over 300 lexical selection rules to map English words to their Korean equivalents.
  • English-Korean Structural Transformation: Another set of tree transformation grammars converts the English-specific LSS into a Korean-specific dependency structure. This step is crucial for adapting to Korean's SOV structure and different head-dependent relationships (e.g., Korean verbs or adjectives often act as heads).
  • Korean Generation:
    • Syntactic Generation: Reconstructs the Korean word order and determines necessary postpositions based on the dependency structure, rules, and lexical information.
  • Morphological Generation: Synthesizes Korean morphemes according to context and combination rules, handling the complex morphological structure of Korean.

Key novelties include the Grammar Writing Language and its Environment, which facilitates the development of tree transformations for LSS manipulation, and the Lexical Semantic Structure (LSS) itself as a language-independent intermediate representation.

Datasets

The system was trained and evaluated using sentences primarily selected from the IEEE computer magazine (September 1991 issue). Additional test sentences were sourced from HP test sentences and the Longman English Grammar Textbook. The paper mentions a test set of 1,708 sentences with lengths up to 26 words.

Key Insights and Ideas

  • LSS as a Language-Agnostic Intermediate Representation: The concept of LSS is a core idea, aiming to decouple the syntactic analysis from the target language generation by providing a dependency-based structure that is less tied to specific language surface forms.
  • The Importance of Grammar Engineering Tools: The development of a Grammar Writing Language and its environment highlights the critical role of specialized tools in managing the complexity of grammar development for MT systems, especially for languages with significant structural differences.
  • Addressing Syntactic Divergence: The paper explicitly tackles the challenge of mapping English's prepositional phrases and SVO order to Korean's postpositional phrases and SOV order through structural transformations.
  • Ambiguity Resolution is Crucial: The authors acknowledge the significant ambiguity in English syntactic analysis and the need for sophisticated heuristics and semantic markers to resolve it, which in turn impacts Korean postposition selection.

Key Results

The paper reports a success rate of approximately 95% for sentences shorter than 15 words, decreasing to 75% for sentences of 26 words. Success is defined as translations that are natural and require no additional post-editing ('best'), or are structurally correct but have minor lexical errors ('good'). 'Poor' translations have structural or lexical errors but are still understandable. The system was tested on 1,708 sentences.

Limitations and Future Work

The authors acknowledge several limitations and areas for future work:

  • Processing of Non-Continuous Idiomatic Expressions: Detecting and representing these expressions from sentences remains a challenge.
  • Processing Korean Sentence Styles: Capturing the nuances of different Korean sentence styles (e.g., honorifics) is difficult.
  • Ambiguity in English Syntactic Analysis: The current reliance on ad hoc heuristics and lexical semantic markers for ambiguity resolution (like PP attachment) is noted as a problem, impacting Korean postposition selection.
  • Robustness for Ill-formed Sentences: The current system assumes well-formed English input, but real-world text (like IEEE articles) often contains non-sentential elements, requiring more robust processing.
  • Word Correspondency Selection: The strategy for selecting word correspondences between alternatives (using semantic markers and frequency scoring) is considered to lack a strong strategy.

Future work includes scaling up dictionaries and grammar rules, controlling system consistency, and downsizing the system for PC environments.

Summary updated

BibTeX
@inproceedings{choi1994english,
  author = {Key-Sun Choi and Seungmi Lee and Hiongun Kim and Deok-bong Kim and Cheoljung Kweon and Gilchang Kim},
  title = {{An English-to-Korean Machine Translator: MATES/EK}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1994},
}
Figure