TL;DR
This paper introduces a novel method for translating English "delexical structures" (DS) into Japanese, a common linguistic phenomenon where verbs like "take" or "give" carry little lexical meaning, with the meaning residing in the accompanying noun (e.g., "take an approach"). Their approach combines rule-based synthesis with an idiosyncratic approach for difficult cases, achieving a remarkable translation accuracy of nearly 100% on news broadcast texts.
What Problem Does This Paper Address?
English frequently uses "delexical structures" (DS), where common verbs like "take," "have," "make," and "give" combine with a deverbal noun to convey meaning. For instance, "make an address" doesn't mean literally "create an address" but rather "to give a speech." The verbs in these structures are "delexical" – they've lost most of their original lexical meaning and primarily serve to carry grammatical information like tense and person, with the deverbal noun carrying the semantic weight.
Translating these DS accurately into Japanese is crucial for machine translation systems, especially for news broadcasts where they are prevalent. A common pitfall is translating the delexical verb literally, leading to awkward and incorrect Japanese. For example, "make an address" literally translated might become "enzetsu wo tsukuru" (create a speech), which sounds unnatural. The challenge lies in capturing the intended meaning of the DS, not just the literal meaning of its components.
Methods
The authors propose a hybrid approach to DS translation, combining a "synthetic approach" with an "idiosyncratic approach."
- Synthetic Approach: This method aims to synthesize the translation of a DS by using the word sense of each component (delexical verb and deverbal noun) and applying syntactic and semantic rules. This is the primary method for translating "ordinary DS" (type-1 and type-2).
- Idiosyncratic Approach: For DS that are difficult to translate synthetically (idiomatic DS), this approach involves listing them with their specific Japanese translations in a lexicon. This is a more direct, lookup-based method.
The authors categorize DS into three types:
- Type-1 DS: Verbalizes the primary meaning of a deverbal noun.
- Type-2 DS: Verbalizes peripheral meanings of a deverbal noun, often requiring specific collocations with the delexical verb.
- Idiomatic DS: Does not represent the meaning of its individual components and resists transformation.
The translation strategy leverages parallels between English and Japanese, particularly the common use of "sahen-meishi" (verbal nouns) in Japanese, which can be verbalized by adding "suru" (to do). They also developed rules for passivization, recognizing that some English DS require passive translation in Japanese to maintain meaning.
Datasets
The research utilized AP news texts for evaluation. Specifically, they analyzed 21 months' worth of AP texts to count "take" + "noun" collocations, identifying 87 deverbal nouns. For the translation experiment, they selected March 1991 AP texts that contained the targeted delexical verbs and deverbal nouns. They compiled two dictionaries: one with necessary markers for 302 deverbal nouns and 4 delexical verbs, and another without these modifications, to test the effectiveness of their mechanism.
Key Insights and Ideas
- Delexical Structures are Pervasive: The study highlights the high frequency of DS in English texts, with "take" alone appearing over 20,000 times in a subset of AP texts, and a significant portion of these involving deverbal nouns.
- Meaning Resides in the Noun: The core insight is that in DS, the delexical verb is largely a grammatical carrier, and the semantic content is primarily conveyed by the deverbal noun.
- Hybrid Approach is Effective: Combining rule-based synthesis with a lexicon for idiomatic cases provides a robust solution for translating DS, balancing generality with the need to handle exceptions.
- Japanese "Sahen-Meishi" Facilitates Translation: The prevalence of "sahen-meishi" in Japanese, which can easily be verbalized, makes it a suitable target language for translating many English DS.
- Passivization is Crucial: Recognizing and correctly translating passive constructions for certain DS is vital to avoid meaning reversal. The authors identified conditions for passive substitution based on the characteristics of the delexical verb and deverbal noun.
Key Results
The proposed DS translation method, implemented in an English-to-Japanese machine translation system, demonstrated remarkable improvements in translation quality.
- High Accuracy: The experimental results showed a "Correct Rate" of nearly 100% for translations involving the DS translation mechanism, compared to the baseline system which produced awkward translations by literal verb translation.
- Specific DV Performance: The mechanism achieved high correct rates for individual delexical verbs: 92% for "give," 95% for "have," and 99% for both "make" and "take."
- Qualitative Improvements: Examples provided show significant improvements, such as correctly translating "take the lead" as "riido wo suru" (to go ahead) instead of the literal and incorrect "namari wo toru" (to pick up lead). Passive DS were also translated correctly, preserving the intended meaning.
Limitations and Future Work
- Lexical Management: The authors acknowledge that the idiosyncratic approach, while necessary for idiomatic DS, increases the number of lexical entries, potentially complicating lexical management.
- Rule Completeness: For the synthetic approach, they note that not all DS can be translated this way, as the necessary rules have not yet been fully factored out.
- Verb/Noun Conversion: The verbalization mechanism, which simply attaches "suru" to nouns, can lead to awkwardness in some cases (e.g., "inshou wo shite" instead of "inshou wo ataete"). Further elaboration using noun/verb conversion tables is suggested.
- Broader DS Coverage: The current work focuses on a limited set of delexical verbs ("take," "make," "have," "give") and their associated deverbal nouns. Future work could explore other delexical verbs (e.g., "put," "cast," "receive") and a wider range of DS.
- Other DS Components: The role of other DS components like articles, possessive pronouns, and prepositions in translation needs further exploration.
- Semantic Nuances: Capturing subtle differences between similar verbs (e.g., "make" vs. "give," "take" vs. "have") and their impact on translation requires deeper investigation.
Summary updated