Detecting the Untranslatable Colloquial Expressions of Japanese Verbs in Cross-Language Instant Messaging

PACLIC 28 !532 Detecting the Untranslatable Colloquial Expressions of Japanese Verbs in Cross-Language Instant Messaging Yuchang Cheng1, Masaru Fuji1, Tomoki Nagase1, Minoru Uegaki2, Isaac Okada2 1Speech & Language Technologies lab., Media Processing Systems Laboratories, Fujitsu Laboratories Limited., Kawasaki, Japan 2System Engineering Knowledge Improvement div., System Engineering Technology Unit, Fujitsu Limited., Tokyo, Japan 1{cheng.yuchang, fuji.masaru, nagase.tomoki}@jp.fujitsu.com, 2{uegaki.minoru, isaac-okada}@jp.fujitsu.com Abstract Using instant messenger in real time commu- nication is widespread worldwide. However, in the communication of using cross-language (Japanese - other languages) instant messag- ing, the use of colloquial expressions usually degrades the efficiency of communication. The contributions of our research can be split into two parts: (1) we analyzed the in-house conversation logs of business correspondence to obtain the cause of the failure of translation in cross-language instant messaging con- versations; (2) we proposed an automatic sys- tem to detect the untranslatable colloquial ex- pressions of Japanese verbs that are the most significant cause of the failure of Japanese- Chinese (and Japanese-English) machine translation in instant message conversation. 1 Introduction In recent years, using instant messenger in real- time communication has become common world- wide, and there have been advances in the use of machine translation for efficient communications in multi-lingual conversation. There are an increas- ing number of global enterprises that use instant messenger for real-time business correspondence. The users of instant messenger can use machine translation service to overcome the language barri- er with foreign language speakers in real-time business correspondence. However, the unneces- sary conversation arising from the discrepancy of intention increases when the quality of the machine translation is insufficient. Recently, it has been concluded that the current technology of natural language processing can achieve a high level of accuracy only in processing standard linguistic expressions such as newspaper articles. However, non-canonical linguistic expres- sions are frequently used in instant message con- versation. We believe that the translation quality of instant message conversation decreases due to the use of non-canonical language expressions. In our research, we first analyzed a log of instant message conversation. The conversation log is in Japanese and it is collected from our in-house business correspondence. We observed the attrib- utes in the conversation and calculated the appear- ance probability of the attributes. In order to ob- serve the effect that the attributes have on the qual- ity of machine translation, we translated the Japanese utterance 1 in the conversation log into English using Japanese-English translation soft- ware. We found that the attribute “the colloquial ex- pressions of Japanese verbs” is an important factor that causes machine translation quality to decrease. In Japanese, the string expression of a verb com- poses of two parts. One is the stem, which explains the core concept of the verb, and the other is a ter- minal expression that explains other information of the verb (such as voice type and tense/aspect). For example, the verb “買った(bought)” consists of the stem “買(buy)” and the termination “った(past tense)”. The colloquial expressions of a verb usual- ly occur as a non-canonical …

Content truncated to 500 words