Automatic Aquisition of Semantic Attributes for User Defined Words m Japanese to English Machine Translation

Automatic Aquisition of Semantic Attributes for User Defined Words m Japanese to English Machine Translation Satoru Ikehara *, Satoshi Shirai * LAkio Yokoo * Francis Bond * and Yoshie Omi * * • NTT Communication .Science Laboratories, 1-2356 Take, Yokosuka-shi, Kanagawa, Japan • *NTT Advanced Technology Corporation, 90-6 Kawakami-cho, Totsuka-ku Yokohama-shi, Kanagawa, Japan Email: ikehara{shirai, ayokoo, bond}@nttkb.ntt.jp Abstract This paper proposes a method that automatically acquires the SAs (semantic attributes) of user defined words. Applying this method to the compilation of a user dictionary targeting newspaper article sentences and sentences of software design documents has revealed that the automatically determined SAs include 50 to 80% of the correct attributes. Translation experiments confirmed that the automatically acquired SAs improved translation quality by 6-13%. 1. Introduction When working with a MT(machine translation) system, users compile user dictionaries for the words which are not registered in the system dictionaries or for those with inappropriate translations [1]. But when registering new words in a dictionary, there is a need to give not just the index word and translated word, but also syntactic, semantic and various other information. Systems aiming at higher levels of translation quality require more detailed and accurate information [2,3], and it is no simple task for laymen to give such information. In particular, semantic information usually requires the skill of professionals. In this paper, attention has been focused on the characteristics of user defined words. A method is proposed where for index words (noun words or compound nouns) in the original language that users seek to register, one need give only the translation in the target language to permit the system apply the knowledge held in the system dictionaries, automatically judge the type of noun and determine the SAs of the word for the noun types. 2. System Dictionaries and User Dictionaries 2.1 Semantic Dictionaries in ALT-J/E Here, we shall refer to the dictionary prepared in advance by the MT system as a system dictionary and the dictionary prepared and used by users as a user dictionary. Semantic Attribute Systems bu%NO tic I 2,800 Attributes -~ 200 Attributes -~ bute System 100 Attributes For the case of the Japanese to English MT system ALT-J/E[4], the relationship between the system dictionaries, the user dictionaries and word SAs are shown in Figure 1. In the semantic dictionaries, semantic information is written using SAs. Common nouns in the semantic word dictionary are given common noun SAs (generally more than one). For proper nouns, both common noun SAs and proper noun S/ks(beth more than one) are given. Verbal SAs are given to sentence patterns registered in the semantic structure dictionary [5]. 3. Determination of Semantic Attributes A method of determining the SAs of user defined words is shown in Figure 2. This method works using the information held in the system dictionaries when index words (Japanese expressions) of user defined words and their translations(English) have been provided by the user. • Japanese Index Wo~ts .English Translation ( Selection of Head Noun ) ~t n ( …

Content truncated to 500 words