Bilingual dictionary

JUJ.e--da.ta base

The initiation phase is concerned with input of the SL sentence, dictionary look-up and morphological processing. The analysis phase involves syntactic and semantic parsing, which are accom­plished through linguistic models and case frame. The transfer phase com­prises rules for converting the parsed SL sentence into the TL sentence. The

The dictionary is bilingual: it con­tains morphological, syntactic and se­mantic information needed for the analysis and generation of a sentence. This includes information about the category of words, their semantic fea­tures and oase frames. Entries in the lexicon are of eight types, and their semantic features are based on graded concepts. The case frame provides a means to find out the logical rela­tionship between components in, the sentence.

At present, the dictionary contains a limited set of lexical entries, which are grouped into models with the ini­tial letter as the index for access.

In ECTST, rules are represented in the form of meta-model. It may contain one or more data items as shown belows where A,B,0 represent information on category of the word, while x,y,z and atb,c are parameters indicating multi­ple pieces of information and tree structure separately. In many oases, semantic features or information on grammatical gender, number or case can be used as parameters.

Generally, hundreds or even thousands of grammar rules are necessary for language analysis and transformation. It is impractical to check whether all these rules can be applied to ail in­termediate processings. To make the processing effioient,rules are grouped into packets, each dealing with speci­fic corresponding linguistic phenomena ranging from a phrase, a clause to a ;:..entenoe„

The work of analysis is performed by the parse module- It constitutes a unified syntactic anc semantic analy­sis of the input sentence, and its output rervp-v a;- an input tes transfer. Tpe yarv\v.■'' -'ro^r.^. jntorrrot h tb.o l;.t— -••p i'enb^ricf- scccr*'ii;:~ to -the availa­ble 1 ir:ir-tic models. The^e come in Federal variable», f-uoh as neta-:r;0(k'lf", csfe fra'-'ss etc.,

Th,:; rnr-'fi rncdul e consiptr. of a rurrber of routines, each respo^tible for a snecial wor> . One ratine i or example perform? morphological analysis of woran; another, for syntactic and ~'e~ other, for application of case frames, and so on. Analysis are carried cut from the surface structure throu-'h the deep structure. h deei. structure .T<ay comprise roles of six essential types: SUBject, CBJeot,lKVolvtownt,fITuation, S TA te and OAtHation. They make up a basic frame with the '"KEBicator as shown in Fig. 1.

In .application, a role may come into it? subcategories whenever necessary, SITuation, for example, may fall into

Tl"e, LOOation, DEGree or COFdition; INVolvement into PfsCipient, PARtioi-pant or GOAljand CAUsation into BAF-is, î UHpose or RE'ult ana no on.

In analysis, the deep structure is obtained via preferential weightings caculated from grammatical data and frame labels. The following sentence? for example would result in two dif­ferent internal structures even though their sirrfa.ee structures are similar.

1. We found John very well.

2. We know John very well.

In the case frame, their deep struc­tures would be:

When an internal structure results, it serves as the basis for transfer.

TRANSFER AND GENi'ffATION

In the transfer stage, transformation is accomplished in two steps s first the internal structure of the f'L sen­tence is generated into a tree, with nodes that indicate their proper se­mantic and grammatical order in tha TL sentence; then lexicon rules are in­voked to transfer the HL entry, on the basis of its context, into the TL entry. The advantage of such a pro­cedure are obvious. It makes the rules more flexible, especially those coping with sentences with similar in­ternal structures. Moreover, rules can be added, modified or changed as needed with no resultant effect on one another.

When a target tree is built up, it is then scanned by way of left recursion and with word s appended to its termi­nal nodes aa required. Finally, a character string is obtained; its out­put is a sentence in TL.

CONCLUSION

Efforts made over the past decade have achieved considerable progress in machine translation. First syntactic parsing was pursued, then semantic and context analysis was advocated. ECTST has profited from both theories.

Natural language is essentially a discrete information system. It often allows multiple syntactic interpreta­tions. To minimize the possibility of multiple interpretation, we introduce a minimum amount of semantic informa­tion and adopt the case frame. This provides flexible facility in syntac­tic analysis and helps to distinguish structural ambiguity. In this ap­proach, the translation accuracy is up to 90/3 or more, which can be raised if corresponding information and/or rules are modified and supplemented.

1. Pan Yu and Yuan Y. Sung (87).

"Automatic English-Chinese Transla­tor," Proceeding of Interna­tional Conference on Chinese In­formation Processing, pp. 52-51

'?. Jonothan Slocum Diemens Corpora­tion and Winfield S. Bennett (82). "The LHC Machine Translation Sys­tem: An Application of State-of-the Art Text and Natural Language Pro­cessing Techniques to the Transla­tion of Technical Maimuals t " Work™ ir:> Paper LHC-32-1 July 1982

3. Mei-Du ! i and Gary G. Erickson. " Modular Dictionary Design For Chines8-To-Knglish Machine Transla­tion," GAPlJAK Research Inc. 990 Highland Drive, Solana Beach, Cali­fornia 92075-2472, r,p. 249-252

5. Makoto HAGAO (82), "A Purvey of Natural Language Processing and Ma­chine Translation in Japan," Compu­ter Science à Technology, pp. 64-70 T. Kila<?awa(eri.) OHMS HA, LTD. & North-Holland Pub. Co. 6. T-!akoto NAGAO (86). "Current Status and and Future Trends ir> Machine Translation," Future Generations Computer System 2 (1986) pp.77-32. Elsevier Science Publishers B.V. (North Holland)