Handling Low Translatability in Machine Translation of Long Sentences

Handling Low Translatability in Machine Translation of Long Sentences Svetlana SHEREMETYEVA LanA Consulting ApS Mynstersvej 7A, 2 Copenhagen, Denmark, DK-1827 lanaconsult@mail.dk Abstract. Long and complex sentences are normally “low translatable” sentences due to high ambiguity. We describe a methodology of handling such sentences in APTrans, - a system for machine translation of patent claims between English and other languages. APTrans builds on patent data hard-coded in the system lexicon and grammars coached in diverse grammatical frameworks (PSG and predicate/argument DG). We motivate a specific partial parse, - in our analyzer parsing is reduced to a phrase level and a level of individual simple clauses. The load of detecting a clause hierarchy in a complex sentence is shifted to the generator. Transfer combines an interlingual and a syntactic transfer approaches. The methodology is universal in the sense that it could be used for different domains, languages and applications. We introduce special features of APTrans on the example of English/ Danish language pair. 1. Introduction Low translatability indicators, such as - included and parallel structures, ambiguous PP attachments, etc., (Underwood and Jongejan, 2001) characteristic of long sentences, are the reasons that currently no commercial MT system can translate patent claims adequately. A device for producing a spray of electrically charged particles comprising means defining a location from which the spray is generated, and a voltage generator for producing high voltage between said location and the surroundings, characterized in that said voltage generator comprises a large solid state array of radiation sensitive voltage producing elements interconnected to produce high voltage... Fig. 1. A Fragment of a US patent claim. Predicates of included clauses are bold-faced. In many MT systems long sentences are broken up along punctuation, and the segments are parsed separately (Kim and Ehara, 1994). This method can be incorrect due to the punctuation ambiguity. The long sentence problem is sometimes approached by being very selective about which sentences to parse as in (Hobbs and Bear, 1995), where statistical filter is used to pre- process a text. In the Pattern-based English- Korean MT (Roh et al., 2003) long sentences are handled by using chunking information on the phrase-level of the parse result to which a sentence pattern is applied directly. A patent specific research in MT where the problem of low translatable sentences is addressed by suggesting an interactive analysis module has been done for Russian to English by (Sheremetyeva and Nirenburg, 1999). The most recent attempt to cope with low translatability of a patent claim is a Japanese-English patent MT system, which merges the English claim authoring system AutoPat (Sheremetyeva, 2003) and a Japanese PC-Transfer application (Neumann, 2005). The APTrans system presented in the current paper integrates some of the transfer and generation techniques described in the last cited works but relies on an automatic analyzer bypassing some of low translatability problems. Although the correlation between the sentence length and ambiguity is clear, the great portion of ambiguities occurs in treatment the higher (clause) nodes in the syntactic tree, on the …

Content truncated to 500 words