Collocations in a Rule-Based MT System: A Case Study Evaluation of their Translation Adequacy

Proceedings of the 13th Annual Conference of the EAMT, pages 128–135, Barcelona, May 2009 Collocations in a Rule-Based MT System: A Case Study Evaluation of Their Translation Adequacy Eric Wehrli, Violeta Seretan, Luka Nerima, Lorenza Russo Language Technology Laboratory University of Geneva Switzerland {firstname.lastname}@unige.ch Abstract Collocations constitute a subclass of multi-word expressions that are particu- larly problematic for machine translation, due 1) to their omnipresence in texts, and 2) to their morpho-syntactic proper- ties, allowing virtually unlimited variation and leading to long-distance dependen- cies. Since existing MT systems incorpo- rate mostly local information, these are ar- guably ill-suited for handling those collo- cations whose items are not found in close proximity. In this article, we describe an integrated environment in which colloca- tions (and possibly their translation equiv- alents) are first identified from text cor- pora and stored in the lexical database of a translation system, then they are employed by this system, which is capable of deal- ing with syntactic transformations as it is based on a deep linguistic approach. We compare the performance of our system (in terms of collocation translation adequacy) with that of two major MT systems, one statistical, and the other rule-based. Our results confirm that syntactic variation af- fects translation quality and show that a deep syntactic approach is more robust in this sense, especially for languages with freer word order (e.g., German) and richer morphology (e.g., Italian) than English. 1 Introduction Collocations, typical word combinations in a given syntactic relation (e.g., warm greeting, distinct preference, [to] wreak havoc, [to] believe firmly, c⃝ 2009 European Association for Machine Translation. [to] break a record) constitute a well-known prob- lem for machine translation. Their identification in the source text and their proper processing by MT systems is the key factor in producing a more acceptable output (Orliac and Dillinger, 2003). Over the past two decades, intensive efforts have been made to devise accurate techniques for collo- cation extraction from corpora; see, for instance, Church and Hanks (1990), Smadja (1993), Lin (1998), Evert (2004), among many others. Yet, the existing MT systems generally do not integrate collocational resources, or they are not designed to handle collocations in a specific and appropriate manner, as required by their high morpho-syntactic potential. Therefore, such systems often achieve an un- satisfactory literal translation, especially when the collocation items are not found in the canonical or- der or in close proximity.1 For instance, Exam- ple (1b) show the French translation returned by a major MT system, freely available online, for the English sentence in (1a), where the order of the verb and object of the collocation break - record is changed due to passivisation, and there are several words occurring between the two. (1)a. Records are made to be broken. b. ∗Les dossiers sont faites pour ˆetre rompu. [The files are made for be broken.] Since the system tested fails to identify that records and broken are part of a collocation, it is unable to propose a correct translation (in this case …

Content truncated to 500 words