A knowledge-based approach to multiwords processing in machine translation: the English-Italian dictionary of multiwords

This poster presents a knowledge-based approach to the identification and translation of multiword expressions (MWEs) from English to Italian. The main assumption of the methodology proposed is that the proper treatment of MWEs in MT calls for a computational approach which must be, at least partially, knowledge-based, and in particular should be grounded on an explicit linguistic description of MWEs, both using a dictionary and a set of rules. Empirical approaches bring interesting complementary robustness-oriented solutions but taken alone, they can hardly cope with this complex linguistic phenomenon for various reasons. For instance, statistical approaches fail to identify and process non high-frequent MWEs in texts or, on the contrary, they are not able to recognise strings of words as single meaning units, even if they are very frequent. Furthermore, MWEs change continuously both in number and in internal structure with idiosyncratic morphological, syntactic, semantic, pragmatic and translational behaviours. The hypothesis is that a linguistic approach can complement probabilistic methodologies to help identify and translate MWEs correctly since hand-crafted and linguistically-motivated resources, in the form of electronic dictionaries and local grammars, obtain accurate and reliable results for NLP purposes. The methodology adopted for this research work is mainly based on the following elements: • an NLP environment which allows the development and testing of the linguistic resources. • an electronic E-I MWE dictionary, based on an accurate linguistic description that accounts for different types of MWEs and their semantic properties by means of well-defined steps: identification, interpretation, disambiguation and finally application. • a set of local grammars We will provide details about the methodology that can be applied to the identification and translation of MWEs. 1. NooJ: an NLP environment for the development and testing of MWE linguistic resources NooJ is a freeware linguistic-engineering development platform used to develop large-coverage formalised descriptions of natural languages and apply them to large corpora, in real time. The knowledge bases used by this tool are: electronic dictionaries (simple words, MWEs and frozen expressions) and grammars represented by organised sets of graphs to formalise various linguistic aspects such as semi-frozen phenomena (local grammars), syntax (grammars for phrases and full sentences) and semantics (named entity recognition, transformational analysis). NooJ’s linguistic engine includes several computational devices used both to formalise linguistic phenomena and parse texts such as FSTs, FSAs, Recursive Transition Networks (RTNs), Enhanced Recursive Transition Networks (ERTNs), Regular Expressions (RegExs), Context Free Grammars (CFGs). NooJ is a tool that is particularly suitable for processing different types of MWEs and several experiments have already been carried out in this area: for instance, Machonis (2007 and 2008), Anastasiadis, Papadopoulou & Gavriilidou (2011), Aoughlis (2011) and finally Vietri (2008). These are only a few examples of the various analysis performed in the last few years on MWE using NooJ as an NLP development and testing environment. 2. The Dictionary of English-Italian MWEs The EIMWE.dic is a dictionary used to represent and recognise various types of MWEs. This dictionary is based on a contrastive English-Italian analysis of continuous and discontinuous MWEs with different degrees of variability of co-occurrence among word compositionality and different syntactic structures. The translation of MWEs requires the knowledge of the correct equivalent in the target language which is hardly ever the result of a literal translation. Given their arbitrariness, MT has to rely on the availability of ready solutions in both languages in order to perform an accurate translation process. Each entry of the dictionary is given a coherent linguistic description consisting of: • the grammatical category for each constituent of the MWE: noun (N), Verb (V), adjective (A), preposition (PREP), determiner (DET), adverb (ADV), conjunction (CONJ); • one or more inflectional and/or derivational paradigms (e.g. how to conjugate verbs, how to nominalise them), preceded by the tag +FLX; • one or more syntactic properties (e.g. “+transitive” or +N0VN1PREPN2); • one or more semantic properties (e.g. distributional classes such as “+Human”, domain classes such as “+Politics”); • the translation into Italian. The EIMWE.dic contains different types of MWE POS patterns. The main part of the dictionary consists of phrasal verbs, support verb constructions, idiomatic expressions and collocations. In the poster, the main verb structures are explained with examples extracted from the British National Corpus, from the Internet by means of the WebCorp LSE application or with our own examples together with the Italian translations. Finally, the corresponding dictionary entry for each example of MWE POS pattern is provided.