[New publication] Multiword Units in Machine Translation and Translation Technology

In this chapter, we take a closer look at students’ post-editing of multiword units (MWUs) from English into Dutch. The data consists of newspaper articles post-edited by translation students as collected by means of advanced keystroke logging tools.

We discuss the quality of the machine translation (MT) output for various types of MWUs, and compare this with the final post-edited quality. In addition, we examine the external resources consulted for each type of MWU. Results indicate that contrastive MWUs are harder to translate for the MT system, and harder to correct by the student post-editors than non-contrastive MWUs. We further find that consulting a variety of external resources helps student post-editors solve MT problems.

We present several Verb + Noun collocation integration methods using linguistic information, aiming to improve the results of a French-Romanian factored statistical machine translation system (FSMT). The system uses lemmatised, tagged and sentence-aligned legal parallel corpora. Verb + Noun collocations are frequent word associations, sometimes discontinuous, related by syntactic links and with non-compositional sense (Gledhill, 2007). Our first strategy extracts collocations from monolingual corpora, using a hybrid method which combines morphosyntactic properties and frequency criteria. The second method applies a bilingual collocation dictionary to identify collocations. Both methods transform collocations into single tokens before alignment. The third method applies a specific alignment algorithm for collocations. We evaluate the influence of these collocation alignment methods on the results of the lexical alignment and of the FSMT system.

Multilingual Information Extraction requires significant Multiword Expressions (MWE) processing as many such items are multiwords. The lexical representation of MWEs supports large bilingual lexicons (for Persian, Pashto, Turkish, Arabic); multiwords are represented like single words, extended by two annotations: MWE head, and lemma plus part of speech for the MWE parts. In text analysis, MWEs are recognised as part of the parsing process, mot as pre- or post-processing components. The analysis design extends the X-bar scheme by a level for multiword rules. In transfer, MWEs are translated as elementary nodes like single word lemmata, to present key concepts for relevance judgement in Information Extraction. Evaluation shows that 90% of the MWE patterns in the lexicon can be analysed with about 150 MWE-specific rules, and that more than 90% of text document tokens are covered by the proposed integrated single and multiword processing.

This article describes a new word alignment gold standard for German nominal compounds and their multiword translation equivalents in English, French, Italian, and Spanish. The gold standard contains alignments for each of the ten language pairs, resulting in a total of 8,229 bidirectional alignments. It covers 362 occurrences of 137 different German compounds randomly selected from the corpus of European Parliament plenary sessions, sampled according to the criteria of frequency and morphological complexity. The standard serves for the evaluation and optimisation of automatic word alignments in the context of spotting translations of German compounds. The study also shows that in this text genre, around 80% of German noun types are morphological compounds indicating potential multiword units in their parallel equivalents.

As compounds pose a problem for applications that rely on precise word alignments, we developed a state-of-the-art compound splitter for Dutch that makes use of corpus frequency information and linguistic knowledge. Domain-adaptation techniques are used to combine large out-of-domain and dynamically compiled in-domain frequency lists. As compounds are not always translated compositionally, we developed a novel methodology for word alignment. We train the word alignment models twice: a first time on the original data set and a second time on the data set in which the compounds are split into their component parts. Experiments show that the compound splitter combined with the novel word alignment technique considerably improves bilingual terminology extraction results.

This paper outlines a methodology and a system for collocation retrieval and translation from parallel and comparable corpora, developed with translators and language learners in mind. It is based on a phraseology framework, applies statistical techniques, and employs source tools and online resources. The collocation retrieval and translation has proved successful for English and Spanish and can be easily adapted to other languages. The evaluation results are promising and future goals are proposed. Furthermore, conclusions are drawn on the nature of comparable corpora and how they can be better exploited to suit particular needs of target users.

Grounded in phraseology and corpus linguistics, this paper aims to explore the use of bilingual lexical bundles to improve the degree of naturalness and textual fit of translated texts. More specifically, this study attempts to identify lexical bundles, that is, recurrent sequences of 3–7 words with similar discursive functions in a purpose-designed comparable corpus of English and Polish patient information leaflets, with 100 text samples in each language. Because of cross-linguistic differences, we additionally apply a number of formal criteria in order to filter out the bundles in each subcorpus. The results show that bilingual lexical bundles with overlapping discourse functions in texts and extracted from comparable corpora hold unexplored potential for machine translation, computer-assisted translation and bilingual lexicography.

Idiomatic expressions are types of MWUs in which the meaning of the unit does not equal the cummulative meaning of its parts. They are culturally dependent, so the translation cannot be inferred from the expression itself. Croatian language has a very rich idiomatic structure. A few such expressions can be understood in direct translation but most are different from the literal translations. As the idioms are rooted in the tradition of the language and society from which they hail, they need special treatment in computational linguistics. Using NooJ as an NLP tool, we describe different types of Croatian idioms that will help us recognize them in texts. Idioms recognition should be given special treatment, being the major task in translation.

This paper presents a corpus-based analysis of constructions consisting of the verb doći ‘to come’ followed by the prepositions do ‘to’ or na ‘onto’ and a noun. Croatian lexicography mainly describes the verb doći as an intransitive verb of motion. Other uses of this verb are listed in the idioms section. This paper will address the association between the verb-preposition constructions and nouns that follow them. The paper will take frequency data into consideration and attempt to distinguish between primary and additional meanings of constructions doći do ‘to come to’ and doći na ‘to come onto’. The research will be based on hrWaC 2.0 (Croatian Web Corpus, Version 2), the Croatian National Corpus and the Croatian Language Repository corpora.

Collocation identification and anaphora resolution are widely recognised as major issues for natural language processing, and particularly for machine translation. This paper focuses on their intersection domain, that is verb-object collocations in which the object has been pronominalised. To handle such cases, an anaphora resolution procedure must link the direct object pronoun to its antecedent. The identification of a collocation can then be made on the basis of the verb and its object or its antecedent. Results obtained from the translation of a large corpus will be discussed, as well as an evaluation of the precision of the anaphora resolution procedure for this specific task.