Knowledge Intensive Word Alignment with KNOWA

Knowledge Intensive Word Alignment with KNOWA Emanuele PIANTA and Luisa BENTIVOGLI ITC-irst Via Sommarie, 18 38050 Povo - Trento Italy {pianta,bentivo}@itc.it Abstract In this paper we present KNOWA, an English/Italian word aligner, developed at ITC-irst, which relies mostly on information contained in bilingual dictionaries. The performances of KNOWA are compared with those of GIZA++, a state of the art statistics-based alignment algorithm. The two algorithms are evaluated on the EuroCor and MultiSemCor tasks, that is on two English/Italian publicly available parallel corpora. The results of the evaluation show that, given the nature and the size of the available English-Italian parallel corpora, a language-resource-based word aligner such as KNOWA can outperform a fully statistics-based algorithm such as GIZA++. 1 Introduction Aligning a text and its translation (also known as bitext) at the word level is a basic Natural Language Processing task that has found various applications in recent years. Word level alignments can be used to build bilingual concordances for human browsing, to feed machine learning-based translation algorithms, or as a basis for sense disambiguation algorithms or for automatic projection of linguistic annotations from one language to another. A number of word alignment algorithms have been presented in the literature, see for instance (Véronis, 2000) and (Melamed, 2001). Shared evaluation procedures have been established, although there are still open issues on some evaluation details (Ahrenberg et al. 2000). Most of the known alignment algorithms are statistics-based and do not exploit external linguistic resources, or use them to a very limited extent. The main attractive of such algorithms is that they are language independent, and only require a parallel corpus of reasonable size to be trained. However, word alignment can be used for different purposes and in different application scenarios; different kinds of alignment strategies produce different kinds of results (for instance in terms of precision/recall) which can be more or less suitable to the goal to be achieved. The requirement of having a parallel corpus of adequate size available for training the statistics- based algorithms may be difficult to meet, given that parallel corpora are a precious but often rare resource. For the most common languages, such as English, French, German, Chinese, etc., reference parallel corpora of adequate size are available, and indeed statistics-based algorithms are evaluated on such reference corpora. Unfortunately, if one needs to replicate in a different corpus the results obtained for the reference corpora, finding a parallel corpus of adequate size can be difficult even for the most common languages. Consider that one of the most appealing features of statistics- based algorithms is their ability to induce alignment models for bitexts belonging to very specific domains, an ability which seems to be out of reach for algorithms based on generic linguistic resources. However, for the statistics-based algorithms to achieve their objective, a parallel corpus for the specific domain needs to be available, a requirement that in some cases cannot be met easily. For these reasons, we claim that in some cases algorithms based on external, linguistics resources, …

Content truncated to 500 words