CATEx – A Project Proposal

by Johann Gamper

Introduction and Motivation

Due to the equal status of the Italian and the German language in South Tyrol, a great part of legal and administrative documents has to be written in both languages: usually they are first laid down in Italian and then translated into German. A prerequisite for high quality translations is a consistent and comprehensive terminology, i.e. a domain-specific dictionary, which however does not exist yet1 . While organized terminological activities have been neglected for decades, various institutions coined different German terms without prior consultation and created their own in-house dictionaries. The result of this has been a "terminological chaos", lots of inconsistencies, duplication of efforts, and poor quality translations. Since 1994 the Commission for Terminology (TERKOM) has been working in cooperation with the scientific area I of the European Academy of Bolzano (EAB) on a standardization of the legal and administrative terminology in South Tyrol (LA-terminology) . The European Academy of Bolzano is focusing on descriptive terminology – Italian/German term pairs are collected from bilingual texts, e.g. the Italian "codici" and their translations, and preferences for German terms are proposed. In particular cases further examinations are performed such as a comparison with the Austrian/German/Swiss law systems. On the basis of this work, the TERKOM establishes one German term for each Italian term. A few years of experience at the European Academy of Bolzano have shown that exhaustive analysis of the huge number of relevant texts and terminology extraction (experts estimate 40.000-70.000 terms) require the use of advanced computational methods: the CATEx project was born.

Objectives

Referring to figure 1, CATEx aims to provide a framework to support the tasks described in the boldface box. The main objectives are as follows:

1. Development of a computational frame work for corpus-based terminological work.
2. Production of linguistic resources: a corpus of bilingual texts and an LA-terminology.
3. Creation of a sophisticated interface for user access.

Text Database/Corpus

The basis for collecting and describing LA-terminology is a domain-specific parallel corpus of representative Italian/German texts in machine-readable form which cover the whole area of law and administration and show the use of the terms in various contexts (table 1). Some of these documents are widely accepted as standard and used as reference books by translators. In order to facilitate various automatic text processing tasks, implicit information in the corpus has to be made explicit – a procedure generally known as text mark-up or text annotation. Four levels of mark-up can be distinguished: document-wide markup (bibliographic information, etc.), gross structural markup (chapter, section, etc.), sub-paragraph markup (sentences, words, etc.) and linguistic annotation (part-of-speech tags, alignment of parallel texts, etc.). Marking up the texts according to the EAGLES recommendations on corpus encoding, an application of SGML, facilitates the exchange of the corpus and the use of already existing text processing software. The corpus will be implemented as a sophisticated text database.

Data/Knowledge Extraction

Once the corpus of annotated texts is ready, automatic methods to recognize and extract various pieces of data and knowledge are needed. First of all, we want to extract terms and their translation equivalents. For that purpose word sequences in the source language have to be identified as terms and their translations in the target language have to be determined. Further, we want to extract conceptual characteristics, for example the relations of hyponymy (e.g. X is a kind of Y) and meronymy (e.g. X is part of Y). Such relations allow a hierarchical organization of terminology facilitating the search process as well as the use of terminology by other programs. Finally, we need additional information like a definition of a term, a small context example, and the relative sources. To automate the data/knowledge extraction process, a lot of information is required including a definition of what a term is, which word patterns represent what relation, as well as linguistic annotations explicitly encoded in the corpus (part-of-speech tags, alignments, etc.). Even if these tasks are not fully automatic, the algorithms compute a list of candidates to be post-edited by a human. The effectiveness of this approach is incomparable to that of manual extraction, where a person has to scan the entire document.

Terminological Database

User Interface

• Multilingual navigation aids including queries in the Italian and German languages
• Display of parallel documents
• Complex queries: words or compound words, word combinations in a context, etc.
• Restriction of queries to parts of the corpus/terminological database

Concluding Remarks

References