Using Foreign Inclusion Detection to Improve Parsing Performance

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 151–160, Prague, June 2007. c⃝2007 Association for Computational Linguistics Using Foreign Inclusion Detection to Improve Parsing Performance Beatrice Alex, Amit Dubey and Frank Keller School of Informatics University of Edinburgh 2 Buccleuch Place Edinburgh, EH8 9LW, UK {balex,adubey,keller}@inf.ed.ac.uk Abstract Inclusions from other languages can be a significant source of errors for monolin- gual parsers. We show this for English in- clusions, which are sufficiently frequent to present a problem when parsing German. We describe an annotation-free approach for accurately detecting such inclusions, and de- velop two methods for interfacing this ap- proach with a state-of-the-art parser for Ger- man. An evaluation on the TIGER cor- pus shows that our inclusion entity model achieves a performance gain of 4.3 points in F-score over a baseline of no inclusion de- tection, and even outperforms a parser with access to gold standard part-of-speech tags. 1 Introduction The status of English as a global language means that English words and phrases are frequently bor- rowed by other languages, especially in domains such as science and technology, commerce, adver- tising, and current affairs. This is an instance of lan- guage mixing, whereby inclusions from other lan- guages appear in an otherwise monolingual text. While the processing of foreign inclusions has re- ceived some attention in the text-to-speech (TTS) lit- erature (see Section 2), the natural language process- ing (NLP) community has paid little attention both to the problem of inclusion detection, and to poten- tial applications thereof. Also the extent to which inclusions pose a problem to existing NLP methods has not been investigated. In this paper, we address this challenge. We focus on English inclusions in German text. Anglicisms and other borrowings from English form by far the most frequent foreign inclusions in German. In spe- cific domains, up to 6.4% of the tokens of a Ger- man text can be English inclusions. Even in regular newspaper text as used for many NLP applications, English inclusions can be found in up to 7.4% of all sentences (see Section 3 for both figures). Virtually all existing NLP algorithms assume that the input is monolingual, and does not contain for- eign inclusions. It is possible that this is a safe assumption, and inclusions can be dealt with ac- curately by existing methods, without resorting to specialized mechanisms. The alternative hypothe- sis, however, seems more plausible: foreign inclu- sions pose a problem for existing approaches, and sentences containing them are processed less ac- curately. A parser, for example, is likely to have problems with inclusions – most of the time, they are unknown words, and as they originate from another language, standard methods for unknown words guessing (suffix stripping, etc.) are unlikely to be successful. Furthermore, the fact that inclusions are often multiword expressions (e.g., named enti- ties) means that simply part-of-speech (POS) tag- ging them accurately is not sufficient: if the parser posits a phrase boundary within …

Content truncated to 500 words