Bridging Languages through Etymology: The case of cross language text categorization

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 651–659, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Bridging Languages through Etymology: The case of cross language text categorization Vivi Nastase and Carlo Strapparava Human Language Technologies, Fondazione Bruno Kessler Trento, Italy {nastase, strappa}@fbk.eu Abstract We propose the hypothesis that word ety- mology is useful for NLP applications as a bridge between languages. We support this hypothesis with experiments in cross- language (English-Italian) document cat- egorization. In a straightforward bag-of- words experimental set-up we add etymo- logical ancestors of the words in the docu- ments, and investigate the performance of a model built on English data, on Italian test data (and viceversa). The results show not only statistically significant, but a large improvement – a jump of almost 40 points in F1-score – over the raw (vanilla bag-of- words) representation. 1 Introduction When exposed to a document in a language he does not know, a reader might be able to glean some meaning from words that are the same (e.g. names) or similar to those in a language he knows. As an example, let us say that an Italian speaker is reading an English text that contains the word expense, which he does not know. He may be re- minded however of the Latin word expensa which is also the etymological root of the Italian word spesa, which usually means “cost”/”shopping”, and may thus infer that the English word refers to the cost of things. In the experiments presented here we investigate whether an automatic text cat- egorization system could benefit from knowledge about the etymological roots of words. The cross language text categorization (CLTC) task consists of categorizing documents in a target language Lt using a model built from labeled examples in a source language Ls. The task becomes more diffi- cult when the data consists of comparable corpora in the two languages – documents on the same top- ics (e.g. sports, economy) – instead of parallel cor- pora – there exists a one-to-one correspondence between documents in the corpora for the two lan- guages, one document being the translation of the other. To test the usefulness of etymological in- formation we work with comparable collec- tions of news articles in English and Ital- ian, whose articles are assigned one of four categories: culture and school, tourism, qual- ity of life, made in Italy. We perform a progres- sion of experiments, which embed etymological information deeper and deeper into the model. We start with the basic set-up, representing the doc- uments as bag-of-words, where we train a model on the English training data, and use this model to categorize documents from the Italian test data (and viceversa). The results are better than ran- dom, but quite low. We then add the etymological roots of the words in the data to the bag-of-words, and notice a large – 21 points – increase in per- formance in terms of F1-score. We then use the bag-of-words …

Content truncated to 500 words