Part of Speech Tagging for Text Clustering in Swedish

Text clustering could be very useful both as an intermediate step in a large natural language processing system and as a tool in its own right. The result of a clustering algorithm is dependent on the text representation that is used. Swedish has a fairly rich morphology and a large number of homographs. This possibly leads to problems in Information Retrieval in general. We investigate the impact on text clustering of adding the part-of-speech-tag to all words in the the common term-bydocument matrix. The experiments are carried out on a few different text sets. None of them give any evidence that part-of-speech tags improve results. However, to represent texts using only nouns and proper names gives a smaller representation without worsen results. We also investigate the effect of lemmatization and the use of a stoplist, both of which improves results significantly in some cases.