Shallow Parsing on the Basis of Words Only: A Case Study

Shallow parsing on the basis of words only: A case study Antal van den Bosch and Sabine Buchholz ILK / Computational Linguistics and AI Tilburg University Tilburg, The Netherlands � Antal.vdnBosch,S.Buchholz � @kub.nl Abstract We describe a case study in which a memory-based learning algorithm is trained to simultaneously chunk sentences and assign grammatical function tags to these chunks. We compare the algo- rithm’s performance on this parsing task with varying training set sizes (yielding learning curves) and different input repre- sentations. In particular we compare in- put consisting of words only, a variant that includes word form information for low- frequency words, gold-standard POS only, and combinations of these. The word- based shallow parser displays an appar- ently log-linear increase in performance, and surpasses the flatter POS-based curve at about 50,000 sentences of training data. The low-frequency variant performs even better, and the combinations is best. Com- parative experiments with a real POS tag- ger produce lower results. We argue that we might not need an explicit intermediate POS-tagging step for parsing when a suffi- cient amount of training material is avail- able and word form information is used for low-frequency words. 1 Introduction It is common in parsing to assign part-of-speech (POS) tags to words as a first analysis step provid- ing information for further steps. In many early parsers, the POS sequences formed the only input to the parser, i.e. the actual words were not used except in POS tagging. Later, with feature-based grammars, information on POS had a more central place in the lexical entry of a word than the identity of the word itself, e.g. MAJOR and other HEAD fea- tures in (Pollard and Sag, 1987). In the early days of statistical parsers, POS were explicitly and often ex- clusively used as symbols to base probabilities on; these probabilities are generally more reliable than lexical probabilities, due to the inherent sparseness of words. In modern lexicalized parsers, POS tagging is of- ten interleaved with parsing proper instead of be- ing a separate preprocessing module (Collins, 1996; Ratnaparkhi, 1997). Charniak (2000) notes that hav- ing his generative parser generate the POS of a con- stituent’s head before the head itself increases per- formance by 2 points. He suggests that this is due to the usefulness of POS for estimating back-off prob- abilities. Abney’s (1991) chunking parser consists of two modules: a chunker and an attacher. The chunker divides the sentence into labeled, non-overlapping sequences (chunks) of words, with each chunk con- taining a head and (nearly) all of its premodi- fiers, exluding arguments and postmodifiers. His chunker works on the basis of POS information alone, whereas the second module, the attacher, also uses lexical information. Chunks as a sepa- rate level have also been used in Collins (1996) and Ratnaparkhi (1997). This brief overview shows that the main reason for the use of POS tags in parsing is that they provide Computational Linguistics (ACL), Philadelphia, July 2002, pp. 433-440. Proceedings of the 40th Annual Meeting of …

Content truncated to 500 words