Cascaded Grammatical Relation Assignment

In this paper we discuss cascaded MemoryBased grammatical relations assignment. In the first stages of the cascade, we find chunks of several types (NP,VP,ADJP,ADVP,PP) and label them with their adverbial function (e.g. local, temporal). In the last stage, we assign grammatical relations to pairs of chunks. We studied the effect of adding several levels to this cascaded classifier and we found that even the less performing chunkers enhanced the performance of the relation finder. 1 I n t r o d u c t i o n When dealing with large amounts of text, finding structure in sentences is often a useful preprocessing step. Traditionally, full parsing is used to find structure in sentences. However, full parsing is a complex task and often provides us with more information then we need. For many tasks detecting only shallow structures in a sentence in a fast and reliable way is to be preferred over full parsing. For example, in information retrieval it can be enough to find only simple NPs and VPs in a sentence, for information extraction we might also want to find relations between constituents as for example the subject and object of a verb. In this paper we discuss some Memory-Based (MB) shallow parsing techniques to find labeled chunks and grammatical relations in a sentence. Several MB modules have been developed in previous work, such as: a POS tagger (Daelemans et al., 1996), a chunker (Veenstra, 1998; Tjong Kim Sang and Veenstra, 1999) and a grammatical relation (GR) assigner (Buchholz, 1998). The questions we will answer in this paper are: Can we reuse these modules in a cascade of classifiers? What is the effect of cascading? Will errors at a lower level percolate to higher modules? Recently, many people have looked at cascaded and/or shallow parsing and GR assignment. Abney (1991) is one of the first who proposed to split up parsing into several cascades. He suggests to first find the chunks and then the dependecies between these chunks. Grefenstette (1996) describes a cascade of finite-state transducers, which first finds noun and verb groups, then their heads, and finally syntactic functions. Brants and Skut (1998) describe a partially automated annotation tool which constructs a complete parse of a sentence by recursively adding levels to the tree. (Collins, 1997; Ratnaparkhi, 1997) use cascaded processing for full parsing with good results. Argamon et al. (1998) applied Memory-Based Sequence Learning (MBSL) to NP chunking and subject/object identification. However, their subject and object finders are independent of their chunker (i.e. not cascaded). Drawing from this previous work we will explicitly study the effect of adding steps to the grammatical relations assignment cascade. Through experiments with cascading several classifiers, we will show that even using imperfect classifiers can improve overall performance of the cascaded classifier. We illustrate this claim on the task of finding grammatical relations (e.g. subject, object, locative) to verbs in text. The GR assigner uses several sources of information step by step such as several types of XP chunks (NP, VP, PP, ADJP and ADVP), and adverbial functions assigned to these chunks (e.g. temporal, local). Since not all of these entities are predicted reliably, it is the question whether each source leads to an improvement of the overall GR assignment. In the rest of this paper we will first briefly describe Memory-Based Learning in Section 2. In