Experiments with Using Semantical Categories in Parsing Systems

One much used method in syntactical analysis is statistical training of a handwritten grammar. It is a well-known problem of statistical selection that it ignores lexical and semantical preferences in disambigua-tion. We have developed a technique for using semantical preferences in statistical training and present experimental results. We compare the results obtained by using dierent sources of semantical information. 1 Background Recent developments in the eld of broad coverage parsing show two key developments. First, the traditional stochastical grammars where the probability of a production depends only on the left hand side non-terminal are now considered to be too simple, because they depend only on very local information. Second, instead of focussing on the algorithm, it is becoming clear that the selection of the right information used for learning is at least as important. This has been argued in general [5], in relation to context free grammars [2], and in relation to parsing methods that do not require a grammar [3, 10]. It is very easy to see the fundamental problem of stochastical grammars. In a simple sentence such as \He asked about climbing professionally." it is impossible to decide what was professional by just looking at the structure (compare \He asked about paying impatiently .") Any method that performs well needs to take words (or particular properties of words) into account. We have developed a method to train and use stochastical grammars with richer stochastical models. This allows the words in the sentence and their properties to in BLOCKINuence the parsing results. The method was already described in, for example, [7, 6]. In this paper we generalize on this method, and give experimental results. 2 The Algorithm The main equations used for the algorithm are given below. This method has been described in previous work [7, 6], but the algorithm we describe here is more general than what was described in previous work. It is based on the Inside Outside Algorithm [1, 9] with chart-parsing but uses extra information in the edges. We number the rules in the grammar and we write rule number x as R x. When the left hand side nonter-minal of rule x is p we write R p x. We also distinguish the edges by numbering them and write edge number k as e k. When edge e k was produced with rule x we write e x k. We assume every edge is …