Structured Learning for Natural Language Semantic Processing
Abstract. A central topic in Natural Language Processing (NLP) is the design of effective linguistic processes suitable for the target applications. Within this scenario, Convolution Kernels provide a powerful method for the automatic design of these models by defining similarity functions between discrete structures, such as graphs or trees. Their application in NLP allows adopting a Structured Learning paradigm where machine learning algorithms are directly applied to complex structures representing linguistic information, without the need of a complex manual features’ design.
In this Thesis, a study of Convolution Kernels, which aims at jointly modeling syntactic and semantic phenomena of raw texts, is presented. The main topic is the definition of the semantically Smoothed Partial Tree Kernel (SPTK), that is a generalized formulation of one of the most performant Convolution Kernels, i.e. the Tree Kernel (TK), by extending the similarity between tree structures with a function of node similarity. The main characteristic of SPTK is its ability to measure the similarity between syntactic tree structures, which are partially similar and whose nodes can differ but are nevertheless semantically related. One of the most important outcomes is that SPTK allows “embedding” external lexical information in the kernel function only through a similarity function among lexical nodes, namely words. Moreover, SPTK only requires this similarity to be a valid kernel itself. This means that such lexical information can be derived from lexical resources or it can be automatically acquired though a Distributional Analysis of texts, thus improving the applicability of the resulting method.
The SPTK has been evaluated in three complex automatic Semantic Processing tasks: Question Classification in Question Answering, Verb Classification and Semantic Role Labeling. Although these tasks address different problems, state-of-the-art results have been achieved in every evaluation. The attained results demonstrate SPTK`s capacity to derive effective learning systems and, most importantly, without requiring complex activities of features’ design.