Propositional KnowledgeAcquisition and Application to Syntactic and Semantic Parsing

Resumo

Interpretation of natural language is one of the central challenges for the development of an artificial intelligence. In general, interpretation requires to build a context of entailed implicit information (from hearer and speaker background knowledge) that permits to recover the original communicative intention. Natural language processing tasks are concrete realizations of our human ability to comprehend and use language, thus in the long term we will need to provide full interpretation capabilities to machines, starting with the development of methods to acquire and use background knowledge. We explore the use of propositions as background knowledge and its utility for language interpretation. Propositions encode knowledge in the form of assertions using natural language, and provide a straightforward way of expressing knowledge without domain restrictions. Propositional knowledge can be derived directly from meaning representations that, in turn, can be obtained directly from text, and therefore, knowledge and representations can be easily connected to perform the textual inferences required for language interpretation. In this thesis, we propose the automatic acquisition of propositional knowledge from large corpora whose documents are represented as graphs. The frequencies of occurrence permit to express a sense of plausibility. The resulting proposition store supposes a middle ground between meaning representations and structured knowledge bases. This opens new research lines that we address in this work. One the one hand, the connection of the meaning representation with the proposition stores so that they can play the role of the background knowledge that enables an inference. On the other hand, the mapping between proposition stores and structured knowledge bases. We explore these research lines with two specific tasks related to natural language understanding: syntactic and semantic parsing. Specifically for syntactic parsing, we address the problem of appositive correction. Appositives are grammatical dependencies that are often used to express that an instance belongs to a semantic class. We use propositional knowledge to measure the semantic compatibility between entities and entity types with semantic classes. Then we use this information to disambiguate cases where there are several grammatical valid candidates to govern an apposition. Regarding semantic parsing, we build a lexicon that permits to map natural language utterances in the form of propositions with linked data relations, and show how to use this resource in a question answering system. In addition, we propose a method to evaluate grounding and the effect that the lexicon has in the task, independently from the processes of training or querying. Using propositional knowledge for textual inference represents a new paradigm for language interpretation. The goal is to validate this paradigm and to explore from it the main areas involved: meaning representation, knowledge acquisition and textual inferences. Results show that proposition stores are a general purpose resource that permit to address different tasks related to language interpretation, opening new and promising research avenues.