Lexical Semantic Knowledge and Information Extraction

Abstract

In the recent decade there has been a sharp increase in the utilization of machine/deep learning models for the development of Natural Language Processing (NLP) applications, especially focused on language understanding, with end-goals targeted at, but not limited to: information retrieval, machine translation, sentiment analysis, question answering, etc. These applications call for the need of better models for in-depth understanding of the language structures which in turn help development of automated routines that may acquire vast variety of unstructured data from web resources, process the data and convert it to the desired information content. In the recent past many different models have been developed for language-specific information extraction and a better understanding of semantic aspects of the language, however, yet there are some challenges associated that need to be addressed for the improved utility of these models in the down-streaming tasks. In this thesis, we propose new models with the aims to improve upon the existing methods for the lexico-semantic relation and information extraction tasks. For lexico-semantic relation extraction, we work around distinguishing among different lexico-semantic relations captured from unstructured data, i.e., distinguishing antonyms from synonyms and hypernymy detection. For information extraction, we work with improving the Fine-Grained Named Entity Typing (FG-NET), which is a key component for different down-streaming information and relation extraction tasks. Given the fact that the fine-grained type hierarchy follows a hierarchical structure, in Chapter 5, we extend the concepts of FGET-RR to Fine-Grained Named Entity Typing with Refinement in Hyperbolic space (FGNET-RH) that combine the benefits of the non-euclidean geometry (hyperbolic space) along with the graph structures to perform FG-NET in performance-enhanced fashion. Finally, in order to evaluate the reliability of the machine/deep learning systems for information extraction in real-life scenarios, we evaluate the performance of these systems under uncustomized settings.