Harnessing sense-level information for semantically augmented knowledge extraction

Nowadays, building accurate computational models for the semantics of language lies at the very core of Natural Language Processing and Artificial Intelligence. A first and foremost step in this respect consists in moving from word-based to sense-based approaches, in which operating explicitly at the level of word senses enables a model to produce more accurate and unambiguous results. At the same time, word senses create a bridge towards structured lexico-semantic resources, where the vast amount of available machine-readable information can help overcome the shortage of annotated data in many languages and domains of knowledge. This latter phenomenon, known as the knowledge acquisition bottlneck, is a crucial problem that hampers the development of large-scale, data-driven approaches for many Natural Language Processing tasks, especially when lexical semantics is directly involved. One of these tasks is Information Extraction, where an effective model has to cope with data sparsity, as well as with lexical ambiguity that can arise at the level of both arguments and relational phrases. Even in more recent Information Extraction approaches where semantics is implicitly modeled, these issues have not yet been addressed in their entirety. On the other hand, however, having access to explicit sense-level information is a very demanding task on its own, which can rarely be performed with high accuracy on a large scale. With this in mind, in ths thesis we will tackle a two-fold objective: our first focus will be on studying fully automatic approaches to obtain high-quality sense-level information from textual corpora; then, we will investigate in depth where and how such sense-level information has the potential to enhance the extraction of knowledge from open text. In the first part of this work, we will explore three different disambiguation scenar- ios (semi-structured text, parallel text, and definitional text) and devise automatic disambiguation strategies that are not only capable of scaling to different corpus sizes and different languages, but that actually take advantage of a multilingual and/or heterogeneous setting to improve and refine their performance. As a result, we will obtain three sense-annotated resources that, when tested experimentally with a baseline system in a series of downstream semantic tasks (i.e. Word Sense Disam- biguation, Entity Linking, Semantic Similarity), show very competitive performances on standard benchmarks against both manual and semi-automatic competitors. In the second part we will instead focus on Information Extraction, with an emphasis on Open Information Extraction (OIE), where issues like sparsity and lexical ambiguity are especially critical, and study how to exploit at best sense-level information within the extraction process. We will start by showing that enforcing a deeper semantic analysis in a definitional setting enables a full-fledged extraction pipeline to compete with state-of-the-art approaches based on much larger (but noisier) data. We will then demonstrate how working at the sense level at the end of an extraction pipeline is also beneficial: indeed, by leveraging sense-based techniques, very heterogeneous OIE-derived data can be aligned semantically, and unified with respect to a common sense inventory. Finally, we will briefly shift the focus to the more constrained setting of hypernym discovery, and study a sense-aware supervised framework for the task that is robust and effective, even when trained on heterogeneous OIE-derived hypernymic knowledge.