Semantic vector representations of senses, concepts and entities and their applications in natural language processing

Representation learning lies at the core of Artificial Intelligence (AI) and Natural Language Processing (NLP). Most recent research has focused on develop representations at the word level. In particular, the representation of words in a vector space has been viewed as one of the most important successes of lexical semantics and NLP in recent years. The generalization power and flexibility of these representations have enabled their integration into a wide variety of text-based applications, where they have proved extremely beneficial. However, these representations are hampered by an important limitation, as they are unable to model different meanings of the same word. In order to deal with this issue, in this thesis we analyze and develop flexible semantic representations of meanings, i.e. senses, concepts and entities. This finer distinction enables us to model semantic information at a deeper level, which in turn is essential for dealing with ambiguity. In addition, we view these (vector) representations as a connecting bridge between lexical resources and textual data, encoding knowledge from both sources. We argue that these sense-level representations, similarly to the importance of word embeddings, constitute a first step for seamlessly integrating explicit knowledge into NLP applications, while focusing on the deeper sense level. Its use does not only aim at solving the inherent lexical ambiguity of language, but also represents a first step to the integration of background knowledge into NLP applications. Multilinguality is another key feature of these representations, as we explore the construction language-independent and multilingual techniques that can be applied to arbitrary languages, and also across languages. We propose simple unsupervised and supervised frameworks which make use of these vector representations for word sense disambiguation, a key application in natural language understanding, and other downstream applications such as text categorization and sentiment analysis. Given the nature of the vectors, we also investigate their effectiveness for improving and enriching knowledge bases, by reducing the sense granularity of their sense inventories and extending them with domain labels, hypernyms and collocations.