Exponential Word Embeddings: Models and Approximate Learning

The digital era floods us with an excessive amount of text data. To make sense of such data automatically, there is an increasing demand for accurate numerical word representations. The complexity of natural languages motivates to represent words with high dimensional vectors. However, learning in a high dimensional space is challenging when the amount of training data is noisy and scarce. Additionally, lingual dependencies complicate learning, mostly because computational resources are limited and typically insufficient to account for all possible dependencies. This thesis addresses both challenges by following a probabilistic machine learning approach to find vectors, word embeddings, performing well under aforementioned limitations. An important finding of this thesis is that by bounding the length of the vector that represents a word as well as penalizing the discrepancy between vectors representing different words make a word embedding robust, which is especially beneficial when noisy and little training data is available. This finding is important because it shows how current word embedding methods are sensitive to small variations in the training data. Although, one might conclude from this finding that more data is not necessary anymore, this thesis does show that training on multiple sources, such as dictionaries and thesaurus, improves performance. But, each data source should be treated carefully, and it is important to weigh informative parts of each data source appropriately. To deal with lingual dependencies, this thesis introduces statistical negative sampling with which the learning objective of a word embedding can be approximated. There are many degrees of freedom in the approximated learning objective, and this thesis argues that current embedding strategies are based on weak heuristics to constrain these freedoms. Novel and more theoretical founded constraints are being proposed to constrain the approximations that are based on global statistics and maximum entropy. Finally, many words in a natural language have multiple meanings, and current word embeddings do not address this because they are built on a common assumption that one vector per word representation can capture all word meanings. This thesis shows that a representation based on multiple vectors per word easily overcomes this limitation by having different vectors representing the different meanings of a word. Taken together, this thesis proposes new insights and a more theoretical foundation for word embeddings which are important to create more powerful models able to deal with the complexity of natural languages.