Supervised and unsupervised methods for learning representations of linguistic units
Sascha Rothe · 2017
Word representations, also called word embeddings, are generic representations, often high-dimensional vectors. They map the discrete space of words into a continuous vector space, which allows us to handle rare or even unseen events, e.g. by considering the nearest neighbors. Many Natural Language Processing tasks can be improved by word representations if we extend the task specific training data by the general knowledge incorporated in the word representations. The first publication investigates a supervised, graph-based method to create word representations. This method leads to a graph-theoretic similarity measure, CoSimRank, with equivalent formalizations that show CoSimRank’s close relationship to Personalized Page-Rank and SimRank. The new formalization is efficient because it can use the graph-based word representation to compute a single node similarity without having to compute the similarities of the entire graph. We also show how we can take advantage of fast matrix multiplication algorithms. In the second publication, we use existing unsupervised methods for word representation learning and combine these with semantic resources by learning representations for non-word objects like synsets and entities. We also investigate improved word representations which incorporate the semantic information from the resource. The method is flexible in that it can take any word representations as input and does not need an additional training corpus. A sparse tensor formalization guarantees efficiency and parallelizability. In the third publication, we introduce a method that learns an orthogonal transformation of the word representation space that focuses the information relevant for a task in an ultradense subspace of a dimensionality that is smaller by a factor of 100 than the original space. We use ultradense representations for a Lexicon Creation task in which words are annotated with three types of lexical information – sentiment, concreteness and frequency. The final publication introduces a new calculus for the interpretable ultradense subspaces, including polarity, concreteness, frequency and part-of-speech (POS). The calculus supports operations like “−1 × hate = love” and “give me a neutral word for greasy” (i.e., oleaginous) and extends existing analogy computations like “king − man + woman = queen”.