Learning Lexical Embeddings with Syntactic and Lexicographic Knowledge

We propose two improvements on lexical association used in embedding learning: factorizing individual dependency relations and using lexicographic knowledge from monolingual dictionaries. Both proposals provide low-entropy lexical cooccurrence information, and are empirically shown to improve embedding learning by performing notably better than several popular embedding models in similarity tasks. 1 Lexical Embeddings and Relatedness Lexical embeddings are essentially real-valued distributed representations of words. As a vectorspace model, an embedding model approximates semantic relatedness with the Euclidean distance between embeddings, the result of which helps better estimate the real lexical distribution in various NLP tasks. In recent years, researchers have developed efficient and effective algorithms for learning embeddings (Mikolov et al., 2013a; Pennington et al., 2014) and extended model applications from language modelling to various areas in NLP including lexical semantics (Mikolov et al., 2013b) and parsing (Bansal et al., 2014). To approximate semantic relatedness with geometric distance, objective functions are usually chosen to correlate positively with the Euclidean similarity between the embeddings of related words. Maximizing such an objective function is then equivalent to adjusting the embeddings so that those of the related words will be geometrically closer. The definition of relatedness among words can have a profound influence on the quality of the resulting embedding models. In most existing studies, relatedness is defined by co-occurrence within a window frame sliding over texts. Although supported by the distributional hypothesis (Harris, 1954), this definition suffers from two major limitations. Firstly, the window frame size is usually rather small (for efficiency and sparsity considerations), which increases the false negative rate by missing long-distance dependencies. Secondly, a window frame can (and often does) span across different constituents in a sentence, resulting in an increased false positive rate by associating unrelated words. The problem is worsened as the size of the window increases since each false-positive n-gram will appear in two subsuming false-positive (n+1)-grams. Several existing studies have addressed these limitations of window-based contexts. Nonetheless, we hypothesize that lexical embedding learning can further benefit from (1) factorizing syntactic relations into individual relations for structured syntactic information and (2) defining relatedness using lexicographic knowledge. We will show that implementation of these ideas brings notable improvement in lexical similarity tasks.