Bridging the gap between vectorial and mental word representations

In this paper, we combine two apparently disparate techniques for word representation: (a) word embeddings, which learn vectors from co-occurency statistics on large corpora (b) word associations, which capture mental representation of language. Existing literature has mainly considered these representations distinctly while proving that they respectively perform better at (a) similarity and (b) analogy tasks. We propose to combine these two sources of information and derive several hybrid models in order to obtain good performance on both tasks. In particular, our best method pairs Shortest Paths methods with an improved-version of Retrofitting models: it reaches an averaged 5% improvement over state-of-the-art on a standard test of 5 similarity and 2 analogy datasets.