A New Distance Geometry Method for Constructing Word and Sentence Vectors

Leo Liberti (1)

1 LIX - Laboratoire d'informatique de l'École polytechnique [Palaiseau]

We present a new methodology for producing low-dimensional word vectors based on distance geometry and dimensional reduction techniques. We use these word vectors in order to construct sentence vectors. We evaluate their usefulness in a sentence classification task performed by a simple artificial neural network. Compared to n-gram incidence vectors, the new methodology is shown to yield lower loss values.

Mots clés: graph embeddings; natural language processing; word vectors; neural networks; clustering

Domaines: Informatique et langage [cs.CL]; Traitement du texte et du document

Leo Liberti. A New Distance Geometry Method for Constructing Word and Sentence Vectors. WWW '20: The Web Conference 2020 (DL4G satellite workshop), 2020, Taipei, Taiwan. pp.679-685, ⟨10.1145/3366424.3391267⟩. ⟨hal-02869238⟩