Distributional and Neural Models for Extracting Manipulation-Relevant Relations from Text Corpora

Tracking #: 1542-2754

This paper is currently under review
Soufian Jebbara
Valerio Basile
Elena Cabrio
Philipp Cimiano

Responsible editor: 
Guest Editors ML4KBG 2016

Submission type: 
Full Paper
In this paper we present a novel approach based on neural network techniques to extract common sense knowledge from text corpora. We apply this approach to extract of common sense knowledge about everyday objects that can be used by intelligent machines, e.g. robots, to support planning of tasks that involve object manipulation. The knowledge we extract is constituted by relations that relate some object (type) to some other entity such as its typical location or typical use. Our approach builds on the paradigm of distributional semantics and frames the task of extracting such relations as a ranking problem. Given a certain object, the goal is to extract a ranked list of locations or uses ranked by `prototypicality'. This ranking is computed via a similarity score in the distributional space. We compare different techniques for constructing this semantic distributional space. On the one hand, we use the well known SkipGram model to embed words into a low-dimensional distributional space, using cosine similarity to rank the various candidates. We also consider an approach in the spirit of latent semantic indexing that relies on the NASARI approach to compute low-dimensional representations that are also ranked by cosine similarity. While both methods were already proposed in earlier work, as main contribution in this paper we present a neural network approach in which the ranking or scoring function is directly learned using a supervised approach. We compare all these approaches showing superiority of the neural network approach for some evaluation measures compared to the other two approaches described above for the construction of the distributional space.
Full PDF Version: 
Under Review