Embedding Words in Non-Vector Space with Unsupervised Graph Learning

TL;DR

This paper introduces GraphGlove, an unsupervised method for learning word representations as weighted graphs, capturing hierarchical structures more naturally than traditional vector embeddings, and demonstrating superior performance on linguistic tasks.

Contribution

It proposes a novel end-to-end unsupervised graph-based approach for word embeddings that better encodes hierarchical language structures.

Findings

GraphGlove outperforms vector-based methods on similarity and analogy tasks.

The learned graph structures are hierarchical and resemble WordNet.

The graph geometry is complex, with diverse local topologies.

Abstract

It has become a de-facto standard to represent words as elements of a vector space (word2vec, GloVe). While this approach is convenient, it is unnatural for language: words form a graph with a latent hierarchical structure, and this structure has to be revealed and encoded by word embeddings. We introduce GraphGlove: unsupervised graph word representations which are learned end-to-end. In our setting, each word is a node in a weighted graph and the distance between words is the shortest path distance between the corresponding nodes. We adopt a recent method learning a representation of data in the form of a differentiable weighted graph and use it to modify the GloVe training algorithm. We show that our graph-based representations substantially outperform vector-based methods on word similarity and analogy tasks. Our analysis reveals that the structure of the learned graphs is hierarchical and similar to that of WordNet, the geometry is highly non-trivial and contains subgraphs with different local topology.