Learning Graph Embeddings from WordNet-based Similarity Measures

Proceedings of the Eighth Joint Conference on Lexical and Computational Semantics (*SEM), pages 125–135 Minneapolis, June 6–7, 2019. c⃝2019 Association for Computational Linguistics 125 Learning Graph Embeddings from WordNet-based Similarity Measures Andrey Kutuzov1, Mohammad Dorgham2, Oleksiy Oliynyk2, Chris Biemann2, and Alexander Panchenko2,3 1Language Technology Group, University of Oslo, Oslo, Norway 2Language Technology Group, University of Hamburg, Hamburg, Germany 3Skolkovo Institute of Science and Technology, Moscow, Russia andreku@ifi.uio.no {5dorgham,6oliinyk,biemann,panchenko}@informatik.uni-hamburg.de Abstract We present path2vec, a new approach for learning graph embeddings that relies on struc- tural measures of pairwise node similarities. The model learns representations for nodes in a dense space that approximate a given user- defined graph distance measure, such as e.g. the shortest path distance or distance measures that take information beyond the graph struc- ture into account. Evaluation of the proposed model on semantic similarity and word sense disambiguation tasks, using various WordNet- based similarity measures, show that our ap- proach yields competitive results, outperform- ing strong graph embedding baselines. The model is computationally efficient, being or- ders of magnitude faster than the direct com- putation of graph-based distances. 1 Introduction Developing applications making use of large graphs, such as networks of roads, social media users, or word senses, often involves the design of a domain-specific graph node similarity measure sim : V × V → R defined on a set of nodes V of a graph G = (V, E). For instance, it can represent the shortest distance from home to work, a com- munity of interest in a social network for a user, or a semantically related sense to a given synset in WordNet (Miller, 1995). There exist a wide vari- ety of such measures greatly ranging in their com- plexity and design from simple deterministic ones, e.g. based on shortest paths in a network (Lea- cock and Chodorow, 1998) to more complex ones, e.g. based on random walks (Fouss et al., 2007; Pilehvar and Navigli, 2015; Lebichot et al., 2018). Naturally, the majority of such measures rely on walks along edges E of the graph, often resulting in effective, but prohibitively inefficient measures requiring complex and computationally expensive graph traversals. Also, there are measures that in addition take e.g. corpus information into account beyond what is directly given in the graph, see e.g. (Budanitsky and Hirst, 2006). We propose a solu- tion to this problem by decoupling development and use of graph-based measures. Namely, once a node similarity measure is defined, we learn vec- tor representations of nodes that enable efficient computation of this measure. We represent nodes in a graph with dense embeddings that are good in approximating such custom, e.g. application- specific, pairwise node similarity measures. Sim- ilarity computations in a vector space are several orders of magnitude faster than computations di- rectly using the graph. Additionally, graph em- beddings can be of importance in privacy-sensitive network datasets, since in this setup, explicitly storing edges is not required anymore. The main advantage over other graph embeddings is that our model can learn a custom …

Content truncated to 500 words