Absinth: A small world approach to word sense induction

ABSINTH: A small world approach to word sense induction Victor Zimmermann1, Maja Hoffmann2 Leipzig University1, Leipzig University of Applied Sciences (HTWK)2 victor.zimmermann@uni-leipzig.de maja katharina.hoffmann@stud.htwk-leipzig.de Abstract ABSINTH1 provides a novel unsupervised graph-based approach to word sense induction. This work combines small world co¨occurrence networks with a graph propagation algorithm to induce per-word sense assignment vectors over a lexicon that can be aggregated for clas- sification of whole snippets. 1 Introduction As late as twelve years after publication, the graph- based approach to word sense induction proposed in V´eronis (2004) was still cited as ’state-of-the- art’ (Tripodi and Pelillo 2017, Ustalov et al. 2017) and only recently surpassed by neural substitution- based approaches (Amrami and Goldberg 2018, Amrami and Goldberg 2019). Our goal with this work is to evaluate an approach native to small- world graphs for the word sense induction task. We build on the principles laid out in Hyperlex (V´eronis, 2004) with a more dynamic feature set and a graph propagation algorithm previously used for sentiment analysis (Hamilton et al., 2016). Our system, ABSINTH1, provides a simple two- step approach to SemEval-2013 Task 11 (Navigli and Vannella, 2013). To achieve this, we utilise the properties of small world graphs for language (Cancho and Sol´e, 2001) in general and semantic relations (Newman, 2003) in particular. We extract senses using the root hub algorithm proposed in V´eronis (2004). For word sense disambiguation we use the sense inventory created in previous steps and a graph propagation algorithm to assign each node a sense distribution vector. Lastly, the vectors of each word in a given context are summed up and the context is assigned the sense of the best cumulative weight. 1Association Based Semantic Induction Tools for root Hub propagation Parameter ABSINTH Hyperlex Baseline Min. context 4 4 4 Min. #nodes Avg. #nodes 10 9 Min. #edges Avg. #edges 5 3 Max. weight 0.9 0.9 0.9 Table 1: Minimum context size, minimum number of nodes, minimum number of edges and maximum edge weight for our system, Hyperlex and our Baseline. In addition to the SemEval scoring methods to eval- uate our results we use characteristic path length and global clustering co¨efficient to evaluate the properties of our co¨occurrence graphs. Our system achieves better results in three out of four metrics than a classifier similar to Hyperlex without label propagation. 2 Related Work Graph-based approaches to word sense induction have been successfully used since the early 2000s (V´eronis 2004, Di Marco and Navigli 2013, Am- playo et al. 2019). V´eronis proposes the use of root hub detection and minimum spanning trees (Kruskal, 1956) to induce senses and disambiguate search results. The usefulness of small world graph properties for sense disambiguation has previously been shown in Newman (2003). The term ’small world’ was introduced by Travers and Milgram, using it to de- scribe the connectedness of acquaintance networks (Travers and Milgram, 1969). According to their findings, the average path length between two peo- ple living in the United States lies around five or six, even though …

Content truncated to 500 words