Using Collaborative Filtering to Model Argument Selection

Proceedings of Recent Advances in Natural Language Processing, pages 629–639 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_071 629 Using Collaborative Filtering to Model Argument Selection Sagar Indurkhya Massachusetts Institute of Technology, Cambridge MA 02139, USA indurks@mit.edu & Virtualitics, Inc Abstract This study evaluates whether model-based Collaborative Filtering (CF) algorithms, which have been extensively studied and widely used to build recommender systems, can be used to predict which common nouns a pred- icate can take as its complement. We find that, when trained on verb-noun co-occurrence data drawn from the Corpus of Contempo- rary American-English (COCA), two popular model-based CF algorithms, Singular Value Decomposition and Non-negative Matrix Fac- torization, perform well on this task, each achieving an AUROC of at least 0.89 and sur- passing several different baselines. We then show that the embedding-vectors for verbs and nouns learned by the two CF models can be quantized (via application of k-means clus- tering) with minimal loss of performance on the prediction task while only using a small number of verb and noun clusters (relative to the number of distinct verbs and nouns). Fi- nally we evaluate the alignment between the quantized embedding vectors for verbs and the Levin verb classes, finding that the alignment surpassed several randomized baselines. We conclude by discussing how model-based CF algorithms might be applied to learning re- strictions on constituent selection between var- ious lexical categories and how these (learned) models could then be used to augment a (rule- based) constituency grammar. 1 Introduction In learning a language, a child solves many dif- ficult puzzles using limited input data (Berwick et al., 2011; Piattelli-Palmarini and Berwick, 2012; Lasnik and Lidz, 2017). One such puzzle involves the child deciding whether a given verb can take a particular noun as its complement when the child has never previously observed that verb and that noun co-occur in a sentence. To illustrate this puz- zle, let us consider an example - suppose the child learner has heard the following sentences: (a) “I lied and and said that I would not smash the windshield.” (b) “The robber was not planning to smash every plate.” (c) “I want to take this hammer and smash the precious vase!” (d) “They are going to shatter the windshield!” (e) “The boy who was busy staring at his phone will trip over and shatter his mother’s favorite plate.” (f) “Did you shatter the blue vase?” (g) “She knew that Susan would break her expen- sive new windshield.” (h) “I saw the man in the red sweater break the delicate plate.” Now suppose that the child learner has never heard a sentence in which the verb “break” takes the noun “vase” as its complement. How should the child decide whether the following production is licit? (i) “He is going to trip over and break the vase!” One strategy that the child might employ is as follows. First the child observes that the verbs “smash” and “shatter” behave similarly by noting that both verbs can select any of the three nouns …

Content truncated to 500 words