Learning Compositionality Functions on Word Embeddings for Modelling Attribute Meaning in Adjective-Noun Phrases

Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 54–64, Valencia, Spain, April 3-7, 2017. c⃝2017 Association for Computational Linguistics Learning Compositionality Functions on Word Embeddings for Modelling Attribute Meaning in Adjective-Noun Phrases Matthias Hartung Fabian Kaupmann Soufian Jebbara Philipp Cimiano Semantic Computing Group CITEC, Bielefeld University {mhartung,fkaupmann,sjebbara,cimiano}@techfak.uni-bielefeld.de Abstract Word embeddings have been shown to be highly effective in a variety of lexical se- mantic tasks. They tend to capture mean- ingful relational similarities between in- dividual words, at the expense of lack- ing the capabilty of making the underly- ing semantic relation explicit. In this pa- per, we investigate the attribute relation that often holds between the constituents of adjective-noun phrases. We use CBOW word embeddings to represent word mean- ing and learn a compositionality function that combines the individual constituents into a phrase representation, thus captur- ing the compositional attribute meaning. The resulting embedding model, while be- ing fully interpretable, outperforms count- based distributional vector space models that are tailored to attribute meaning in the two tasks of attribute selection and phrase similarity prediction. Moreover, as the model captures a generalized layer of at- tribute meaning, it bears the potential to be used for predictions over various attribute inventories without re-training. 1 Introduction Attributes such as SIZE, WEIGHT or COLOR are part of the building blocks of representing knowl- edge about real-world entities or events (Barsalou, 1992). In natural language, formal attributes find their counterpart in attribute nouns which can be used in order to generalize over individual proper- ties, e.g., big or small in case of SIZE, blue or red in case of COLOR (Hartung, 2015). In order to ascribe such properties to entities or events, adjective-noun phrases are a very frequent linguistic pattern. In these constructions, attribute meaning is conveyed only implicitly, i.e., with- out being overtly realized at the phrasal surface. Hence, attribute selection has been defined as the task of predicting the hidden attribute meaning ex- pressed by a property-denoting adjective in com- position with a noun (Hartung and Frank, 2011b), as in the following examples: (1) a. hot summer → TEMPERATURE b. hot debate → EMOTIONALITY c. hot soup → TASTE/TEMPERATURE Previous work on this task has largely been car- ried out in distributional semantic models (cf. Har- tung (2015) for an overview). In the face of the re- cent rise of distributed neural representations as a means of capturing lexical meaning in NLP tasks (Collobert et al., 2011; Mikolov et al., 2013a; Pen- nington et al., 2014), our goal in this paper is to model attribute meaning based on word embed- dings. In particular, we use CBOW embeddings of adjectives and nouns (Mikolov et al., 2013a) as underlying word representations and train a com- positionality function in order to compute a phrase representation that is predictive of the implicitly conveyed attribute meaning. In fact, word embeddings (also referred to as predict models) have been shown to be highly ef- fective in …

Content truncated to 500 words