Automatic Food Categorization from Large Unlabeled Corpora and Its Impact on Relation Extraction

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 673–682, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Automatic Food Categorization from Large Unlabeled Corpora and Its Impact on Relation Extraction Michael Wiegand and Benjamin Roth and Dietrich Klakow Spoken Language Systems Saarland University D-66123 Saarbr¨ucken, Germany {Michael.Wiegand|Benjamin.Roth|Dietrich.Klakow}@lsv.uni-saarland.de Abstract We present a weakly-supervised induc- tion method to assign semantic informa- tion to food items. We consider two tasks of categorizations being food-type classi- fication and the distinction of whether a food item is composite or not. The cate- gorizations are induced by a graph-based algorithm applied on a large unlabeled domain-specific corpus. We show that the usage of a domain-specific corpus is vi- tal. We do not only outperform a manually designed open-domain ontology but also prove the usefulness of these categoriza- tions in relation extraction, outperforming state-of-the-art features that include syn- tactic information and Brown clustering. 1 Introduction In view of the large interest in food in many parts of the population and the ever increasing amount of new dishes/food items, there is a need of au- tomatic knowledge acquisition. We approach this task with the help of natural language processing. We investigate different methods to assign cate- gories to food items. We focus on two categoriza- tions, being a classification of food items to cat- egories of the Food Guide Pyramid (U.S. Depart- ment of Agriculture, 1992) and a categorization of whether a food item is composite or not. We present a semi-supervised graph-based ap- proach to induce these food categorizations from an unlabeled domain-specific text corpus crawled from the Web. The method only requires mini- mal manual guidance for the initialization of the algorithm with seed terms. It depends, however, on an automatically constructed high-quality sim- ilarity graph. For that we choose a pattern-based representation that outperforms a distributional- based representation. For initialization, we ex- amine some manually compiled seed words and a very few simple surface patterns to automati- cally induce such expressions. As a hard baseline, we compare the effectiveness of using a general- purpose ontology for the same types of categoriza- tions. Apart from an intrinsic evaluation, we also examine the categories in relation extraction. The contributions of this paper are a method re- quiring minimal supervision for a comprehensive classification of food items and a proof of con- cept that the knowledge that can thus be gained is beneficial for relation extraction. Even though we focus on a specific domain, the induction method can be easily translated to other domains. In par- ticular, other life-style domains, such as fashion, cosmetics or home & gardening, show parallels since comparable textual web data are available and similar relation types (e.g. that two items fit together or can be substituted by each other) exist. Our experiments are carried out on German data but our findings should carry over to other lan- guages since the issues we address are (mostly) language universal. For general accessibility, all examples are …

Content truncated to 500 words