Text-Based Joint Prediction of Numeric and Categorical Attributes of Entities in Knowledge Bases

Proceedings of Recent Advances in Natural Language Processing, pages 1194–1202, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_137 1194 Text-Based Joint Prediction of Numeric and Categorical Attributes of Entities in Knowledge Bases V Thejas BITS Pilani, India Abhijeet Gupta and Sebastian Pad´o IMS, University of Stuttgart, Germany Abstract Collaboratively constructed knowledge bases play an important role in information systems, but are essentially always incomplete. Thus, a large number of models has been devel- oped for Knowledge Base Completion, the task of predicting new attributes of entities given partial descriptions of these entities. Vir- tually all of these models either concentrate on numeric attributes (<Italy,GDP,2T$>) or they concentrate on categorical attributes (<Tim Cook,chairman,Apple>). In this paper, we propose a simple feed- forward neural architecture to jointly predict numeric and categorical attributes based on embeddings learned from textual occurrences of the entities in question. Following insights from multi-task learning, our hypothesis is that due to the correlations among attributes of dif- ferent kinds, joint prediction improves over separate prediction. Our experiments on seven FreeBase domains show that this hypothesis is true of the two attribute types: we find substantial improve- ments for numeric attributes in the joint model, while performance remains largely unchanged for categorical attributes. Our analysis indi- cates that this is the case because categorical attributes, many of which describe member- ship in various classes, provide useful ’back- ground knowledge’ for numeric prediction, while this is true to a lesser degree in the in- verse direction. 1 Introduction Collaboratively constructed knowledge bases (CCKBs) such as WikiData (Vrandeˇci´c and Kr¨otzsch, 2014), YAGO (Suchanek et al., 2008), FreeBase (Bollacker et al., 2008) or DBPedia (Bizer et al., 2009), capture world knowledge in the shape of a graph structure where nodes de- note entities and edges denote attributes (Hitzler et al., 2009). Their collaborative construction im- portantly enables them to avoid the scaling prob- lems encountered by expert-constructed knowledge bases. Thus, CCKBs have come to play an impor- tant role in information systems, forming the basis for a wide range of natural language processing applications (Hovy et al., 2013) such as question answering (Berant et al., 2013; Krishnamurthy and Mitchell, 2015) or representation learning for enti- ties (Toutanova et al., 2015; Yaghoobzadeh et al., 2018). The most crucial shortcoming of CCKBs is their incompleteness (Min et al., 2013; West et al., 2014) – not just with respect to the entities that they cover, but also with respect to the attributes present for entities that are nominally covered. This is not surprising: When a contributor to a knowledge base adds an entity, they will probably concentrate on the most salient attributes (e.g., for a scientist, field or affiliation), while other attributes (such as parents or place of birth) may be added later or never. This realization has led to a large boost to work in the area of knowledge base completion, that is, the prediction of attributes of entities that are currently missing from the CCKB (Bordes et al., 2013; Socher et al., 2013; Min et …

Content truncated to 500 words