Few-Shot Multi-Hop Relation Reasoning over Knowledge Bases

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 580–585 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 580 Few-Shot Multi-Hop Relation Reasoning over Knowledge Bases Chuxu Zhang1, Lu Yu3, Mandana Saebi2, Meng Jiang2, Nitesh V. Chawla2 1Brandeis University, 2University of Notre Dame, 3King Abdullah University of Science and Technology 1chuxuzhang@brandeis.edu, 2{msaebi, mjiang2, nchawla}@nd.edu, 3lu.yu@kaust.edu.sa Abstract Multi-hop relation reasoning over knowledge base is to generate effective and interpretable relation prediction through reasoning paths. The current methods usually require sufficient training data (fact triples) for each query re- lation, impairing their performances over few- shot relations (with limited triples) which are common in knowledge base. To this end, we propose FIRE, a novel few-shot multi- hop relation learning model. FIRE applies reinforcement learning to model the sequen- tial steps of multi-hop reasoning, besides per- forms heterogeneous structure encoding and knowledge-aware search space pruning. The meta-learning technique is employed to op- timize model parameters that could quickly adapt to few-shot relations. Empirical study on two datasets demonstrate that FIRE outper- forms state-of-the-art methods. 1 Introduction Nowadays large scale knowledge bases (KB), e.g., NELL (Mitchell et al., 2018) or Freebase (Bol- lacker et al., 2008), are serving as useful resources for many natural language processing applications such as semantic search or question answering. Due to the nature of incompleteness (Bordes et al., 2013), it is essential to automate the KB comple- tion. One typical problem is fact (triple) prediction. For example, given a query “What is the national- ity of Barack Obama?” denoted as (Barack Obama, Nationality, ?), the task is to infer USA as the an- swer. There have been a lot of work for solving this problem by embedding learning approaches (Bor- des et al., 2013; Socher et al., 2013; Yang et al., 2015) or deep learning models (Dettmers et al., 2018; Schlichtkrull et al., 2018). The fact prediction ignores the compositional relations in KB and results answer that lacks of interpretation. Accordingly, an alternative prob- lem, multi-hop relation reasoning, was presented. The task is to infer facts using multi-hop reasoning paths, e.g., (Barack Obama, BornIn, Hawaii) ∧ (Hawaii, LocateIn, USA) → (Barack Obama, Na- tionality, USA). A number of recent models (Xiong et al., 2017; Das et al., 2018; Lv et al., 2019) for- mulate the problem as sequential decision process and leverage reinforcement learning to achieve con- siderable performance. Most current multi-hop relation reasoning mod- els require a good amount of training data (fact triples) for each query relation. However, the rela- tion frequency distribution in KB is usually long- tail (Xiong et al., 2018), showing that a large por- tion of relations only have few-shot fact triples for model training. Despite that some few-shot relation learning methods (Chen et al., 2019; Lv et al., 2019; Zhang et al., 2020) have been proposed recently, they target at fact prediction only or their perfor- mance is suboptimal due to deficiency in capturing heterogeneous structural information and pruning search space in KB. In this work, we aim at …

Content truncated to 500 words