Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence (IJCAI-17) Learning to Explain Entity Relationships by Pairwise Ranking with Convolutional Neural Networks Jizhou Huangy;z∗, Wei Zhangz, Shiqi Zhaoz, Shiqiang Dingz, Haifeng Wangz yResearch Center for Social Computing and Information Retrieval, Harbin Institute of Technology, China zBaidu Inc., Beijing, China fhuangjizhou01, zhangwei32, zhaoshiqi, dingshiqiang01, [email protected] Abstract Providing a plausible explanation for the relation- ship between two related entities is an important task in some applications of knowledge graphs, such as in search engines. However, most exist- ing methods require a large number of manually labeled training data, which cannot be applied in large-scale knowledge graphs due to the expensive data annotation. In addition, these methods typi- cally rely on costly handcrafted features. In this paper, we propose an effective pairwise ranking model by leveraging clickthrough data of a Web search engine to address these two problems. We Figure 1: An example of entity recommendation with entity first construct large-scale training data by leverag- relationship explanation. ing the query-title pairs derived from clickthrough data of a Web search engine. Then, we build a pairwise ranking model which employs a convo- nations can significantly increase the understandability of the lutional neural network to automatically learn rel- recommendations and user engagement [Fang et al., 2011; evant features. The proposed model can be easily Huang et al., 2016]. It is therefore important to find descrip- trained with backpropagation to perform the rank- tive sentences that explain the relationship between two enti- ing task. The experiments show that our method ties. significantly outperforms several strong baselines. Although there is a growing interest in predicting relations between entities, e.g., [Bordes et al., 2013; Wang et al., 2014; 1 Introduction Lin et al., 2015; Ji et al., 2016], the relation types defined by Due to its heterogeneity, semantic richness, and large scale, the knowledge graph, such as “Spouse” and “Offspring”, can knowledge graph has been widely used in search applications, hardly be directly used as descriptions in the aforementioned such as search results enrichment with semantic information application in search engines, which are not explanatory or [Singhal, 2012] and entity recommendation [Blanco et al., informative enough. To better explain why two entities are 2013; Yu et al., 2014]. Presenting evidence on why two en- related in a knowledge graph, it is important to provide more tities are related in a knowledge graph can help in building detailed evidence about the relationship between them. To the and promoting trust between the users and the entity recom- best of our knowledge, the problem of automatically finding a mendation system [Voskarides et al., 2015]. Figure 1 shows descriptive sentence that explains a given relationship of two an example of Baidu1 Web search engine’s entity recommen- entities has not been well addressed. dations for the query “Obama”. In this example, a short Although using a template-based method to match or gen- sentence is presented under each entity, e.g., “Barack and erate a descriptive sentence for this task is a straightforward Michelle married in 1992 and have two daughters”, provid- method, it has two major limitations. First, it is usually re- ing users an explanation of the relationship between the rec- quired to manually label a number of seeds or candidate tem- ommended entity “Michelle Obama” and the queried entity plates for each relationship. Therefore, it can hardly be ap- “Obama”.2 Presenting related entities with plausible expla- plied in large-scale knowledge graphs due to the expensive ∗Corresponding author. annotation. Second, this method can achieve high precision 1https://www.baidu.com/ but low recall due to strict or limited templates. 2We translate the example from Chinese to English for the sake Recently, Voskarides et al. [2015] first studied the problem of understanding. of explaining relationships between entities. They modeled 4018 Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence (IJCAI-17) this task as a ranking problem, and proposed a learning to information to help users understand the relevance and rela- rank (LTR) method with a rich set of handcrafted features tionship between an entity and a query. The major distinction to rank the sentences within Wikipedia3 articles according to between our work and this study is that our method focuses on how well they describe the relationship between the entities. explaining the relationship between two entities, rather than Although the method achieved promising results in the ex- between an entity and a query. Fang et al. [2011] proposed periments, it has two major drawbacks when applied to large- to produce a ranked list of relationships that describe how a scale, real-world problems. pair of entities are related based on a knowledge graph. For First, large-scale training data is essential for building example, for “Angelina Jolie” and “Brad Pitt”, it generates ranking models based on supervised machine learning tech- “Spouse” and “Co-starring” as the explanations of their con- niques. To train the ranking model, a total of 5,689 sentences nections. Banko et al. [2007] and Fader et al. [2011] pro- for 1,476 entity pairs are manually labeled as training and posed to extract a large set of relational tuples using Open test data in [Voskarides et al., 2015]. However, it gets too Information Extraction (IE) paradigm. Our work is different expensive if we want to generate a much larger scale training in that we extract sentences that are eligible for explaining the dataset, which is essential in training neural network models relationship of an entity pair, rather than recognize the types as we do in this paper. of the relationships. Second, this method employs elaborately designed fea- tures. However, since errors exist inevitably in the feature 3 Methodology extraction process, the use of handcrafted features leads to error propagation or accumulation [Zeng et al., 2015]. With We study the problem of explaining relationships between en- the recent revival of interest in neural networks, many re- tities by leveraging clickthrough data, and model this task as a searchers have tried to use neural networks to automatically sentence ranking problem. More specifically, given a relation learn features. It has been shown that this methd achieves triplet qs consisted of a pair of entities eh and et, and a rela- significant improvements compared with manually designed tionship rk between them, i.e., qs = (eh; rk; et), our task is features in several tasks, e.g., [Socher et al., 2012; Kim, 2014; to extract a set of candidate sentences S = fs1; s2; ··· ; sng Zeng et al., 2015; Santos et al., 2015]. Inspired by these stud- that contain eh and et, and learn a ranking function to rank ies, we also employ a convolutional neural network (CNN) to the sentences according to how well they describe the re- automatically learn relevant features in our task. lationship rk between eh and et. In this paper, we focus To address the two problems above, this paper proposes an on the ranking problem, and the candidate sentences are ex- effective pairwise ranking model by leveraging clickthrough tracted using the keyword based retrieval method proposed data. We first automatically extract large-scale training data in [Voskarides et al., 2015]. In what follows, we first intro- by leveraging the query-title pairs derived from clickthrough duce the strategy we used to acquire large amounts of data for data of a Web search engine. No manually labeled training training ranking models. Then, we present our pairwise rank- data are needed. More specifically, we consider a special kind ing model that employs a CNN to automatically learn relevant of query which consists of a relation triplet (an entity pair and features, which is referred to as PR-CNN hereafter. their relationship), and view the corresponding clicked titles of the query as potential descriptions of the relationship be- 3.1 Training Data Acquisition tween two entities mentioned in the query. We then build a The acquisition of training data in the task of explaining rela- pairwise ranking model that employs a CNN to automatically tionships between entity pairs is crucial. Since publicly avail- learn relevant features, in which no handcrafted features are able datasets are quite limited, previous work [Voskarides et needed. Finally, we use the trained model to rank the sen- al., 2015] manually labeled the training data by asking human tences according to how well they describe the relationship assessors to assess the quality of each candidate description. between the entities. The experimental results on a manually However, this method is expensive and limited in quantity. constructed test set show that our method significantly out- Therefore, the training and evaluation are both restricted to a performs several strong baselines. small number of labeled instances, making it difficult to meet the demand of large-scale, real-world applications. To alle- 2 Related Work viate this problem, we propose to acquire large amounts of training data by leveraging clickthrough data of a Web search Previous work that is the closest to our task is selecting sen- engine. tences that describe a particular relationship of two entities in The motivation of our approach is as follows. Given a [Voskarides et al., 2015]. However, there are some significant query qs that consists of an entity relation triplet (eh; rk; et), differences. First, we employ a CNN to automatically learn if a user clicks on a document title t that contains both entities relevant features directly from the training examples, rather eh and et of the triplet when searching for qs, then it is likely than use handcrafted features. Second, we propose an effec- that the title t is a description of the entity relation triplet.