Cross-lingual Knowledge Projection Using Machine Translation and Target-side Knowledge Base Completion

Proceedings of the 27th International Conference on Computational Linguistics, pages 1508–1520 Santa Fe, New Mexico, USA, August 20-26, 2018. 1508 Cross-lingual Knowledge Projection Using Machine Translation and Target-side Knowledge Base Completion Naoki Otani∗ Carnegie Mellon University notani@cs.cmu.edu Hirokazu Kiyomaru Kyoto University kiyomaru@nlp.ist.i.kyoto-u.ac.jp Daisuke Kawahara Kyoto University dk@i.kyoto-u.ac.jp Sadao Kurohashi Kyoto University kuro@i.kyoto-u.ac.jp Abstract Considerable effort has been devoted to building commonsense knowledge bases. However, they are not available in many languages because the construction of KBs is expensive. To bridge the gap between languages, this paper addresses the problem of projecting the knowledge in English, a resource-rich language, into other languages, where the main challenge lies in projection ambi- guity. This ambiguity is partially solved by machine translation and target-side knowledge base completion, but neither of them is adequately reliable by itself. We show their combination can project English commonsense knowledge into Japanese and Chinese with high precision. Our method also achieves a top-10 accuracy of 90% on the crowdsourced English–Japanese bench- mark. Furthermore, we use our method to obtain 18,747 facts of accurate Japanese commonsense within a very short period. 1 Introduction Commonsense has been considered to play a vital role in language understanding (LoBue and Yates, 2011), and considerable effort has been devoted to building knowledge bases (KBs) that organize com- monsense (Zang et al., 2013). The largest multi-lingual commonsense KB is ConceptNet (Liu and Singh, 2004b). ConceptNet maintains knowledge as a triple of two concepts and relationship between them, which we call fact. The characteristic of ConceptNet is that concepts are represented in undisambiguated forms of words or phrases, which facilitates commonsense acquisition and inference in practice (Liu and Singh, 2004a) and recently has proven useful for building word representations (Speer et al., 2017; Camacho-Collados et al., 2017). We target ConceptNet in this paper. A major problem lies in a large gap of quantity and quality between languages. The latest release (v5.5.0) of ConceptNet has 2,828,394 unique English facts1, but the number of Japanese facts in Con- ceptNet is only 69,902 (≈ 2.5%) even though Japanese knowledge takes up the eighth-largest portion of the database. This problem is not specific to ConceptNet. English KBs are typically larger and of higher quality than other languages. Although an adequate amount of knowledge of named entities is often available in many languages thanks to semi-structured text on the web such as Wikipedia in- fobox (Lehmann et al., 2014), commonsense is hard to obtain due to the lack of tractable and objective information (Gordon and Van Durme, 2013). It is not realistic to develop large knowledge resources in every language from scratch because of cost constraints. Instead, this paper focuses on cross-lingual knowledge projection. We translate English commonsense facts into a target language, aiming to gain large commonsense resources in the target language efficiently. The main challenge is projection ambiguity. For example, consider translating (bat, CapableOf, fly) shown in Figure 1. Bat has four Japanese translations such as koumori [bat (animal)] and batto [bat (stick)]. Fly has 27 translations such as …

Content truncated to 500 words