Selective Decoding for Cross-lingual Open Information Extraction

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 832–842, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Selective Decoding for Cross-lingual Open Information Extraction Sheng Zhang Johns Hopkins University zsheng2@jhu.edu Kevin Duh Johns Hopkins University kevinduh@cs.jhu.edu Benjamin Van Durme Johns Hopkins University vandurme@cs.jhu.edu Abstract Cross-lingual open information extrac- tion is the task of distilling facts from the source language into representations in the target language. We propose a novel encoder-decoder model for this problem. It employs a novel selective de- coding mechanism, which explicitly mod- els the sequence labeling process as well as the sequence generation process on the decoder side. Compared to a stan- dard encoder-decoder model, selective de- coding significantly increases the perfor- mance on a Chinese-English cross-lingual open IE dataset by 3.87-4.49 BLEU and 1.91-5.92 F1. We also extend our ap- proach to low-resource scenarios, and gain promising improvement. 1 Introduction Cross-lingual open information extraction is de- fined as the task of extracting facts from the source language (e.g., Chinese text in Fig. 1(a)) and rep- resenting them in the target language (e.g. English predicate-argument information in Fig. 1(c))1. It is a challenging task and of great importance to solve the cross-lingual portability issues of various NLP systems which are in the support of open informa- tion extraction (Sudo et al., 2004). Additionally, there is often a great demand for rapid access to information across languages, especially when a large-scale incident occurs (Lu et al., 2016). Conventional solutions decompose the task as a pipeline of machine translation followed by 1The predicate-argument information is normally repre- sented by relation tuples. Here, we use a richer representation (i.e., a tree structure) adopted by PredPatt (White et al., 2016), a lightweight tool for identifying predicate-argument infor- mation, available at https://github.com/hltcoe/ PredPatt. P [ Soldiers A started P mortars A ] P P [ Soldiers A firing P ] P [(Soldiers:ah) started:ph firing:ph (mortars:ah )]] [(Soldiers:ah) ��.������������ (a) (d) (e) started firing Soldiers Soldiers mortars ARG ARG ARG ARG SOMETHING (c) ������������������������������� (b) Figure 1: Example of cross-lingual open IE: Chinese input text (a), English translation (b), English predicate-argument information (c), lin- earized PredPatt output (d) and output with sep- arated predicate and argument labels (e). open information extraction (or vice versa), which causes a deviation since machine translation at- taches equal importance to adequacy and fluency of the intermediate translation results (Snover et al., 2009), whereas the final goal focuses more on extracting correct predicates and arguments. Recently Zhang et al. (2017a) proposes an end- to-end solution that outperforms the conventional pipeline solutions. They recast cross-lingual open IE as a sequence-to-sequence learning problem by converting the target facts in the tree struc- ture (Fig. 1(c)) into a linear form called lin- earized PredPatt2 (Fig. 1(d)), and employ a stan- 2Linearized PredPatt is converted from the PredPatt tree structure by taking an in-order traversal of every node in the tree. See Zhang et al. (2017a) for details. 832 dard encoder-decoder model to address the …

Content truncated to 500 words