Learning Semantic Correspondences from Noisy Data-text Pairs by Local-to-Global Alignments

Proceedings of the 28th International Conference on Computational Linguistics, pages 3050–3059 Barcelona, Spain (Online), December 8-13, 2020 3050 Learning Semantic Correspondences from Noisy Data-text Pairs by Local-to-Global Alignments Feng Nie1,3∗ Jinpeng Wang2 Chin-Yew Lin2 1Interactive Entertainment Group, Tencent Inc, Shenzhen, China 2Microsoft Research Asia 3Sun Yat-Sen University {fengniesysu,wjppku}@gmail.com cyl@microsoft.com Abstract Learning semantic correspondences between structured input data (e.g., slot-value pairs) and associated texts is a core problem for many downstream NLP applications, e.g., data-to-text generation. Large-scale datasets recently proposed for generation contain loosely corresponding data text pairs, where part of spans in text cannot be aligned to its incomplete paired input. To learn semantic correspondences from such datasets, we propose a two-stage local-to-global alignment (L2GA) framework. First, a local model based on multi-instance learning is applied to build alignments for texts spans that can be directly grounded to its paired structured input. Then, a novel global model built upon a memory-guided conditional random field (CRF) layer aims to infer missing alignments for text spans not supported by paired incomplete inputs, where the memory is designed to leverage alignment clues provided by the local model to strengthen the global model. In this way, the local model and global model can work jointly to learn semantic correspondences in the same framework. Experimental results show that our proposed method can be generalized to both restaurant and computer domains and improve the alignment accuracy. 1 Introduction Discovering semantic correspondences between structured input (e.g., a set of slot-value pairs) and associated descriptions is a central step for many downstream NLP tasks. For example, in data-to-text generation, these correspondences indicate which subset of input data is verbalized in the description texts (what to say) and how data are described in natural language (how to say) (Angeli et al., 2010; Perez-Beltrachini and Lapata, 2018). Recent neural methods for data-to-text generation require using large-scale training corpus. These datasets are often automatically constructed from parallel structured data and descriptions in websites. Some of the collected instances inevitably contain noise where input structured data are semantically inequivalent with associated texts (Perez-Beltrachini and Gardent, 2017; Nie et al., 2019). Fig. 1 depicts an example. The slot-value pair Rating:low in the input meaning representation (MR) is contradicted with the text span highly recommended in descriptions, while the text span restaurant in descriptions should refer to a slot-value pair EatType:restaurant, which is not in the paired MR. Previous work (Barzilay and Lapata, 2005; Liang et al., 2009; Perez-Beltrachini and Lapata, 2018) aligns text spans in descriptions by only focusing on building semantic correspondences with the paired MR. However, in the scenario of semantically inequivalent MR-text pairs, due to the noise of input, there is not sufficient information to induce alignments for text spans such as restaurant by merely looking at its current input. To tackle the challenge of inducing correspondences (i.e., alignments) from semantically inequivalent data text pairs, we propose a local-to-global Alignment (L2GA) framework. This framework is composed of a local and a global model. The local model grounds text spans in …

Content truncated to 500 words