Imitation learning for language generation from unaligned data

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1101–1112, Osaka, Japan, December 11-17 2016. Imitation learning for language generation from unaligned data Gerasimos Lampouras Department of Computer Science University of Sheffield, UK g.lampouras@sheffield.ac.uk Andreas Vlachos Department of Computer Science University of Sheffield, UK a.vlachos@sheffield.ac.uk Abstract Natural language generation (NLG) is the task of generating natural language from a meaning rep- resentation. Rule-based approaches require domain-specific and manually constructed linguistic resources, while most corpus based approaches rely on aligned training data and/or phrase tem- plates. The latter are needed to restrict the search space for the structured prediction task defined by the unaligned datasets. In this work we propose the use of imitation learning for structured prediction which learns an incremental model that handles the large search space while avoid- ing explicitly enumerating it. We adapted the Locally Optimal Learning to Search (Chang et al., 2015) framework which allows us to train against non-decomposable loss functions such as the BLEU or ROUGE scores while not assuming gold standard alignments. We evaluate our approach on three datasets using both automatic measures and human judgements and achieve results comparable to the state-of-the-art approaches developed for each of them. Furthermore, we performed an analysis of the datasets which examines common issues with NLG evaluation. 1 Introduction Natural language generation (NLG) is the task of generating natural language from a machine- interpretable meaning representation (MR). Traditionally, NLG systems tend to be rule-based and require domain-specific language resources (i.e. sentence plans, aggregation rules, ordering information, etc.) to be manually authored for a particular task, e.g. weather reporting (Reiter et al., 2005), route navigation (Dale et al., 2003), or ontology descriptions (Bontcheva and Wilks, 2004; Androutsopoulos et al., 2013). From a machine learning perspective (ML) perspective, NLG is a complex structured prediction task due to its large output space, which is the set of all possible NL utterances. To limit the output space, many ML-based approaches rely on aligned training datasets (Mairesse et al., 2010; Dethlefs et al., 2013) that specify the words that each MR element is responsible for, and phrase templates. While these enable a variety of models to be trained, their applicability is restricted as they need additional manual annotation. More recently, ML-based approaches focused on learning NLG models from unaligned training data. Duˇsek and Jurc´ıcek (2015) learn how to incrementally generate deep-syntax dependency trees of candi- date sentence plans specifying which MR elements to be mentioned and the overall sentence structure, however the surface realization (i.e. the actual NL word generation) is performed using rules based on the frames of an English-Czech dependency treebank, a resource that may not be available for certain domains or languages. In another approach, Wen et al. (2015) use a Long Short-term Memory (LSTM) network to learn from unaligned data and jointly address sentence planning and surface realization. They augment each cell of the LSTM with a gate that conditions it on the input MR. However the LSTM is trained using cross-entropy …

Content truncated to 500 words