Training a Natural Language Generator From Unaligned Data

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 451–461, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Training a Natural Language Generator From Unaligned Data Ondˇrej Dušek and Filip Jurˇcíˇcek Charles University in Prague, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostranské námˇestí 25, CZ-11800 Prague, Czech Republic {odusek,jurcicek}@ufal.mff.cuni.cz Abstract We present a novel syntax-based natural language generation system that is train- able from unaligned pairs of input mean- ing representations and output sentences. It is divided into sentence planning, which incrementally builds deep-syntactic de- pendency trees, and surface realization. Sentence planner is based on A* search with a perceptron ranker that uses novel differing subtree updates and a simple fu- ture promise estimation; surface realiza- tion uses a rule-based pipeline from the Treex NLP toolkit. Our first results show that training from unaligned data is feasible, the outputs of our generator are mostly fluent and rele- vant. 1 Introduction We present a novel approach to natural lan- guage generation (NLG) that does not require fine- grained alignment in training data and uses deep dependency syntax for sentence plans. We include our first results on the BAGEL restaurant recom- mendation data set of Mairesse et al. (2010). In our setting, the task of a natural language generator is that of converting an abstract meaning representation (MR) into a natural language utter- ance. This corresponds to the sentence planning and surface realization NLG stages as described by Reiter and Dale (2000). It also reflects the in- tended usage in a spoken dialogue system (SDS), where the NLG component is supposed to trans- late a system output action into a sentence. While the content planning NLG stage has been used in SDS (e.g., Rieser and Lemon (2010)), we believe that deciding upon the contents of the system’s ut- terance is generally a task for the dialogue man- ager. We focus mainly on the sentence planning part in this work, and reuse an existing rule-based surface realizer to test the capabilities of the gen- erator in an end-to-end setting. Current NLG systems usually require a sepa- rate training data alignment step (Mairesse et al., 2010; Konstas and Lapata, 2013). Many of them use a CFG or operate in a phrase-based fashion (Angeli et al., 2010; Mairesse et al., 2010), which limits their ability to capture long-range syntactic dependencies. Our generator includes alignment learning into sentence planner training and uses deep-syntactic trees with a rule-based surface re- alization step, which ensures grammatical correct- ness of the outputs. Unlike previous approaches to trainable sentence planning (e.g., Walker et al. (2001); Stent et al. (2004)), our generator does not require a handcrafted base sentence planner. This paper is structured as follows: in Section 2, we describe the architecture of our generator. Sec- tions 3 and 4 then provide further details on its main components. In Section 5, we describe our experiments on the BAGEL data …

Content truncated to 500 words