Statistical Approaches to Adaptive Natural Language Generation

Abstract

Employing statistical models of users, generation contexts and of natural languages themselves has several potentially beneficial features: the ability to train models on real data, the availability of precise mathematical methods for optimisation, and the capacity to adapt robustly to previously unseen situations. This chapter will describe recent statistical approaches to generating natural language in spoken dialogue systems (SDS), covering several methods developed in the CLASSIC project: Reinforcement Learning (RL) approaches as developed by [22, 25, 33, 35, 44] and language generation using graphical models and active learning with data collected through crowd-sourcing [36]. The former approach has been applied to higher-level decisions in natural language generation (NLG) such as content and attribute selection [33, 35, 44, 45], as well as some low-level decisions such as lexical choice [22, 25], while the latter focusses on data-driven surface realisation. (See [50] and Fig. 6.1 for a representation of the NLG decision structure for SDS.) In this chapter, the RL approaches are presented in Sects. 6.3 and 6.4, and work using graphical models is discussed in Sect. 6.5. We also discuss recent work which jointly optimises higher-level and lower-level NLG decisions, using combinations of these types of approaches [12, 13], in Sect. 6.6. The reinforcement learning approach presented below shares many features with more traditional planning approaches to NLG, but it uses statistical machine learn-ing models to develop adaptive NLG components for SDS. Rather than emulating human behaviour in generation (which can be sub-optimal), these methods can even find strategies for NLG which improve upon human performance. Some very encouraging test results have recently been obtained with real users of systems developed using these methods, which we will survey here. For example, in an evaluation with real users (see Sect. 6.3.5), a trained information presentation Input: Communicative Act from Dialogue Management Decisions: text planning content selection information structure … Decisions: sentence planning syntactic structure lexical choice alignment prosodic structure … Output Fig. 6.1 NLG decisions in a SDS strategy significantly improved dialogue task completion, with up to a 9.7 % increase compared to a deployed dialogue system which used conventional, hand-coded presentation prompts. Turning to surface realisation, using factored language models, a fully data-driven statistical language generator was developed to produce utterances in a tourist information domain (see Sect. 6.5). Its output was shown not to differ significantly from human paraphrases (over 200 test utterances). Furthermore, active learning from semantically labelled utterances collected through crowd-sourcing was shown to significantly improve performance on sparse training sets. This chapter will explain the motivations behind these approaches, and related work, and will describe several case studies which illustrate their advantages, with reference to recent empirical results in the areas of information presentation and referring expression generation (REG). This presentation will include discussion of user simulations for training NLG components (Sect. 6.4.2), reward functions for training NLG, and evaluation methods and metrics (Sect. 6.3.5). Finally, we provide a critical outlook for future work in this direction, in Sect. 6.7.