Conditional Random Fields for Responsive Surface Realisation using Global Features

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1254–1263, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Conditional Random Fields for Responsive Surface Realisation using Global Features Nina Dethlefs, Helen Hastie, Heriberto Cuay´ahuitl and Oliver Lemon Mathematical and Computer Sciences Heriot-Watt University, Edinburgh n.s.dethlefs | h.hastie | h.cuayahuitl | o.lemon@hw.ac.uk Abstract Surface realisers in spoken dialogue sys- tems need to be more responsive than con- ventional surface realisers. They need to be sensitive to the utterance context as well as robust to partial or changing generator inputs. We formulate surface realisation as a sequence labelling task and combine the use of conditional random fields (CRFs) with semantic trees. Due to their extended notion of context, CRFs are able to take the global utterance context into account and are less constrained by local features than other realisers. This leads to more natural and less repetitive surface realisa- tion. It also allows generation from partial and modified inputs and is therefore ap- plicable to incremental surface realisation. Results from a human rating study confirm that users are sensitive to this extended no- tion of context and assign ratings that are significantly higher (up to 14%) than those for taking only local context into account. 1 Introduction Surface realisation typically aims to produce out- put that is grammatically well-formed, natural and cohesive. Cohesion can be characterised by lexical or syntactic cues such as repetitions, substitutions, ellipses, or connectives. In automatic language generation, such properties can sometimes be dif- ficult to model, because they require rich context- awareness that keeps track of all (or much) of what was generated before, i.e. a growing generation history. In text generation, cohesion can span over the entire text. In interactive settings such as gen- eration within a spoken dialogue system (SDS), a challenge is often to keep track of cohesion over several utterances. In addition, since interactions are dynamic, generator inputs from the dialogue manager can sometimes be partial or subject to subsequent modification. This has been addressed by work on incremental processing (Schlangen and Skantze, 2009). Since dialogue acts are passed on to the generation module as soon as possible, this can sometimes lead to incomplete generator inputs (because the user is still speaking), or in- puts that are subject to later modification (because of an initial ASR mis-recognition). In this paper, we propose to formulate surface realisation as a sequence labelling task. We use conditional random fields (Lafferty et al., 2001; Sutton and McCallum, 2006), which are suitable for modelling rich contexts, in combination with semantic trees for rich linguistic information. This combination is able to keep track of dependen- cies between syntactic, semantic and lexical fea- tures across multiple utterances. Our model can be trained from minimally labelled data, which re- duces development time and may (in the future) facilitate an application to new domains. The domain used in this paper is a pedestrian walking around a city looking for information and recommendations for local restaurants from …

Content truncated to 500 words