Natural Language Generation with Tree Conditional Random Fields

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 400–409, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Natural Language Generation with Tree Conditional Random Fields Wei Lu1, Hwee Tou Ng1,2, Wee Sun Lee1,2 1Singapore-MIT Alliance 2Department of Computer Science National University of Singapore luwei@nus.edu.sg {nght,leews}@comp.nus.edu.sg Abstract This paper presents an effective method for generating natural language sentences from their underlying meaning represen- tations. The method is built on top of a hybrid tree representation that jointly encodes both the meaning representation as well as the natural language in a tree structure. By using a tree conditional random field on top of the hybrid tree representation, we are able to explicitly model phrase-level dependencies amongst neighboring natural language phrases and meaning representation components in a simple and natural way. We show that the additional dependencies captured by the tree conditional random field allows it to perform better than directly inverting a previously developed hybrid tree semantic parser. Furthermore, we demonstrate that the model performs better than a previ- ous state-of-the-art natural language gen- eration model. Experiments are performed on two benchmark corpora with standard automatic evaluation metrics. 1 Introduction One of the ultimate goals in the field of natural lan- guage processing (NLP) is to enable computers to converse with humans through human languages. To achieve this goal, two important issues need to be studied. First, it is important for comput- ers to capture the meaning of a natural language sentence in a meaning representation. Second, computers should be able to produce a human- understandable natural language sentence from its meaning representation. These two tasks are re- ferred to as semantic parsing and natural language generation (NLG), respectively. In this paper, we use corpus-based statistical methods for constructing a natural language gener- ation system. Given a set of pairs, where each pair consists of a natural language (NL) sentence and its formal meaning representation (MR), a learn- ing method induces an algorithm that can be used for performing language generation from other previously unseen meaning representations. A crucial question in any natural language pro- cessing system is the representation used. Mean- ing representations can be in the form of a tree structure. In Lu et al. (2008), we introduced a hybrid tree framework together with a probabilis- tic generative model to tackle semantic parsing, where tree structured meaning representations are used. The hybrid tree gives a natural joint tree rep- resentation of a natural language sentence and its meaning representation. A joint generative model for natural language and its meaning representation, such as that used in Lu et al. (2008) has several advantages over var- ious previous approaches designed for semantic parsing. First, unlike most previous approaches, the generative approach models a simultaneous generation process for both NL and MR. One el- egant property of such a joint generative model is that it allows the modeling of both semantic parsing and natural language generation within the same process. Second, the generative process pro- ceeds as a …

Content truncated to 500 words