Efficient algorithm for Context Sensitive Aggregation in Natural Language generation

Proceedings of Recent Advances in Natural Language Processing, pages 84–89, Hissar, Bulgaria, 12-14 September 2011. Efficient Algorithm for Context Sensitive Aggregation in Natural Language Generation Hemanth Sagar Bayyarapu Language Technologies Research Center International Institute of Information Technology Hyderabad, AP, India - 500032 hemanth.sagar@research.iiit.ac.in Abstract Aggregation is a sub-task of Natural Lan- guage Generation (NLG) that improves the conciseness and readability of the text outputted by NLG systems. Till date, approaches towards the aggregation task have been predominantly manual (man- ual analysis of domain specific corpus and development of rules). In this paper, a new algorithm for aggregation in NLG is proposed, that learns context sensitive ag- gregation rules from a parallel corpus of multi-sentential texts and their underly- ing semantic representations. Addition- ally, the algorithm accepts external con- straints and interacts with the surface re- alizer to generate the best output. Ex- periments show that the proposed con- text sensitive probablistic aggregation al- gorithm performs better than the determin- istic hand crafted aggregation rules. 1 Introduction Aggregation is the process in which two or more linguistic structures are merged to form a single sentence. It helps in generating concise and flu- ent text and hence is an essential component in any NLG system (Reiter and Dale 2000). Fig- ure 1(a) presents an example of de-aggregated text while Figure 1(b) shows its aggregated counter- part. Clearly, the aggregated text is fluent while the de-aggregated text is artificial with lot of re- dundancy. Reiter (1994) proposed a consensus pipeline ar- chitecture for NLG systems with three stages: • Content-Determination: Selects the informa- tion (propositions) to be conveyed and orga- nizes the information in a rhetorically coher- ent manner. • Sentence-Planning: Generates referring ex- pressions, combines multiple propositions, selects appropriate lexical items and syntac- tic structures for each (aggregated) proposi- tion and adds cohesion devices (eg, discourse markers) to make the text flow smoothly. • Surface-Realizer: Converts the lexicalized linguistic structure into a linearized string while ensuring grammaticality, proper punc- tuation, correct morphology. Figure 1: Example showing de-aggregated text and its equivalent aggregated text. The input to the process of aggregation, a sub- module of Sentence-Planning in the consensus ar- chitecture described above, is a set of propositions selected by Content-Determination module which are organized using rhetorical relations between the propositions. Typical NLG systems use a two- stage aggregation process (Wilkinson, 1995). In the first stage, i.e., semantic grouping, the input set of propositions are partitioned into multiple sets, each of which is realized as a sentence. In the sec- ond stage, decisions related to actual realization of each set partition are taken. 84 The essential idea behind semantic grouping is that the propositions that form a set and get real- ized as a meaningful sentence are related some- how. For example in Figure 1, the first two propo- sitions (Bacteria are unicellular. Bacteria are prokaryotic.) are two assertive sentences about Bacteria and hence are aggregated. But it is not true that these two propositions will always be ag- gregated into a single …

Content truncated to 500 words