Incremental Text Structuring with Online Hierarchical Ranking

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 83–91, Prague, June 2007. c⃝2007 Association for Computational Linguistics Incremental Text Structuring with Online Hierarchical Ranking Erdong Chen, Benjamin Snyder and Regina Barzilay Computer Science and Artificial Intelligence Laboratory Massachusetts Institute of Technology {edc,bsnyder,regina}@csail.mit.edu Abstract Many emerging applications require doc- uments to be repeatedly updated. Such documents include newsfeeds, webpages, and shared community resources such as Wikipedia. In this paper we address the task of inserting new information into exist- ing texts. In particular, we wish to deter- mine the best location in a text for a given piece of new information. For this process to succeed, the insertion algorithm should be informed by the existing document struc- ture. Lengthy real-world texts are often hier- archically organized into chapters, sections, and paragraphs. We present an online rank- ing model which exploits this hierarchical structure – representationally in its features and algorithmically in its learning proce- dure. When tested on a corpus of Wikipedia articles, our hierarchically informed model predicts the correct insertion paragraph more accurately than baseline methods. 1 Introduction Many emerging applications require documents to be repeatedly updated. For instance, newsfeed ar- ticles are continuously revised by editors as new in- formation emerges, and personal webpages are mod- ified as the status of the individual changes. This re- vision strategy has become even more prevalent with the advent of community edited web resources, the most notable example being Wikipedia. At present this process involves massive human effort. For in- stance, the English language version of Wikipedia averaged over 3 million edits1 per month in 2006. Even so, many articles quickly become outdated. A system that performs such updates automatically could drastically decrease maintenance efforts and potentially improve document quality. Currently there is no effective way to automati- cally update documents as new information becomes available. The closest relevant text structuring tech- nique is the work on sentence ordering, in which a complete reordering of the text is undertaken. Pre- dictably these methods are suboptimal for this new task because they cannot take advantage of existing text structure. We introduce an alternative vision of text struc- turing as a process unfolding over time. Instead of ordering sentences all at once, we start with a well- formed draft and add new information at each stage, while preserving document coherence. The basic operation of incremental text structuring is the inser- tion of new information. To automate this process, we develop a method for determining the best loca- tion in a text for a given piece of new information. The main challenge is to maintain the continuity and coherence of the original text. These proper- ties may be maintained by examining sentences ad- jacent to each potential insertion point. However, a local sentence comparison method such as this may fail to account for global document coherence (e.g. by allowing the mention of some fact in an inappro- priate section). This problem is especially acute …

Content truncated to 500 words