A Cascaded Machine Learning Approach to Interpreting Temporal Expressions

Proceedings of NAACL HLT 2007, pages 420–427, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics A Cascaded Machine Learning Approach to Interpreting Temporal Expressions David Ahn Joris van Rantwijk Maarten de Rijke ISLA, University of Amsterdam Kruislaan 403, 1098 SJ Amsterdam, The Netherlands {ahn, rantwijk, mdr}@science.uva.nl Abstract A new architecture for identifying and in- terpreting temporal expressions is intro- duced, in which the large set of com- plex hand-crafted rules standard in sys- tems for this task is replaced by a series of machine learned classifiers and a much smaller set of context-independent seman- tic composition rules. Experiments with the TERN 2004 data set demonstrate that overall system performance is comparable to the state-of-the-art, and that normaliza- tion performance is particularly good. 1 Introduction In order to fully understand a piece of text, we must understand its temporal structure. The first step toward such an understanding is identifying ex- plicit references to time. We focus on the task of automatically annotating temporal expressions (or timexes)—both identifying them in text and inter- preting them to determine what times they refer to. Timex annotation is more than normalizing date ex- pressions. First, time consists of more than calen- dar dates and clock times—it also includes points of finer and coarser granularity, durations, and sets of times. Second, the expressions that refer to time are not just full date and time expressions—they may be underspecified, ambiguous, and anaphoric. Building a system for the full timex identifica- tion and interpretation task can be tedious, requiring a great deal of manual effort. The 2004 Temporal Expression Recognition and Normalization (TERN) evaluation1 evaluated systems on two tasks: timex 1http://timex2.mitre.org/tern.html recognition (identification) alone and recognition and normalization (interpretation) together. All the full-task systems were rule-based systems; the top performing full-task system uses in excess of one thousand hand-crafted rules, which probe words and their contexts in order to both identify timexes and to assemble information necessary to interpret them (Negri and Marseglia, 2004). By contrast, machine learned systems dominated the recognition-only task and even achieved slightly better recognition scores than their rule-based counterparts. We seek to demonstrate that a timex annotation system that performs both recognition and normal- ization need not be a tangle of rules that serve dou- ble duty for identification and interpretation and that mix up context-dependent and context-independent processing. We propose a novel architecture that clearly separates syntactic, semantic, and prag- matic processing and factors out context-dependent from context-independent processing. Factoring out context-dependent disambiguation into separate classification tasks introduces the opportunity for using machine learning, which supports our main goal: building a portable, trainable timex annota- tion system in which the role of hand-crafted rules is minimized. The system we present here (avail- able from http://ilps.science.uva.nl/ Resources/timextag/) achieves the goal of making use of only a small set of hand-crafted, context-independent rules to achieve state-of-the-art normalization performance. In the following section, we define what a timex is. We give an overview of our system architecture in §3 and describe the components in …

Content truncated to 500 words