Automatic Extraction of Briefing Templates

Automatic Extraction of Briefing Templates Dipanjan Das Mohit Kumar Language Technologies Institute Carnegie Mellon University 5000 Forbes Avenue Pittsburgh, PA 15213, USA {dipanjan, mohitkum, air}@cs.cmu.edu Alexander I. Rudnicky Abstract An approach to solving the problem of au- tomatic briefing generation from non-textual events can be segmenting the task into two major steps, namely, extraction of briefing templates and learning aggregators that col- late information from events and automati- cally fill up the templates. In this paper, we describe two novel unsupervised approaches for extracting briefing templates from hu- man written reports. Since the problem is non-standard, we define our own criteria for evaluating the approaches and demonstrate that both approaches are effective in extract- ing domain relevant templates with promis- ing accuracies. 1 Introduction Automated briefing generation from non-textual events is an unsolved problem that currently lacks a standard approach in the NLP community. Broadly, it intersects the problem of language generation from structured data and summarization. The prob- lem is relevant in several domains where the user has to repeatedly write reports based on events in the domain, for example, weather reports (Reiter et al., 2005), medical reports (Elhadad et al., 2005), weekly class project reports (Kumar et al., 2007) and so forth. On observing the data from these domains, we notice a templatized nature of report items. Ex- amples (1)-(3) demonstrate equivalents in a particu- lar domain (Reiter et al., 2005). (1) [A warm front] from [Iceland] to [northern Scotland] will move [SE] across [the northern North Sea] [today and tomorrow] (2) [A warm front] from [Iceland] to [the Faeroes] will move [ENE] across [the Norwegian Sea] [this evening] (3) [A ridge] from [the British Isles] to [Iceland] will move [NE] across [the North Sea] [today] In each sentence, the phrases in square brackets at the same relative positions form the slots that take up different values at different occasions. The cor- responding template is shown in (4) with slots con- taining their respective domain entity types. Instan- tiations of (4) may produce (1)-(3) and similar sen- tences. This kind of sentence structure motivates an approach of segmenting the problem of closed do- main summarization into two major steps of auto- matic template extraction and learning aggregators, which are pattern detectors that assimilate informa- tion from the events, to populate these templates. (4) [PRESSURE ENTITY] from [LOCATION] to [LOCATION] will move [DIRECTION] across [LOCATION] [TIME] In the current work we address the first problem of automatically extracting domain templates from hu- man written reports. We take a two-step approach to the problem; first, we cluster report sentences based on similarity and second, we extract template(s) cor- responding to each cluster by aligning the instances in the cluster. We experimented with two indepen- dent, arguably complementary techniques for clus- tering and aligning – a predicate argument based ap- proach that extracts more general templates contain- ing one predicate and a ROUGE (Lin, 2004) based 265 approach that can extract templates containing mul- tiple verbs. As we will see below, …

Content truncated to 500 words