Probabilistic Task Content Modeling for Episodic Textual Narratives

Episodic knowledge is often stored in the form of textual narratives written in natural language. However, a large repository of such narratives will contain both repetitive and novel knowledge. In this paper, we propose an approach for discovering interesting pieces of knowledge by using a priori task knowledge. By considering the narratives as generated by an underlying task structure, the elements of the task can be regarded as topics that generate the text. Then, by capturing task content in a probabilistic model, the model can be used, e.g., to identify the semantic orientation of textual phrases. An evaluation for a real world corpus of episodic narratives provides strong evidence for the feasibility of the proposed approach. Episodic Textual Narratives Discussing the use of knowledge in knowledge systems, (Richter 1998) distinguishes among three types of knowledge: background, contextual, and episodic knowledge. Out of all types of knowledge systems, episodic knowledge— which is of narrative character, because it tells the story of something that happened in the past—is directly employed in Case-Based Reasoning (CBR) systems only. Since episodic knowledge has a narrative character, a natural means of preserving it is in the form of textual narratives written in natural language by human users. Extraction of valuable pieces of knowledge from such narratives, which can serve as cases in the context of a Textual CBR (TCBR) system, is the focus of our research. A common way of extracting knowledge from text documents in the context of a TCBR system is by considering a priori domain knowledge (Lenz 1999). The underlying idea is that text can be regarded as a container of domain objects (or information entities) and by using different types of knowledge acquisition, text documents can be reduced to a set of such information entities. In contrast, we take an alternative perspective. We consider text as generated by an underlying process, which consists of a series of related events. Domain objects are then participants of such events. By recognizing events and their participants in text, it is not only possible to discover domain objects, but also their respective roles in the event. In this way, text will not be a Copyright c © 2007, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. mere set of information entities, but a network of interconnected, semantically labeled entities. Additionally, a second difference to (Lenz 1999)’s approach is that instead of a priori domain knowledge, we use a priori task knowledge to perform knowledge extraction. Task Knowledge: An Example Consider the task MONITOR-and-DIAGNOSE, a task that starts with the monitoring of an object and continues with diagnosis only if something problematic is observed during the monitoring step. In general, there is a distinct temporal order in the way the MONITOR-and-DIAGNOSE task is performed. 1. Some entities of interest are observed and the respective findings are noticed. 2. These findings are then explained and evaluated. 3. If findings are evaluated as negative, actions for maintenance are recommended. These three steps can be referred to as observation (OBS), explanation (EXP), and take action (ACT), and can be regarded as events that occur during the execution of the MONITOR-and-DIAGNOSE task. In the same way in which these events occur in reality, they will be described in written form, too. Each of the events is constituted by relations between different elements of the task. For example, an OBS relates an observed object to a finding, or an EXP relates a symptom to a possible cause. In our previous research, we have described an active learning approach (named LARC) that learns to annotate episodic narratives with task knowledge roles such as observed object, finding, cause, etc. (Mustafaraj, Hoof, & Freisleben 2006). Probabilistic Task Content Modeling The previous description of a task as a series of interconnected events and participant roles constitutes an abstract model of task structure. An instantiation of the task structure in a real situation produces the task content. The instantiation consists of the verbalization of the abstract events and roles with concrete sentences and phrases. Because realworld events and natural language are of stochastic nature, a probabilistic model is an appropriate means for capturing task content.