We present a novel method for discovering causal relations between events encoded in text. In order to determine if two events from the same sentence are in a causal relation or not, we first build a graph representation of the sentence that encodes lexical, syntactic, and semantic information. From such graph representations we automatically extract multiple graph patterns (or subgraphs). The patterns are sorted according to their contribution to the expression of intra-sentential causality between events. To decide whether a pair of events is in a causal relation, we employ a binary classifier that uses the graph patterns. Our experimental results indicate that capturing causal event relations using graph patterns outperforms existing methods.
TopIntroduction
Automatic discovery of causal relations between textual events is a central task for various applications in Natural Language Processing (NLP). Specifically, tasks that require some form of reasoning such as probabilistic reasoning (Narayanan, 1997), common sense reasoning (Mueller, 2007), and question answering (Girju, 2003). An example of a causal relation between two events is presented in the following sentence:
This example encodes a special type of causal relation called enablement, in which the event wound enables the event let loose to happen. In this case, the event let loose will take place only if the event wound occurs. Moreover, causal relations are closely related with temporal relations that hold between two events. For example, the enablement relation from sentence S1 corresponds to a before temporal relation established between the wound event and the let loose event.
We cast causal relation detection as a classification task. Most existing approaches for discovering causal relations use machine learning classifiers based on unstructured linguistic features. In this chapter, we describe a method that captures the contextual information of the two events. Contextual information is captured by lexical, syntactic, and semantic features as well as dependencies that exist between these features. Features are structured into graphs, enabling us to discover graph patterns that account for causality between pairs of intra-sentential events.
The remainder of the chapter is organized as follows. We first describe the related work. Next we detail the graph representation that captures the structural dependencies of lexical, syntactic, and semantic features characterizing pairs of events that belong to a causal relation. In that section we detail each type of feature as well. The following section discusses the method of discovering syntactic and lexico-semantic graph patterns that characterize causality. Patterns are used for training a binary classifier which decides whether there exists a causal relation between pairs of intra-sentential events. In a separate section we discuss the experimental results and the error analysis. We also discuss the future research directions before summarizing the conclusions of the chapter.
TopCausality can be expressed in texts in many ways. Moreover, linguists and researchers have established that there are several forms of causality, including purpose, reason, and enablement (Bejan & Harabagiu, 2008a). Traditionally, causality in text has been viewed as a discourse phenomenon or as a semantic-specific function. Methods that have focused on the discourse-level information attempted to recognized causal relations between discourse units in the same way as they would recognize other discourse relations (e.g. elaboration, parallel, and violated-expectations). For example, Marcu & Echihabi (2002) have used a naïve Bayes classifier trained on pairs of words from discourse segments separated by cue phrases such as thus and because. Their approach achieved very high scores when evaluated on cause-explanation-evidence relations.