Heuristic Based Extraction of Causal Relations from Annotated Causal Cue Phrases

Heuristic Based Extraction of Causal Relations from Annotated Causal Cue Phrases By Matthew J. Hausknecht This work focuses on the detection and extraction of Causal Relations from open domain text starting with annotated Causal Cue Phrases (CCPs). It is argued that the problem of causality extraction should be decomposed into two distinct subtasks. First, it is necessary to identify Causal Cue Phrases (CCPs) inside of a body of text. Second, using these CCPs, the cause and effect phrases of each causal relation must be extracted. To prove that CCPs are an essential part of causality extraction, it is experimentally demonstrated that the accuracy of cause and effect phrase extraction dramatically increases when CCP knowledge is utilized. A 31% increase in accuracy of cause and effect phrase extraction of two equivalent CRF machine learning algorithms is found when simple, word-based knowledge of CCPs is taken into account. Furthermore, it is shown that cause and effect phrase extraction can be performed accurately and robustly without the aid of complex machine learning techniques. A simple, heuristic based extraction algorithm, centering around three distinct classes of CCPs, is introduced. This algorithm achieves an accuracy of 87% on the task of extracting cause and effect phrases. While the problem of identifying CCPs in open domain text is not addressed, it is hypothesized that this task is far easier than identifying cause and effect phrases alone because the space of all possible CCPs is far smaller than that of all causal relations. Finally, this work contributes a free, publicly accessible corpus explicitly annotated with both intra-sentential causal relations and corresponding Causal Cue Phrases. It is our hope that this resource may see future use as a standard corpus for the task of causality extraction. Heuristic Based Extraction of Causal Relations from Annotated Causal Cue Phrases