Detecting and Explaining Causes From Text For a Time Series Event

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2758–2767 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Detecting and Explaining Causes From Text For a Time Series Event Dongyeop Kang, Varun Gangal, Ang Lu, Zheng Chen, Eduard Hovy Language Technology Institute Carnegie Mellon University {dongyeok,vgangal,alu1,zhengc1,hovy}@cs.cmu.edu Abstract Explaining underlying causes or effects about events is a challenging but valu- able task. We define a novel problem of generating explanations of a time series event by (1) searching cause and effect relationships of the time series with tex- tual data and (2) constructing a connect- ing chain between them to generate an ex- planation. To detect causal features from text, we propose a novel method based on the Granger causality of time series be- tween features extracted from text such as N-grams, topics, sentiments, and their composition. The generation of the se- quence of causal entities requires a com- monsense causative knowledge base with efficient reasoning. To ensure good in- terpretability and appropriate lexical usage we combine symbolic and neural repre- sentations, using a neural reasoning algo- rithm trained on commonsense causal tu- ples to predict the next cause step. Our quantitative and human analysis show em- pirical evidence that our method success- fully extracts meaningful causality rela- tionships between time series with textual features and generates appropriate expla- nation between them. 1 Introduction Producing true causal explanations requires deep understanding of the domain. This is beyond the capabilities of modern AI. However, it is possible to collect large amounts of causally related events, and, given powerful enough representational vari- ability, to construct cause-effect chains by select- ing individual pairs appropriately and linking them together. Our hypothesis is that chains composed Figure 1: Example of causal features for Face- book’s stock change in 2013. The causal features (e.g., martino, k-rod) rise before the Facebook’s rapid stock rise in August. of locally coherent pairs can suggest overall cau- sation. In this paper, we view causality as (common- sense) cause-effect expressions that occur fre- quently in online text such as news articles or tweets. For example, “greenhouse gases causes global warming” is a sentence that provides an ‘atomic’ link that can be used in a larger chain. By connecting such causal facts in a sequence, the result can be regarded as a causal explanation be- tween the two ends of the sequence (see Table 1 for examples). This paper makes the following contributions: • we define the problem of causal explanation generation, • we detect causal features of a time series event (CSPIKES) using Granger (Granger, 1988) method with features extracted from text such as N-grams, topics, sentiments, and their com- position, • we produce a large graph called CGRAPH of lo- cal cause-effect units derived from text and de- velop a method to produce causal explanations by selecting and linking appropriate units, using neural representations to enable unit matching and chaining. 2758 Table 1: Examples of generated causal expla- nation between some temporal causes and target …

Content truncated to 500 words