Causal networks as the backbone for temporal data-to-text

Causal networks have been successfully used to explain the process of narrative comprehension in humans (Trabasso et al., 1989). This motivated their use in the automatic creation of fairy tales (Swartjes and Theune, 2006; Theune et al., 2007). Some have suggested that causal relations also play an important role in improving narrative generation from real-life temporal data (Hunter et al., 2012; Gervás, 2014). Several narrative data-to-text systems already identify and make use of some causal relations (Hallett, 2008; Hunter et al., 2012; Bouayad-Agha et al., 2012; Wanner et al., 2010). Is it possible go one step further in the identification of causal relations and aim at extracting a causal network that could be used to improve the coherence of generated texts? Given that real-life temporal data could be analysed and interpreted automatically so as to produce an appropriate causal network, how could this information be best used to organize a coherent narrative adapted to the communicative needs of a wide range of data-to-text applications? We will here address the second question by proposing a bottom-up document planning method for building the rhetorical structure of coherent narratives from a causal network. The causal network represents the plot of the story to be told by the narrative text. The rhetorical structure represents one way, corresponding to a particular viewpoint, of organizing the text to tell that story. The goal is therefore to translate the semantic information contained in a causal network into a rhetorical structure using parameters that characterize this perspective. The proposed approach uses hierarchical clustering and parameterization by adjacency and ordering preferences. It is robust to partial causal networks and does not require the generated rhetorical structure to be a tree. The document planning stage would be preceded by signal analysis and data interpretation and followed by microplanning and surface realisation in a complete data-to-text pipeline (Reiter, 2007).