3.2 Essential Narrative Models
The most recent and popular computational model of narratives is the Story Intention Graph (SIG) , . SIG model captures structure of a narrative in great detail. The model is composed of three interconnected layers – the Text, the Timeline and the Interpretative layer. Narratives are modeled through these basic units. With the patterns of SIG encodings, narratives are categorized in to 80 pre-defined scenario classes like “Gain,” “Promise Broken,” “Unintended Harm,” etc. The scenario classes represent narrative tropes in terms of SIG relations. SIG captures both the timeline of events in the narratives and and the relationship between the actors or entities in the narrative. Analogical similarity is deduced with combinations of layers in the encoding. Three are three suggested analogy detection approaches. The first is Propositional Similarity, where just the timeline is used. The second is Static Analogy, a top-down approach, where analogy is inferred by mapping encoding with one of pre-defined scenario patterns. The third is Dynamic Analogy, a bottom-up approach, where analogy is inferred by largest isomorphic subgraph. Scheherazade1 is a freely available application program, to encode stories in the form of SIGs. SIG is a good formal model to represent narratives and it has to constructed manually from the text using the tool.
In Halpin et al. , a model is proposed to analyze recall of events and the sequencing in rewritten stories by children. The model was devised to develop a system called “Story Station” to provide guidance in writing stories to children aged between 10-12 years. The computational model compares stories by finding equivalence of lemmatized tokens of events, using WordNet and events order. The merit of rewritten stories is decided by the semantic nearness of recalled events and reproduction of event sequences. Rewritten stories were rated as Excellent, Good, Fair and Poor. The story was rated excellent, when the point of the story and all important sequences are reproduced. Good rating was assigned for stories when main events and links were reproduced. Fair rating was assigned for stories when a major chunk of the story is missing and poor when substantial amount of the plot is missing. A corpus of 103 stories rewritten by children were used to evaluate the model.
In Miller et al. , relationship between entities is computed by finding the similar events in which they appear. Finding the similarity between entities is termed as alignment. Events are extracted and event sets are created manually. Similarity between events is found with the abstract hypernym of tokens appearing in events text and entities appearing in similar events are tagged to be related. An adjacency matrix of entity relationships is created. The adjacency matrix results in a graph of entities as nodes and edges as hypernym relation. From these graphs, using network analysis approaches, similarity between all entities of narrative document is computed.
Reiter et al.  presents an unsupervised and automated narrative structure discovery approach. It was evaluated on ritual description and folktale narratives corpus. A narrative is represented as a graph with entities as nodes and edges as relationship between them. The story is first represented as a sequence of events, and entities appearing in a single event are tagged to be related. Model is constructed using FrameNet for event re-construction, co-reference resolution to relate entities across narrative document and WordNet sense to compare events at a higher abstract level. Three alignment algorithms are suggested: Sequence alignment to align chains of sequences from narrative documents, Graph based predicate alignment to align by similar predicate-argument structures of events from narrative documents and Bayesian model merging to align using Hidden Markov Models (HMM) in event sequences from narrative documents.
1Scheherazade SIG tool: http://www.cs.columbia.edu/~delson/software.shtml