Narratives are central to communication and the human experience. For a computer system to understand a narrative, it must be able to identify the key facts or plot elements that describe what happened or how the world has changed. These element are called events;establishing a document’s events and the relationships between them is central to under-standing a text’s narrative. Events are related to each other temporally and causally by being part of the same story arc. Further, these event sequences typically follow patterns called scripts.
In this thesis, I explore three essential stages for narrative understanding. All three stages form an end-to-end system that starts with plain text documents and ultimately produces scripts, a generalization of narrative structure. The first two stages, event detection and event sequence extraction, analyze a document and extract the key information needed to understand a document’s narrative. The final stage, script learning, generalizes the discovered event sequences to find common patterns between them.
First, I propose a neural network model based on grammar and syntax. It combines a left-to-right reading of the text along with a reading ordered by the sentence’s syntactic tree. The model is an extension of Gated Recurrent Units and uses an attention mechanism to blend both reading modes. This model achieves state-of-the-art performance on a well-studied task.
I present an evaluation that is the first to quantify the substantial variability of neural networks when applied to the nuanced problem of event detection. Two sources of variability are considered: the effect of local optimization of the neural networks’ training procedure and the types of documents used for evaluation and training. I show that the variation involved is often greater than the differences in the state-of-the-art, demonstrating the need for a robust evaluation.
Second, the new task of event sequence extraction is addressed with a novel, interpretable neural network framework. The framework represents the problem as a series of graph trans-formations. By doing so, it allows for various neural network architectures to be combined while mirroring the structure of the task. Several models instantiated from the framework are evaluated against a strong baseline showing a substantial improvement on a difficult task.Further, I demonstrate the framework’s flexibility by evaluating it on the entity relation extraction task.
Finally, I examine using Hidden Markov Models to learn scripts from event sequences with missing data. This formulation of scripts as Hidden Markov Models is novel and the first to explicitly account for missing observations in the context of natural language processing. The models are learned with a bottom-up induction algorithm based on Structural Expectation Maximization. The scripts are evaluated by inferring omitted events in event sequences and are shown to be more effective than an informed baseline.