Understanding stories via event sequence modeling

event sequence modeling;event extraction;entity co-reference;event co-reference;semantic language model;discourse understanding

Peng, Haoruo

关键词: event sequence modeling;    event extraction;    entity co-reference;    event co-reference;    semantic language model;    discourse understanding;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/101526/PENG-DISSERTATION-2018.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Understanding stories, i.e. sequences of events, is a crucial yet challenging natural language understanding (NLU) problem, which requires dealing with multiple aspects of semantics, including actions, entities and emotions, as well as background knowledge. In this thesis, towards the goal of building a NLU system that can model what has happened in stories and predict what would happen in the future, we contribute on three fronts: First, we investigate the optimal way to model events in text; Second, we study how we can model a sequence of events with the balance of generality and specificity; Third, we improve event sequence modeling by joint modeling of semantic information and incorporating background knowledge.Each of the above three research problems poses both conceptual and computational challenges. For event extraction, wefind that Semantic Role Labeling (SRL) signals can be served as good intermediate representations for events, thus giving us the ability to reliably identify events with minimal supervision. In addition, since it is important to resolve co-referred entities for extracted events, we make improvements to an existing co-reference resolution system. To model event sequences, we start from studying within document event co-reference (the simplest flow of events); and then extend to model two other more natural event sequences along with discourse phenomena while abstracting over the specific mentions of predicates and entities. We further identify problems for the basic event sequence models, where we fail to capture multiple semantic aspects and background knowledge. We then improve our system by jointly modeling frames, entities and sentiments, yielding joint representations of all these semantic aspects; while at the same time incorporate explicit background knowledge acquired from other corpus as well as human experience. For all tasks, we evaluate the developed algorithms and models on benchmark datasets and achieve better performance compared to other highly competitive methods.