We describe a method for identifying the speakers of quoted speech in natural-language textual stories. We have assembled a corpus of more than 3,000 quotations, whose speakers (if any) are manually identified, from a collection of 19th and 20th century literature by six authors. Using rule-based and statistical learning, our method identifies candidate characters, determines their genders, and attributes each quote to the most likely speaker. We divide the quotes into syntactic classes in order to leverage common discourse patterns, which enable rapid attribution for many quotes. We apply learning algorithms to the remainder and achieve an overall accuracy of 83%.

This thesis describes new approaches to the formal modeling of narrative discourse. Although narratives of all kinds are ubiquitous in daily life, contemporary text processing techniques typically do not leverage the aspects that separate narrative from expository discourse. We describe two approaches to the problem. The first approach considers the conversational networks to be found in literary fiction as a key aspect of discourse coherence; by isolating and analyzing these networks, we are able to comment on longstanding literary theories. The second approach proposes a new set of discourse relations that are specific to narrative. By focusing on certain key aspects, such as agentive characters, goals, plans, beliefs, and time, these relations represent a theory-of-mind interpretation of a text. We show that these discourse relations are expressive, formal, robust, and through the use of a software system, amenable to corpus collection projects through the use of trained annotators. We have procured and released a collection of over 100 encodings, covering a set of fables as well as longer texts including literary fiction and epic poetry. We are able to inferentially find similarities and analogies between encoded stories based on the proposed relations, and an evaluation of this technique shows that human raters prefer such

Quote extraction and attribution is the task of automatically extracting quotes from text and attributing each quote to its correct speaker. The present state-of-the-art system uses gold standard information from previous decisions in its features, which, when removed, results in a large drop in performance. We treat the problem as a sequence labelling task, which allows us to incorporate sequence features without using gold standard information. We present results on two new corpora and an augmented version of a third, achieving a new state-of-the-art for systems using only realistic features. 1