Discourse structures have a central role in several computational tasks, such as question-answering or dialogue generation. In particular, the framework of the Rhetorical Structure Theory (RST) offers a sound formalism for hierarchical text organization. In this article, we present HILDA, an implemented discourse parser based on RST and Support Vector Machine (SVM) classification. SVM classifiers are trained and applied to discourse segmentation and relation labeling. By combining labeling with a greedy bottom-up tree building approach, we are able to create accurate discourse trees in linear time complexity. Importantly, our parser can parse entire texts, whereas the publicly available parser SPADE (Soricut and Marcu 2003) is limited to sentence level analysis. HILDA outperforms other discourse parsers for tree structure construction and discourse relation labeling. For the discourse parsing task, our system reaches 78.3% of the performance level of human annotators. Compared to a state-of-the-art rule-based discourse parser, our system achieves a performance increase of 11.6%.

Abstract. The Text2Dialogue (T2D) system that we are developing allows digital content creators to generate attractive multi-modal dialogues presented by two virtual agents-by simply providing textual information as input. We use Rhetorical Structure Theory (RST) to decompose text into segments and to identify rhetorical discourse relations between them. These are then "acted out" by two 3D agents using synthetic speech and appropriate conversational gestures. In this paper, we present version 1.0 of the T2D system and focus on the novel technique that it uses for mapping rhetorical relations to question-answer pairs, thus transforming (monological) text into a form that supports dialogues between virtual agents.

Abstract. Identifying discourse relations in a text is essential for various tasks in Natural Language Processing, such as automatic text summarization, question-answering, and dialogue generation. The first step of this process is segmenting a text into elementary units. In this paper, we present a novel model of discourse segmentation based on sequential data labeling. Namely, we use Conditional Random Fields to train a discourse segmenter on the RST Discourse Treebank, using a set of lexical and syntactic features. Our system is compared to other statistical and rule-based segmenters, including one based on Support Vector Machines. Experimental results indicate that our sequential model outperforms current state-of-the-art discourse segmenters, with an F-score of 0.94. This performance level is close to the human agreement F-score of 0.98.

Abstract. The corpora available for training discourse relation classifiers are annotated using a general set of discourse relations. However, for certain applications, custom discourse relations are required. Creating a new annotated corpus with a new relation taxonomy is a timeconsuming and costly process. We address this problem by proposing a semi-supervised approach to discourse relation classification based on Structural Learning. First, we solve a set of auxiliary classification problems using unlabeled data. Second, the learned classifiers are used to extend feature vectors to train a discourse relation classifier. By defining a relevant set of auxiliary classification problems, we show that the proposed method brings improvement of at least 50% in accuracy and F-score on the RST Discourse Treebank and Penn Discourse Treebank, when small training sets of ca. 1000 training instances are employed. This is an attractive perspective for training discourse relation classifiers on domains where little amount of labeled training data is available.

Abstract. This paper describes recent advances on the Text2Dialogue system we are currently developing. Our system enables automatic transformation of monological text into a dialogue. The dialogue is then "acted out" by virtual agents, using synthetic speech and gestures. In this paper, we focus on the monologue-to-dialogue transformation, and describe how it uses textual coherence relations to map text segments to query-answer pairs between an expert and a layman agent. By creating mapping rules for a few well-selected relations, we can produce coherent dialogues with proper assignment of turns for the speakers in a majority of cases.