Being able to identify which rhetorical relations (e.g., contrast or explanation) hold between spans of text is important for many natural language processing applications. Using machine learning to obtain a classifier which can distinguish between different relations typically depends on the availability of manually labelled training data, which is very time-consuming to create. However, rhetorical relations are sometimes lexically marked, i.e., signalled by discourse markers (e.g., because, but, consequently etc.), and it has been suggested (Marcu and Echihabi, 2002) that the presence of these cues in some examples can be exploited to label them automatically with the corresponding relation. The discourse markers are then removed and the automatically labelled data are used to train a classifier to determine relations even when no discourse marker is present (based on other linguistic cues such as word co-occurrences). In this paper, we investigate empirically how feasible this approach is. In particular, we test whether automatically labelled, lexically marked examples are really suitable training material for classifiers that are then applied to unmarked examples. Our results suggest that training on this type of data may not be such a good strategy, as models trained in this way do not seem to generalise very well to unmarked data. Furthermore, we found some evidence that this behaviour is largely independent of the classifiers used and seems to lie in the data itself (e.g., marked and unmarked examples may be too dissimilar linguistically and removing unambiguous markers in the automatic labelling process may lead to a meaning shift in the examples). 1

In this paper we propose a data intensive approach for inferring sentence-internal temporal relations. Temporal inference is relevant for practical NLP applications which either extract or synthesize temporal information (e.g., summarisation, question answering). Our method bypasses the need for manual coding by exploiting the presence of markers like after, which overtly signal a temporal relation. We first show that models trained on main and subordinate clauses connected with a temporal marker achieve good performance on a pseudo-disambiguation task simulating temporal inference (during testing the temporal marker is treated as unseen and the models must select the right marker from a set of possible candidates). Secondly, we assess whether the proposed approach holds promise for the semi-automatic creation of temporal annotations. Specifically, we use a model trained on noisy and approximate data (i.e., main and subordinate clauses) to predict intra-sentential relations present in TimeBank, a corpus annotated rich temporal information. Our experiments compare and contrast several probabilistic models differing in their feature space, linguistic assumptions and data requirements. We evaluate performance against gold standard corpora and also against human subjects. 1.

In this paper we consider the problem of identifying and classifying discourse coherence relations. We report initial results over the recently released Discourse GraphBank (Wolf and Gibson, 2005). Our approach considers, and determines the contributions of, a variety of syntactic and lexico-semantic features. We achieve 81% accuracy on the task of discourse relation type classification and 70 % accuracy on relation identification. 1

This paper introduces a new algorithm to parse discourse within the framework of Rhetorical Structure Theory (RST). Our method is based on recent advances in the field of statistical machine learning (multivariate capabilities of Support Vector Machines) and a rich feature space. RST offers a formal framework for hierarchical text organization with strong applications in discourse analysis and text generation. We demonstrate automated annotation of a text with RST hierarchically organised relations, with results comparable to those achieved by specially trained human annotators. Using a rich set of shallow lexical, syntactic and structural features from the input text, our parser achieves, in linear time, 73.9 % of professional annotators’ human agreement F-score. The parser is 5 % to 12 % more accurate than current state-of-the-art parsers. 1

We report results of experiments which build and refine models of rhetoricalsemantic relations such as Cause and Contrast. We adopt the approach of Marcu and Echihabi (2002), using a small set of patterns to build relation models, and extend their work by refining the training and classification process using parameter optimization, topic segmentation and syntactic parsing. Using human-annotated and automatically-extracted test sets, we find that each of these techniques results in improved relation classification accuracy. 1

Abstract Predicting discourse structure on naturally occurring texts and dialogs is challenging and computationally intensive. Attempts to construct hand-built systems have run into problems both in how to specify the required knowledge and how to perform the necessary computations in an efficient manner. Data-driven approaches have recently shown to be successful for handling challenging aspects of discourse without using lots of fine-grained semantic detail, but they require annotated material for training. We describe our effort to annotate Segmented Discourse Representation Structures on Wall Street Journal texts, arguing that graph-based representations are necessary for adequately capturing the dependencies found in the data. We then explore two data-driven parsing strategies for recovering discourse structures. We show that the generative PCFG model of B&L is inherently limited by its inability to incorporate new features when learning from small data sets, and we show how recent developments in dependency parsing and discriminative learning can be utilized to get around this problem and thereby improve parsing accuracy. Results from exploratory experiments on Verbmobil dialogs and our annotated news wire texts are given; these results suggest that these methods do indeed enhance performance and have the potential for significant further improvements by developing richer feature sets.

Existing works indicate that the absence of explicit discourse connectives makes it difficult to recognize implicit discourse relations. In this paper we attempt to overcome this difficulty for implicit relation recognition by automatically inserting discourse connectives between arguments with the use of a language model. Then we propose two algorithms to use these predicted connectives. One is to use these predicted implicit connectives as additional features in a supervised model. The other is to perform implicit relation recognition based only on these predicted connectives. Results on Penn Discourse Treebank 2.0 show that predicted discourse connectives help implicit relation recognition and the first algorithm can achieve an absolute average f-score improvement of 3 % over a state of the art baseline system. 1

In this paper, we address the task of automatically determining which discourse relation holds between two text spans. We focus on relations that are not explicitly signalled by a discourse marker like but. While lexical models have been found useful for the task, they are also prone to data sparseness problems, which is a big drawback given the scarcity of discourse annotated data. We therefore investigate whether the use of lexical-semantic resources, such as WordNet, can be exploited to back-off to a more general representation of lexical information in cases were data are sparse. We compare such a semantic back-off strategy to morphological generalisations over word forms, such as stemming and lemmatising. 1

We explore the task of predicting which discourse relation holds between two text spans in which the relation is not signalled by an unambiguous discourse marker. It has been proposed that automatically labelled data, which can be derived from examples in which a discourse relation is unambiguously signalled, could be used to train a machine learner to perform this task reasonably well. However, more recent results suggest that there are problems with this approach, probably due to the fact that the automatically labelled data has particular properties which are not shared by the data to which the classifier is then applied. We investigate how big this problem really is and whether the unrepresentativeness of the automatically labelled data can be overcome by performing automatic example and feature selection. 1

In this paper, we present an integrated model of the two central tasks of dialog management: interpreting user actions and generating system actions. We model the interpretation task as a classi�cation problem and the generation task as a prediction problem. These two tasks are interleaved in an incremental parsing-based dialog model. We compare three alternative parsing methods for this dialog model using a corpus of human-human spoken dialog from a catalog ordering domain that has been annotated for dialog acts and task/subtask information. We contrast the amount of context provided by each method and its impact on performance. 1