Describing implicit phenomena in discourse is known to be a problematic task, from both theoretical and empirical perspectives. The present article contributes to this topic by a novel comparative analysis of two prominent annotation approaches to discourse relations (coherence relations) that were carried out on the same texts. We compare the annotation of implicit relations in the Penn Discourse Treebank 2.0, i.e. discourse relations not signaled by an explicit discourse connective, to the recently released analysis of signals of rhetorical relations in the RST Signalling Corpus (RST-SC). The intersection of corresponding pairs of relations is rather a small one, but it shows a clear tendency: unlike the overall signal distribution in the RST-SC, more than half of the signals in the studied intersection are of semantic type, formed mostly by loosely defined lexical chains. Our data transformation allows for a simultaneous depiction and detailed study of the two resources.

Ideally, the users of spoken dialogue systems should be able to speak at their own tempo. Thus, the systems needs to interpret utterances from various users correctly, even when the utterances contain pauses. In response to this issue, we propose an approach based on a posteriori restoration for incorrectly segmented utterances. A crucial part of this approach is to determine whether restoration is required. We use a classification-based approach, adapted to each user. We focus on each user’s dialogue tempo, which can be obtained during the dialogue, and determine the correlation between each user’s tempo and the appropriate thresholds for classification. A linear regression function used to convert the tempos into thresholds is also derived. Experimental results show that the proposed user adaptation approach applied to two restoration classification methods, thresholding and decision trees, improves classification accuracies by 3.0% and 7.4%, respectively, in cross validation.

To support a natural flow of a conversation between humans and automated agents, rhetoric structures of each message has to be analyzed. We classify a pair of paragraphs of text as appropriate for one to follow another, or inappropriate, based on both topic and communicative discourse considerations. To represent a multi-sentence message with respect to how it should follow a previous message in a conversation or dialogue, we build an extension of a discourse tree for it. Extended discourse tree is based on a discourse tree for RST relations with labels for communicative actions, and also additional arcs for anaphora and ontology-based relations for entities. We refer to such trees as Communicative Discourse Trees (CDTs). We explore syntactic and discourse features that are indicative of correct vs incorrect request-response or question-answer pairs. Two learning frameworks are used to recognize such correct pairs: deterministic, nearest-neighbor learning of CDTs as graphs, and a tree kernel learning of CDTs, where a feature space of all CDT sub-trees is subject to SVM learning. We form the positive training set from the correct pairs obtained from Yahoo Answers, social network, corporate conversations including Enron emails, customer complaints and interviews by journalists. The corresponding negative training set is artificially created by attaching responses for different, inappropriate requests that include relevant keywords. The evaluation showed that it is possible to recognize valid pairs in 70% of cases in the domains of weak request-response agreement and 80% of cases in the domains of strong agreement, which is essential to support automated conversations. These accuracies are comparable with the benchmark task of classification of discourse trees themselves as valid or invalid, and also with classification of multi-sentence answers in factoid question-answering systems. The applicability of proposed machinery to the problem of chatbots, social chats and programming via NL is demonstrated. We conclude that learning rhetoric structures in the form of CDTs is the key source of data to support answering complex questions, chatbots and dialogue management.

It is generally acknowledged that discourse markers are used differently in speech and writing, yet many general descriptions and most annotation frameworks are written-based, thus partially unfit to be applied in spoken corpora. This paper identifies the major distinctive features of discourse markers in spoken language, which can be associated with problems related to their scope and structure, their meaning and their tendency to co-occur. The description is based on authentic examples and is followed by methodological recommendations on how to deal with these phenomena in more exhaustive, speech-friendly annotation models.

This study explores strategies in pro and anti-shale organizations’ discourse by combining the Discourse-Historical Approach (Wodak, 2001) with corpus linguistics. With the help of keyword lists, collocations, concordances, and key semantic domains, the representations of shale gas extraction, relevant actors and argumentation schemes in opposing discourses of the pro-shale Marcellus Shale Coalition and anti-shale Americans Against Fracking were analyzed. The findings of the study show that the advocates presented shale gas as a bonus for the crisis-struck American society while backgrounding its environmental impacts. The opponents, on the other hand, represented shale gas as a threat to the American ecosystem and public health through an alarming and scientific discourse. The empirical findings of this study add to a growing body of literature on discursive strategies employed by opposing camps of environmental controversies.

Storytelling is a universal activity, but the way in which discourse structure is used to persuasively convey ideas and emotions may depend on cultural factors. Because first-person accounts of life experiences can have a powerful impact in how a person is perceived, the storyteller may instinctively employ specific strategies to shape the audience’s perception. Hypothesizing that some of the differences in storytelling can be captured by the use of narrative levels and subjectivity, we analyzed over one thousand narratives taken from personal weblogs. First, we compared stories from three different cultures written in their native languages: English, Chinese and Farsi. Second, we examined the impact of these two discourse properties on a reader’s attitude and behavior toward the narrator. We found surprising similarities and differences in how stories are structured along these two dimensions across cultures. These discourse properties have a small but significant impact on a reader’s behavioral response toward the narrator.

It has long been argued that accenting or stressing a pronoun (i.e., making it prosodically prominent) changes its interpretation as compared to its unaccented counterpart. However, recent experimental work demonstrated that this generalization does not apply when the alternative interpretation of the pronoun is not plausible (Taylor et al., 2013). In a series of three experiments that use an offline comprehension task, we show, first, that the lack of reversal is observed when plausibility is controlled for. We furthermore show that a new generalization cannot be formed by excluding cases where the bias towards the unmarked interpretation is strong or cases where the character in the alternative interpretation is low in salience. Instead, we conclude that what constrains the interpretation of accented pronouns is coherence relations, with parallel discourses exhibiting reversal and result discourses not exhibiting reversal. We propose that the difference between coherence relations should be viewed in what would be the minimal change in order to create a ‘surprising’ or expected’ event, which is the characteristic of accenting more generally.

Examples and specifications occur frequently in text, but not much is known about how they function in discourse and how readers interpret them. Looking at how they’re annotated in existing discourse corpora, we find that annotators often disagree on these types of relations; specifically, there is disagreement about whether these relations are elaborative (additive) or argumentative (pragmatic causal). To investigate how readers interpret examples and specifications, we conducted a crowdsourced discourse annotation study. The results show that these relations can indeed have two functions: they can be used to both illustrate/specify a situation and serve as an argument for a claim. These findings suggest that examples and specifications can have multiple simultaneous readings. We discuss the implications of these results for discourse annotation.

Understanding how the social context of an interaction affects our dialog behavior is of great interest to social scientists who study human behavior, as well as to computer scientists who build automatic methods to infer those social contexts. In this paper, we study the interaction of power, gender, and dialog behavior in organizational interactions. In order to perform this study, we first construct the Gender Identified Enron Corpus of emails, in which we semi-automatically assign the gender of around 23,000 individuals who authored around 97,000 email messages in the Enron corpus. This corpus, which is made freely available, is orders of magnitude larger than previously existing gender identified corpora in the email domain. Next, we use this corpus to perform a largescale data-oriented study of the interplay of gender and manifestations of power. We argue that, in addition to one’s own gender, the “gender environment” of an interaction, i.e., the gender makeup of one’s interlocutors, also affects the way power is manifested in dialog. We focus especially on manifestations of power in the dialog structure — both, in a shallow sense that disregards the textual content of messages (e.g., how often do the participants contribute, how often do they get replies etc.), as well as the structure that is expressed within the textual content (e.g., who issues requests and how are they made, whose requests get responses etc.). We find that both gender and gender environment affect the ways power is manifested in dialog, resulting in patterns that reveal the underlying factors. Finally, we show the utility of gender information in the problem of automatically predicting the direction of power between pairs of participants in email interactions.

Temporal information is one of the prominent features that determine the coherence in a discourse. That is why we need an adequate way to deal with this type of information during discourse annotation. In this paper, we will argue that temporal order is a relational rather than a segment-specific property, and that it is a cognitively plausible notion: temporal order is expressed in the system of linguistic markers and is relevant in both acquisition and language processing. This means that temporal relations meet the requirements set by the Cognitive approach of Coherence Relations (CCR) to be considered coherence relations, and that CCR would need a way to distinguish temporal relations within its annotation system. We will present merits and drawbacks of different options of reaching this objective and argue in favor of adding temporal order as a new dimension to CCR.

The connective because can express both highly objective and highly subjective causal relations. In this, it differs from its counterparts in other languages, e.g. Dutch, where two conjunctions omdat and want express more objective and more subjective causal relations, respectively. The present study investigates whether it is possible to anchor the different uses of because in context, examining a large number of syntactic, morphological and semantic cues with a minimal cost of manual annotation. We propose an innovative method of distinguishing between subjective and objective uses of because with the help of information available from an English/Dutch segment of a parallel corpus, which is accompanied by a distributional analysis of contextual features. On the basis of automatic syntactic and morphological annotation of approximately 1500 examples of because, every English sentence is coded semi-automatically for more than twenty contextual variables, such as the part of speech, number, person, semantic class of the subject, modality, etc. We employ logistic regression to determine whether these contextual variables help predict which of the two causal connectives is used in the corresponding Dutch sentences. Our results indicate that a set of semantic and syntactic features that include modality, semantics of referents (subjects), semantic class of the verbal predicate, tense (past vs. non-past) and the presence of evaluative adjectives, are reliable predictors of the more subjective and objective uses of because, demonstrating that this distinction can indeed be anchored in the immediate linguistic context. The proposed method and relevant contextual cues can be used for identification of objective and subjective relationships in discourse.

Discourse relations can either be explicitly marked by discourse connectives (DCs), such as therefore and but, or implicitly conveyed in natural language utterances. How speakers choose between the two options is a question that is not well understood. In this study, we propose a psycholinguistic model that predicts whether or not speakers will produce an explicit marker given the discourse relation they wish to express. Our model is based on two information-theoretic frameworks: (1) the Rational Speech Acts model, which models the pragmatic interaction between language production and interpretation by Bayesian inference, and (2) the Uniform Information Density theory, which advocates that speakers adjust linguistic redundancy to maintain a uniform rate of information transmission. Specifically, our model quantifies the utility of using or omitting a DC based on the expected surprisal of comprehension, cost of production, and availability of other signals in the rest of the utterance. Experiments based on the Penn Discourse Treebank show that our approach outperforms the state-of-the-art performance at predicting the presence of DCs (Patterson and Kehler, 2013), in addition to giving an explanatory account of the speaker’s choice.

This paper reports on an experiment implementing a data-intensive approach to discourse organisation. Its focus is on enumerative structures envisaged as a type of textual pattern in a sequentiality-oriented approach to discourse. On the basis of a large-scale annotation exercise calling upon automatic feature mark-up alongside manual annotation, we explore a method to identify complex discourse markers seen as configurations of cues. The presentation of the background to what is termed "multi-level annotation" is organised around four issues: linearity, complexity of discourse markers, top-down processing, granularity and the multi-level nature of discourse structures. In this context, enumerative structures seem to deserve scrutiny for a number of reasons: they are frequent structures appearing at different granularity levels, they are signalled by a variety of devices appearing to work together in complex ways, and they combine a textual role (discourse organisation) with an ideational role (categorisation). We describe the annotation procedure and experimental framework which resulted in nearly 1,000 enumerative structures being annotated in a diversified corpus of over 600,000 words. The results of two approaches to the rich data produced are then presented: firstly, a descriptive survey highlights considerable variation in length and composition, while showing enumerative structure to be a basic strategy resorted to in all three sub-corpora, and leads to a granularity-based typology of the annotated structures; secondly, recurrent cue configurations—our "complex  markers"—are identified by the application of data mining methods. The paper ends with perspectives for further exploitation of the data, in particular with respect to the semantic characterisation of enumerative structures.

In this paper, we construct and train end-to-end neural network-based dialogue systems usingan updated version of the recent Ubuntu Dialogue Corpus, a dataset containing almost 1 million multi-turn dialogues, with a total of over 7 million utterances and 100 million words. This dataset is interesting because of its size, long context lengths, and technical nature; thus, it can be used to train large models directly from data with minimal feature engineering, which can be both time consuming and expensive. We provide baselines in two different environments: one where models are trained to maximize the log-likelihood of a generated utterance conditioned on the context of the conversation, and one where models are trained to select the correct next response from a list of candidate responses. These are both evaluated on a recall task that we call Next Utterance Classification (NUC), as well as other generation-specific metrics. Finally, we provide a qualitative error analysis to help determine the most promising directions for future research on the Ubuntu Dialogue Corpus, and for end-to-end dialogue systems in general.

Many language learners never acquire truly native-sounding prosody. Previous work has suggested that this involves skill deficits in the dialog-related uses of prosody, and may be attributable to weaknesses with specific prosodic constructions. Using semi-automated methods, we identified 32 of the most common prosodic constructions in English dialog. Examining 90 minutes of six advanced native-Spanish learners conversing in English, there were differences, notably regarding swift turn-taking, alignment, and empathy, but overall their uses of prosodic constructions were largely similar to those of native speakers.