The need for software applications that can assist with mental disorders has never been greater. Individuals suffering from mental illnesses often avoid consultation with a psychotherapist, because they do not realize the need, or because they cannot or will not face the social and economic consequences, which can be severe. Between ideal treatment by a human therapist and self-help websites lies the possibility of a helpful interaction with a language-using computer. A model of empathic response planning for sentence generation in a forthcoming automated psychotherapist is described here. The model combines emotional state tracking, contextual information from the patient’s history and continuously updated therapeutic goals to form suitable conceptual graphs that may then be realized as suitable textual sentences.
The RDF-to-text task has recently gained substantial attention due to the continuous growth of RDF knowledge graphs in number and size. Recent studies have focused on systematically comparing RDF-to-text approaches on benchmarking datasets such as WebNLG. Although some evaluation tools have already been proposed for text generation, none of the existing solutions abides by the Findability, Accessibility, Interoperability, and Reusability (FAIR) principles and involves RDF data for the knowledge extraction task. In this paper, we present BENG, a FAIR benchmarking platform for Natural Language Generation (NLG) and Knowledge Extraction systems with focus on RDF data. BENG builds upon the successful benchmarking platform GERBIL, is opensource and is publicly available along with the data it contains.
Bouayad-Agha et al. (2012) issued a content determination challenge in which researchers were asked to create systems that can automatically select content suitable for a first paragraph in a Wikipedia article from an RDF knowledge base of information about people. This article is a description of the system built at the University of Aberdeen. Our working assumption is that the target text should contain information that is commonly known about the target person. The Wikipedia’s manual of style mentions that “The lead [section] serves as an introduction to the article and a summary of its most important aspects1.” What is most important about a person is likely to be often mentioned in biographies and hence it is more likely to be commonly known. Our system was motivated by the notion of common ground, especially the way it was accounted for by (Clark and Marshall, 1981). Clark and Marshall (1981) introduce two categories of common ground: personal common ground shared by a small group of individuals and communal common ground shared by a community of people. We are most interested in the concept of communal common ground, which arises from the exposure to the same information within a community. For example, if there is a statue in front of your work place, you expect your colleagues to also know about this statue and so the information that there is a statue in front of you workplace becomes a part of the community knowledge (where the community are people who work at the same place). Our hypothesis is that if we take a corpus of documents produced by some large community (e.g., English speakers), we should be able to ap-
We propose a novel end-to-end framework for abstractive meeting summarization. We cluster sentences in the input into communities and build an entailment graph over the sentence communities to identify and select the most relevant sentences. We then aggregate those selected sentences by means of a word graph model. We exploit a ranking strategy to select the best path in the word graph as an abstract sentence. Despite not relying on the syntactic structure, our approach significantly outperforms previous models for meeting summarization in terms of informativeness. Moreover, the longer sentences generated by our method are competitive with shorter sentences generated by the previous word graph model in terms of grammaticality.
This position paper presents an on-goingwork on a natural language generationframework that is particularly tailored fornatural language generation from bodyarea networks. We present an overview of ...
We present an Integer Linear Programming model of content selection, lexicalization, and aggregation that we developed for a system that generates texts from OWL ontologies. Unlike pipeline architectures, our model jointly considers the available choices in these three text generation stages, to avoid greedy decisions and produce more compact texts. Experiments with two ontologies confirm that it leads to more compact texts, compared to a pipeline with the same components, with no deterioration in the perceived quality of the generated texts. We also present an approximation of our model, which allows longer texts to be generated efficiently.
Planning-based approaches to reference provide a uniform treatment of linguistic decisions, from content selection to lexical choice. In this paper, we show how the issues of lexical ambiguity, vagueness, unspecific descriptions, ellipsis, and the interaction of subsective modifiers can be expressed using a belief-state planner modified to support context-dependent actions. Because the number of distinct denotations it searches grows doublyexponentially with the size of the referential domain, we present representational and search strategies that make generation and interpretation tractable.
In this overview paper we present the outcome of the first content selection challenge from open semantic web data, focusing mainly on the preparatory stages for defining the task and annotating the data. The task to perform was described in the challenge’s call as follows: given a set of RDF triples containing facts about a celebrity, select those triples that are reflected in the target text (i.e., a short biography about that celebrity). From the initial nine expressions of interest, finally two participants submitted their systems for evaluation.
Paraphrasing is expressing the same semantic content using different linguistic means. Although previous work has addressed linguistic variations at different levels of language, paraphrasing in Turkish has not been yet thoroughly studied. This paper presents the first study towards Turkish paraphrase alignment. We perform an analysis of different types of paraphrases on a modest Turkish paraphrase corpus and present preliminary results on that analysis from different standpoints. We also explore the impact of human interpretation of paraphrasing on the alignment of paraphrase sentence pairs.
We present the first prototype of a handover report generator developed for the MIME (Managing Information in Medical Emergencies) project. NLG applications in the medical domain have been varied but most are deployed in clinical situations. We develop a mobile device for prehospital care which receives streamed sensor data and user input, and converts these into a handover report for paramedics.
This abstract describes a contribution to the 2013 KBGen Challenge from CNRS/LORIA and the University of Lorraine. Our contribution focuses on an attempt to automate the extraction of a Feature Based Tree Adjoining Grammar equipped with a unification based compositional semantics which can be used to generate from KBGen data.
Pointing gestures are pervasive in human referring actions, and are often combined with spoken descriptions. Combining gesture and speech naturally to refer to objects is an essential task in multimodal NLG systems. However, the way gesture and speech should be combined in a referring act remains an open question. In particular, it is not clear whether, in planning a pointing gesture in conjunction with a description, an NLG system should seek to minimise the redundancy between them, e.g. by letting the pointing gesture indicate locative information, with other, nonlocative properties of a referent included in the description. This question has a bearing on whether the gestural and spoken parts of referring acts are planned separately or arise from a common underlying computational mechanism. This paper investigates this question empirically, using machine-learning techniques on a new corpus of dialogues involving multimodal references to objects. Our results indicate that human pointing strategies interact with descriptive strategies. In particular, pointing gestures are strongly associated with the use of locative features in referring expressions.
We describe a statistical Natural Language Generation (NLG) method for summarisation of time-series data in the context of feedback generation for students. In this paper, we initially present a method for collecting time-series data from students (e.g. marks, lectures attended) and use example feedback from lecturers in a datadriven approach to content selection. We show a novel way of constructing a reward function for our Reinforcement Learning agent that is informed by the lecturers’ method of providing feedback. We evaluate our system with undergraduate students by comparing it to three baseline systems: a rule-based system, lecturerconstructed summaries and a Brute Force system. Our evaluation shows that the feedback generated by our learning agent is viewed by students to be as good as the feedback from the lecturers. Our findings suggest that the learning agent needs to take into account both the student and lecturers’ preferences.
We show that Nakatsu & White’s (2010) proposed enhancements to the SPaRKy Restaurant Corpus (SRC; Walker et al., 2007) for better expressing contrast do indeed make it possible to generate better texts, including ones that make effective and varied use of contrastive connectives and discourse adverbials. After first presenting a validation experiment for naturalness ratings of SRC texts gathered using Amazon’s Mechanical Turk, we present an initial experiment suggesting that such ratings can be used to train a realization ranker that enables higher-rated texts to be selected when the ranker is trained on a sample of generated restaurant recommendations with the contrast enhancements than without them. We conclude with a discussion of possible ways of improving the ranker in future work.
This paper introduces the problem of generating descriptions of n-dimensional spatial data by decomposing it via modelbased clustering. I apply the approach to the error function of supervised classification algorithms, a practical problem that uses Natural Language Generation for understanding the behaviour of a trained classifier. I demonstrate my system on a dataset taken from CoNLL shared tasks.
