This study concerns how to segment a scenario-driven multiparty dialogue and how to label these segments automatically. We apply approaches that have been proposed for identifying topic boundaries at a coarser level to the problem of identifying agenda-based topic boundaries in scenario-based meetings. We also develop conditional models to classify segments into topic classes. Experiments in topic segmentation show that a supervised classification approach that combines lexical and conversational features outperforms the unsupervised lexical chain-based approach, achieving 20% and 12% improvement on segmentating top-level and sub-topic segments respectively. Experiments in topic classification suggest that it is possible to automatically categorize segments into appropriate topic classes given only the transcripts. Training with features selected using the Log Likelihood ratio improves the results by 13.3%.

SummaryObjectives: Consumer Health Informatics (CHI) and the use of Patient-Generated Health Data (PGHD) are rapidly growing focus areas in healthcare. The objective of this paper is to briefly review the literature that has been published over the past few years and to provide a sense of where the field is going. Methods: We searched PubMed and the ACM Digital Library for articles published between 2014 and 2016 on the topics of CHI and PGHD. The results of the search were screened for relevance and categorized into a set of common themes. We discuss the major topics covered in these articles. Results: We retrieved 65 articles from our PubMed query and 32 articles from our ACM Digital Library query. After a review of titles, we were left with 47 articles to conduct our full article survey of the activities in CHI and PGHD. We have summarized these articles and placed them into major categories of activity. Within the domain of consumer health informatics, articles focused on mobile health and patient-generated health data comprise the majority of the articles published in recent years. Conclusions: Current evidence indicates that technological advancements and the widespread availability of affordable consumer-grade devices are fueling research into using PGHD for better care. As we observe a growing number of (pilot) developments using various mobile health technologies to collect PGHD, major gaps still exist in how to use the data by both patients and providers. Further research is needed to understand the impact of PGHD on clinical outcomes. KeywordsConsumer health information/methods; patient-generated health data; mHealth; user-computer interface; consumer participation in delivery of health careYearb Med Inform 2017:152-9 http://dx

Compared to the traditional one-size-fits-all, nomothetic model that generalizes population-evidence for individuals, the proposed N-of-1 model can better capture the individual difference in their stressbehavior pathways. In this paper, we demonstrate it is feasible to perform personalized exercise behavior prediction, mainly made possible by mobile health technology and machine learning analytics.

AMI Meeting Facilitator is a system that performs topic segmentation and extractive summarisation. It consists of three components: (1) a segmenter that divides a meeting into a number of locally coherent segments, (2) a summarizer that selects the most important utterances from the meeting transcripts. and (3) a compression component that removes the less important words from each utterance based on the degree of compression the user specied. The goal of the AMI Meeting Facilitator is two-fold: rst, we want to provide sucient visual aids for users to interpret what is going on in a recorded meeting; second, we want to support the development of downstream information retrieval and information extraction modules with the information about the topics and summaries in meeting segments. The AMI Meeting Segmenter is trained using a set of 50 meetings that are seperate from the input meeting. We rst extract features from the audio and video recording of the input meeting in order to train the Maximum Entropy (MaxEnt) models for classifying topic boundaries and non-topic boundaries. Then we test each utterance in the input meeting on the Segmenter to see if it is a topic boundary or not. The features we use include the following ve categories: (1) Conversational Feature: These include a set of seven conversational features, including the amount of overlapping speech, the amount of silence between speaker segments, the level of similarity of speaker activity, the number of cue words, and the predictions of LCSEG (i.e., the lexical cohesion statistics, the estimated posterior probability, the predicted class). (2) Lexical Feature: Each spurt is represented as a vector space of uni-grams, wherein a vector is 1 or 0 depending on whether the cue word appears in the spurt. (3) Prosodic Feature: These include dialogue-act (DA) rate-of-speech, maximum F0 of the DA, mean energy of the DA, amount of silence in the DA, precedent and subsequent pauses, and duration of the DA. (4) Motion Feature: These include the average magnitude of speaker movements, which is measured by the number of pixels changed, over the frames of 40 ms within the spurt. (5) Contextual Feature: These include the dialogue act types and the speaker role (e.g., project manager, marketing expert). In the dialogue act annotations, each dialogue act is classied as one of the 15 types. SummarizationThe AMI summarizer is trained using a set of 98 scenario meetings. We train a support vector machine (SVM) on these meetings, using 26 features relating to the following categories: (1) Prosodic Features: These include dialogueact (DA) rate-of-speech, maximum F0 of the DA, mean energy of the DA, amount of silence in the DA, precedent and subsequent pauses, 9

Decision making is an important aspect of meetings in organisational settings, and archives of meeting recordings constitute a valuable source of information about the decisions made. However, standard utilities such as playback and keyword search are not sufficient for locating decision points from meeting archives. In this paper, we present the AMI DecisionDetector, a system that automatically detects and highlights where the decision-related conversations are. In this paper, we apply the models developed in our previous work [1], which detects decision-related dialogue acts (DAs) from parts of the transcripts that have been manually annotated as extract-worthy, to the task of detecting decision-related DAs and topic segments directly from complete transcripts. Results show that we need to combine features extracted from multiple knowledge sources (e.g., lexical, prosodic, DA-related, and topical class) in order to yield the model with the highest precision. We have provided a quantitative account of the feature class effects. As our ultimate goal is to operate AMI DecisionDetector in a fully automatic fashion, we also investigate the impacts of using automatically generated features, for example, the 5-class DA features obtained in [2]. keywords: Spoken language understanding, meeting tracking and analysis, argumentation modelling0.