Domestic violence (DV) is not only a major health and welfare issue but also a violation of human rights. In recent years, domestic violence crisis support (DVCS) groups active on social media have proven indispensable in the support services provided to victims and their families. In the deluge of onlinegenerated content, the significant challenge arises for DVCS groups' to manually detect the critical situation in a timely manner. For instance, the reports of abuse or urgent financial help solicitation are typically obscured by a vast amount of awareness campaigns or prayers for the victims. The state-of-the-art deep learning models with the embeddings approach have already demonstrated superior results in online text classification tasks. The automatic content categorization would address the scalability issue and allow the DVCS groups to intervene instantly with the exact support needed. Given the problem identified, the study aims to: 1) construct the novel ''gold standard' dataset from social media with multi-class annotation; 2) perform the extensive experiments with multiple deep learning architectures; 3) train the domain-specific embeddings for performance improvement and knowledge discovery; and 4) produce the visualizations to facilitate models analysis and results in interpretation. The empirical evidence on a ground truth dataset has achieved an accuracy of up to 92% in classes prediction. The study validates an application of cutting edge technology to a real-world problem and proves beneficial to DVCS groups, health care practitioners, and most of all victims.

The paper aims to leverage the highly unstructured user-generated content in the context of pollen allergy surveillance using neural networks with character embeddings and the attention mechanism. Currently, there is no accurate representation of hay fever prevalence, particularly in real-time scenarios. Social media serves as an alternative to extract knowledge about the condition, which is valuable for allergy sufferers, general practitioners, and policy makers. Despite tremendous potential offered, conventional natural language processing methods prove limited when exposed to the challenging nature of user-generated content. As a result, the detection of actual hay fever instances among the number of false positives, as well as the correct identification of non-technical expressions as pollen allergy symptoms poses a major problem. We propose a deep architecture enhanced with character embeddings and neural attention to improve the performance of hay fever-related content classification from Twitter data. Improvement in prediction is achieved due to the character-level semantics introduced, which effectively addresses the out-of-vocabulary problem in our dataset where the rate is approximately 9%. Overall, the study is a step forward towards improved real-time pollen allergy surveillance from social media with state-of-art technology.

Online product reviews underpin nearly all e-shopping activities. The high volume of data, as well as various online review quality, puts growing pressure on automated approaches for informative content prioritization. Despite a substantial body of literature on review helpfulness prediction, the rationale behind specific feature selection is largely under-studied. Also, the current works tend to concentrate on domain- and/or platform-dependent feature curation, lacking wider generalization. Moreover, the issue of result comparability and reproducibility occurs due to frequent data and source code unavailability. This study addresses the gaps through the most comprehensive feature identification, evaluation, and selection. To this end, the 30 most frequently used content-based features are first identified from 149 relevant research papers and grouped into five coherent categories. The features are then selected to perform helpfulness prediction on six domains of the largest publicly available Amazon 5-core dataset. Three scenarios for feature selection are considered: (i) individual features, (ii) features within each category, and (iii) all features. Empirical results demonstrate that semantics plays a dominant role in predicting informative reviews, followed by sentiment, and other features. Finally, feature combination patterns and selection guidelines across domains are summarized to enhance customer experience in today’s prevalent e-commerce environment. The computational framework for helpfulness prediction used in the study have been released to facilitate result comparability and reproducibility.

No abstract

Automatic helpfulness prediction aims to prioritize online product reviews by quality. Existing methods have combined review content and star ratings for automatic helpfulness prediction. However, the relationship between review content and star ratings is not explicitly captured, which limits the capability of rating information in influencing review content. This paper proposes a deep neural architecture to learn the explicit content-rating interaction (ECRI) for automatic helpfulness prediction. Specifically, ECRI explores two methods to interact review content with star ratings and adaptively specify the amount of rating information needed by review content. ECRI is evaluated against state-of-the-art methods on six real-world domains of the Amazon 5-core dataset. Experimental results demonstrate that exploiting the explicit content-rating interaction improves automatic helpfulness prediction. The source code of ECRI can be obtained from https://github.com/ tokawah/ECRI.