Semi-supervised Topic Model Research Articles

The Humanitarian Crisis in the Media: Framing Analysis of Rohingya-Related International News Using BERTopic

ABSTRACT The Rohingya crisis is one of the most severe humanitarian tragedies of our time, yet its portrayal in international media often lacks justification. Utilizing BERTopic, a transformer-based semi-supervised topic modeling technique, we analyzed 8,074 news articles about the Rohingya crisis, spanning newspapers from ten countries between 2009 and 2023. We aimed to observe the framing of the Rohingya crisis across countries and track frames’ evolution over time. We identified eleven frames: Politics, Departure, Repatriation, Human Rights, Resettlement, Refuge, Relief, Crisis, Retribution, Condolence, and Press Freedom, with the Political frame as the most frequent. Notably, host and non-host countries differed in framing the crisis, with Retribution and Crisis frames exclusive to host countries’ media, while Resettlement and Press Freedom frames were exclusive to non-host countries’ media. Host countries focused on Repatriation, Human Rights, and Relief news, whereas non-host countries prioritized Politics, Departure, and Refuge frames. Additionally, neighboring host countries were less likely to publish news on Resettlement. Changes in frames displayed significant variability across countries, indicating a lack of consistently identifiable patterns and suggesting that news media's framing habits may prioritize national interests over the chronological progression of the crisis.

Understanding the rationales and information environments for early, late, and nonadopters of the COVID-19 vaccine

Anti-vaccine sentiment during the COVID-19 pandemic grew at an alarming rate, leaving much to understand about the relationship between people’s vaccination status and the information they were exposed to. This study investigated the relationship between vaccine behavior, decision rationales, and information exposure on social media over time. Using a cohort study that consisted of a nationally representative survey of American adults, three subpopulations (early adopters, late adopters, and nonadopters) were analyzed through a combination of statistical analysis, network analysis, and semi-supervised topic modeling. The main reasons Americans reported choosing to get vaccinated were safety and health. However, work requirements and travel were more important for late adopters than early adopters (95% CI on OR of [0.121, 0.453]). While late adopters’ and nonadopters’ primary reason for not getting vaccinated was it being too early, late adopters also mentioned safety issues more often and nonadopters mentioned government distrust (95% CI on OR of [0.125, 0.763]). Among those who shared Twitter/X accounts, early adopters and nonadopters followed a larger fraction of highly partisan political accounts compared to late adopters, and late adopters were exposed to more neutral and pro-vaccine messaging than nonadopters. Together, these findings suggest that the decision-making process and the information environments of these subpopulations have notable differences, and any online vaccination campaigns need to consider these differences when attempting to provide accurate vaccine information to all three subpopulations.

Taking sides: Corporate social responsibility and political ideology

AbstractIn the United States and other countries, large private firms are increasingly more likely to take public political stances on controversial issues. Firms behave as ideological actors when they take sides in large public debates on social and economic issues. These stances may involve public speech, changes in the terms of service, or in internal reorganization and intrafirm actions. Despite the increase in firms' political actions, firms' ideological orientations remain unexamined. To answer this question, we collect corporate social responsibility statements from Fortune 1000 firms. Using semisupervised topic models, we identify topics that reflect stances on environmental and social issues. We then examine if firms are more likely to take stances on political issues due to pragmatic or ideological factors. We find that while pragmatic considerations play a role, firms' political stances are more driven by the ideological orientation of employees and managers. This research contributes a novel measure of firm ideology and sheds new light on the determinants of corporate political activity.

Improving fraud detection with semi-supervised topic modeling and keyword integration.

Fraud detection through auditors' manual review of accounting and financial records has traditionally relied on human experience and intuition. However, replicating this task using technological tools has represented a challenge for information security researchers. Natural language processing techniques, such as topic modeling, have been explored to extract information and categorize large sets of documents. Topic modeling, such as latent Dirichlet allocation (LDA) or non-negative matrix factorization (NMF), has recently gained popularity for discovering thematic structures in text collections. However, unsupervised topic modeling may not always produce the best results for specific tasks, such as fraud detection. Therefore, in the present work, we propose to use semi-supervised topic modeling, which allows the incorporation of specific knowledge of the study domain through the use of keywords to learn latent topics related to fraud. By leveraging relevant keywords, our proposed approach aims to identify patterns related to the vertices of the fraud triangle theory, providing more consistent and interpretable results for fraud detection. The model's performance was evaluated by training with several datasets and testing it with another one that did not intervene in its training. The results showed efficient performance averages with a 7% increase in performance compared to a previous job. Overall, the study emphasizes the importance of deepening the analysis of fraud behaviors and proposing strategies to identify them proactively.

Integrating Human Insights Into Text Analysis: Semi-Supervised Topic Modeling of Emerging Food-Technology Businesses’ Brand Communication on Social Media

Textual social media data have become indispensable to researchers’ understanding of message strategies and other marketing practices. In a new departure for the field of brand communication, this study adopts and extends a semi-supervised machine-learning approach, guided latent Dirichlet allocation (LDA), which incorporates human insights into the discovery and classification of topics. We used it to analyze tweets from businesses involved with an emerging food technology, cultured meat, and delineated four key message strategies used by these brands: providing functional, educational, corporate social responsibility, and relational content. We further ascertained the relationships between brands and the key topics embedded in their Twitter data. A comparison of model performance suggests that guided LDA can be an advantageous alternative to traditional LDA, which is characterized by high efficiency and immense popularity among researchers, but—because of its unsupervised nature—yields findings that can be difficult to interpret. The present study therefore has critical theoretical and methodological implications for communication and marketing scholars.

Discovering the Research Topics on Construction Safety and Health Using Semi-Supervised Topic Modeling

Safety and health have been one of the major issues in the construction industry worldwide for decades, and the relevant research has correspondingly drawn much attention in the academic field. Considering the expanding size and increasing heterogeneity of this research field, this paper proposes the topic modeling approach to cluster latent topics, extract coherent keywords, and discover evolving trends over the past three decades. Focusing on a total of 1984 articles published in 27 different journal sources until February 2023, this paper applied both unsupervised topic modeling techniques—Latent Dirichlet Allocation (LDA) and Correlation Explanation (CorEx)—and their semi-supervised versions—Guided LDA and Anchored CorEx. The evolving trends and inter-relationship of 15 research topics generated by the Anchored CorEx model (the best-performing model) were analyzed. Top-listed documents of major topics were analyzed to discuss their standalone research focuses. The results of this paper provided helpful insights and implications of existing research and offered potential guides for future research on construction safety and health by helping researchers (1) select research topics of interest and clearing decaying topics; (2) extract the top words of each research topic using systematic approaches; and (3) explore the interconnection of different research topics as well as their standalone focuses.

Automatized visualization methods for the qualitative analysis of teachers' collegial discussions

In this paper, we discuss a set of procedures for the visualization of conversation data, based on machine learning and text mining algorithms. The methodological function of these visualization methods is to support (rather than substitute) humanistic and interpretive qualitative data analysis. The context of our methodological discussion is the study of teachers’ learning communities. Enabling collegial learning among teachers has been argued to play a pivotal role in both teachers’ professional development and students’ learning. The problem is that the interpretive study of collegial discussions among teachers, which are typically considered a fruitful context for the study of teachers' collegial learning, is both time-consuming and cognitively demanding. Several scholars have discussed the potential of including machine learning and text mining in the toolbox of qualitative data analysis as a way to analyze large qualitative data sets and solve the time demand problem. However, the literature discussing the use of machine learning and text mining for qualitative research typically conceives the algorithms as substitutes for the work of the researcher. These automatized approaches are introduced to categorize the data, substituting the researcher in the analysis process. In contrast, we discuss the potential rewards of using automatized methods to support the analysis of teachers’ collegial discussions. We introduce a set of visualization procedures based on semi-supervised topic modeling and supervised classification algorithms and discuss how these visualizations can provide additional elements useful for the analysis of collegial discussions. Whereas these visualizations do not substitute interpretive analysis, we argue that they can support researchers’ interpretations by reducing data, identifying patterns of interactions, and possibly, visualizing group-level learning.

How do Gambling Providers Use the Social Network Twitter in Germany? An Explorative Mixed-Methods Topic Modeling Approach

This study examines the social media activities of gambling providers in Germany, focusing on the platform Twitter. A collection of 34.151 tweets from 13 Twitter accounts was made, representing casinos, sports betting, state lotteries, social lotteries and lottery brokers. We apply an explorative mixed-methods approach, integrating a summative content analysis together with a semi-supervised guided topic modeling approach, to analyse frequency, number of followers, interaction and content of Twitter messages, and work out differences among the individual providers. The results show that Twitter does not seem to be particularly important for gambling providers who are active in Germany. Regarding outreach, frequency of tweets and interaction, Twitter plays a much smaller role than, for example, in the UK. The potential for Twitter to be an advertising, marketing and interaction channel has not yet been fully exploited, which would make it a perfect moment for developing an appropriate regulatory framework. Overall, the results of semi-supervised topic modeling show that providers most often use informative content (news) and combine it with other, usually less neutral content. It is alarming that many providers make little or no use of Responsible Gambling messages. Even though the activities are presently on a low level, they contribute to the normalization of gambling, setting incentives for increasing gambling participation. Children and youth form a large segment of Twitter users. Potential harm can arise especially for this group because of the lack of enforcement of any age limit.

Exploring sentiment divergence on migrant workers through the lens of Sina Weibo

PurposeSince the opening of China (aka, reform and opening-up), a great number of rural residents have migrated to large cities in the past 40 years. Such a one-way population inflow to urban areas introduces nontrivial social conflicts between urban natives and migrant workers. This study aims to investigate the most discussed topics about migrant workers on Sina Weibo along with the corresponding sentiment divergence.Design/methodology/approachAn exploratory-descriptive-explanatory research methodology is employed. The study explores the main topics on migrant workers discussed in social media via manual annotation. Subsequently, guided LDA, a semi-supervised topic modeling approach, is applied to describe the overall topical landscape. Finally, the authors verify their theoretical predictions with respect to the sentiment divergence pattern for each topic, using regression analysis.FindingsThe study identifies three most discussed topics on migrant workers, namely wage default, employment support and urban/rural development. The regression analysis reveals different diffusion patterns contingent on the nature of each topic. In particular, this study finds a positive association between urban/rural development and the sentiment divergence, while wage default exhibits an opposite relationship with sentiment divergence.Originality/valueThe authors combine unique characteristics of social media with well-established theories of social identity and framing, which are applied more to off-line contexts, to study a unique phenomenon of migrant workers in China. From a practical perspective, the results provide implications for the governance of urbanization-related social conflicts.

Knowledge Source Rankings for Semi-Supervised Topic Modeling

Recent work suggests knowledge sources can be added into the topic modeling process to label topics and improve topic discovery. The knowledge sources typically consist of a collection of human-constructed articles, each describing a topic (article-topic) for an entire domain. However, these semisupervised topic models assume a corpus to contain topics on only a subset of a domain. Therefore, during inference, the model must consider which article-topics were theoretically used to generate the corpus. Since the knowledge sources tend to be quite large, the many article-topics considered slow down the inference process. The increase in execution time is significant, with knowledge source input greater than 103 becoming unfeasible for use in topic modeling. To increase the applicability of semisupervised topic models, approaches are needed to speed up the overall execution time. This paper presents a way of ranking knowledge source topics to satisfy the above goal. Our approach utilizes a knowledge source ranking, based on the PageRank algorithm, to determine the importance of an article-topic. By applying our ranking technique we can eliminate low scoring article-topics before inference, speeding up the overall process. Remarkably, this ranking technique can also improve perplexity and interpretability. Results show our approach to outperform baseline methods and significantly aid semisupervised topic models. In our evaluation, knowledge source rankings yield a 44% increase in topic retrieval f-score, a 42.6% increase in inter-inference topic elimination, a 64% increase in perplexity, a 30% increase in token assignment accuracy, a 20% increase in topic composition interpretability, and a 5% increase in document assignment interpretability over baseline methods.

Portability of semantic and spatial\u2013temporal machine learning methods to analyse social media for near-real-time disaster monitoring

Up-to-date information about an emergency is crucial for effective disaster management. However, severe restrictions impede the creation of spatiotemporal information by current remote sensing-based monitoring systems, especially at the beginning of a disaster. Multiple publications have shown promising results in complementing monitoring systems through spatiotemporal information extracted from social media data. However, various monitoring system criteria, such as near-real-time capabilities or applicability for different disaster types and use cases, have not yet been addressed. This paper presents an improved version of a recently proposed methodology to identify disaster-impacted areas (hot spots and cold spots) by combining semantic and geospatial machine learning methods. The process of identifying impacted areas is automated using semi-supervised topic models for various kinds of natural disasters. We validated the portability of our approach through experiments with multiple natural disasters and disaster types with differing characteristics, whereby one use case served to prove the near-real-time capability of our approach. We demonstrated the validity of the produced information by comparing the results with official authority datasets provided by the United States Geological Survey and the National Hurricane Centre. The validation shows that our approach produces reliable results that match the official authority datasets. Furthermore, the analysis result values are shown and compared to the outputs of the remote sensing-based Copernicus Emergency Management Service. The information derived from different sources can thus be considered to reliably detect disaster-impacted areas that were not detected by the Copernicus Emergency Management Service, particularly in densely populated cities.

Economic Narratives and Market Outcomes: A Semi-Supervised Topic Modeling Approach

Economic Narratives and Market Outcomes: A Semi-Supervised Topic Modeling Approach

Economic Narratives and Market Outcomes: A Semi-supervised Topic Modeling Approach

I employ the seeded Latent Dirichlet Process (sLDA) model in natural language processing to extract the narratives discussed by Shiller (2019) from nearly seven million New York Times articles over the past 150 years. The estimation scheme is designed to avoid any look-ahead bias in constructing the monthly narrative weights. Among the narratives considered, the most important one is Panic, which encompasses various stress- and anxiety-related themes including economic downturns, wars, political tensions, and epidemics. I find that Panic and a narrative index that loads heavily on Panic are strong positive predictors of excess U.S. market return and negative predictors of both realized and implied market volatility. I document empirical support for Panic as a proxy for time-varying risk aversion, consistent with a univariate version of the intertemporal capital asset pricing model (ICAPM). The predictability of narratives over market returns holds at both market and portfolio level and at both monthly and daily interval, and importantly is increasing over time.

Gaussian Mixture Variational Autoencoder for Semi-Supervised Topic Modeling

Topic models are widely explored for summarizing a corpus of documents. Recent advances in Variational AutoEncoder (VAE) have enabled the development of black-box inference methods for topic modeling in order to alleviate the drawbacks of classical statistical inference. Most existing VAE based approaches assume a unimodal Gaussian distribution for the approximate posterior of latent variables, which limits the flexibility in encoding the latent space. In addition, the unsupervised architecture hinders the incorporation of extra label information, which is ubiquitous in many applications. In this paper, we propose a semi-supervised topic model under the VAE framework. We assume that a document is modeled as a mixture of classes, and a class is modeled as a mixture of latent topics. A multimodal Gaussian mixture model is adopted for latent space. The parameters of the components and the mixing weights are encoded separately. These weights, together with partially labeled data, also contribute to the training of a classifier. The objective is derived under the Gaussian mixture assumption and the semi-supervised VAE framework. Modules of the proposed framework are appropriately designated. Experiments performed on three benchmark datasets demonstrate the effectiveness of our method, comparing to several competitive baselines.

A novel method to identify emerging technologies using a semi-supervised topic clustering model: a case of 3D printing industry

There have been recent attempts to identify emerging technologies by using topic-based analysis, but many of them have methodological deficiencies. First, analyses are unsupervised, and unsupervised methods cannot incorporate supervised knowledge that is needed to better identify technological domains. Second, those methods lack semantic interpretation, as many of them still remain at word-level analyses, we developed a novel technology-identification method that uses a semi-supervised topic clustering model (Labeled Dirichlet Multi Mixture model) to integrate technological domain knowledge. The model also generates a sentence-level semantic technological topic description through the topic description method (Various-aspects Sentence-level Description) on information extraction. We used this novel method to analyze the technology of the 3D printing industry, and successfully identified emerging technologies by differentiating new topics from the traditional topics, the results effectively demonstrated the semantic technological topic description by showing sentences. This method could be of great interest to technology forecasters and relevant policy-makers.

Communicating Climate Mitigation and Adaptation Efforts in American Cities

City governments have a large role to play in climate change mitigation and adaptation policies, given that urban locales are responsible for disproportionately high levels of greenhouse gas (GHG) emissions and are on the “front lines” of observed and anticipated climate change impacts. This study examines how US mayors prioritize climate policies within the context of the city agenda. Employing a computer-assisted content analysis of over 2886 mayoral press releases related to climate change from 82 major American cities for the period 2010–2016, we describe and explain the extent to which city governments discuss mitigation and adaptation policies in official communications. Specifically, we rely on a semi-supervised topic model to measure key climate policy themes in city press releases and examine their correlates using a multilevel statistical model. Our results suggest that while mitigation policies tend to dominate the city agenda on climate policy, discussion of adaptation efforts has risen dramatically in the past few years. Further, our statistical analysis indicates that partisanship influences city discussion on a range of climate policy areas—including emissions, land use policy, and climate resiliency—while projected vulnerability to climatic risks only influences discussion of climate resiliency and adaptation efforts.

Sentiment-Aware Deep Recommender System With Neural Attention Networks

With the advent of web technology, user-generated textual reviews are becoming increasingly accumulated on many e-commerce websites. These reviews contain not only the user comments on different aspects of the products but also the user sentiments associated with the aspects. Although these user sentiments serve as vital side information for improving the performance of recommender systems, most existing approaches ignore to fully exploit them in modeling the fine-grained user-item interaction for improving recommender system performance. Thus, this paper proposes a sentiment-aware deep recommender system with neural attention network (SDRA), which can capture both the aspects of products and the underlying user sentiments associated with the aspects for improving the recommendation system performance. Particularly, a semi-supervised topic model is designed to extract the aspects of the product and the associated sentiment lexicons from the user textual reviews, which are then incorporated into a long short term memory (LSTM) encoder via an interactive neural attention mechanism for better learning of the user and item sentiment-aware representation. Furthermore, a co-attention mechanism is introduced to better model the fine-grained user-item interaction for improving predictive performance. The extensive experiments on different datasets showed that our proposed SDRA model can achieve better performance over the baseline approaches.

Getting to Why: Semi-Supervised Topic Modeling of Customer Purchase Histories

The design and marketing of new products is fundamentally about understanding a customer's underlying needs. In this paper we describe research to learn needs by analyzing a customer's past purchases. We leverage what companies already (believe that they) know about how customers solve problems in the form of guides, instructions, and/or recipes. Using a semi-supervised form of LDA topic modeling, we align customer purchases with lists of equipment and materials that define established solutions for known problems. From known needs and solutions, we seek to discover new ways that customers are solving (un)known problems in the unlabeled data. We evaluate the approach on purchase-data from 18,000 customers of a multi-billion dollar specialty retailer.

A semi-supervised topic model incorporating sentiment and dynamic characteristic

With the rapid popularization of social applications, various kinds of social media have developed into an important platform for publishing information and expressing opinion. Detecting hidden topics from the huge amount of user-generated contents is of great commerce value and social significance. However traditional text analysis approaches only focus on the statistical correlation between words, but ignore the sentiment tendency and the temporal properties which may have great effects on topic detection results. This paper proposed a Dynamic Sentiment-Topic (DST) model which can not only detect and track the dynamic topics but also analyze the shift of public's sentiment tendency towards certain topic. Expectation-Maximization algorithm was used in DST model to estimate the latent distribution, and we used Gibbs sampling method to sample new document set and update the hyper parameters and distributions. Experiments are conducted on a real dataset and the results show that DST model outperforms the existing algorithms in terms of topic detection and sentiment accuracy.

Sentiment-Aspect Analysis through Semi-Supervised Topic Modeling

Sentiment analysis based on the aspects of products or services is designed to explore subjective information such as attitudes and opinions in user-generated reviews. Although a great many of approaches have been proposed in detecting aspects and the relevant aspect-specific sentiments, most of them detect the latent aspects with no proper classifying them or classify them employing unsupervised topic modeling without predicting the sentiment towards these aspects. This paper proposes a novel sentimentaspect analysis probabilistic modeling framework consisting of Seeding words extraction and semi-supervised topic (SST) model based on Sentence-LDA. More specifically, the proposed methodology starts by capturing seeding words from the websites inherent semi-structured information about products or services description. Then, it employs the captured seeding words to instruct the discovery of aspects and relevant sentiment of products or services simultaneously. Experimental results show that significant improvements have been achieved by the proposed method with respect to other state-ofthe-art methods.