Incorporating Multiple Knowledge Sources for Targeted Aspect-based Financial Sentiment Analysis

2023 | ACM Transactions on Management Information Systems | Citations: 0

Authors: Du, Kelvin; Xing, Frank; Cambria, Erik

Abstract: Combining symbolic and subsymbolic methods has become a promising strategy as re ... Expand

Abstract: Combining symbolic and subsymbolic methods has become a promising strategy as research tasks in AI grow increasingly complicated and require higher levels of understanding. Targeted Aspect-based Financial Sentiment Analysis (TABFSA) is an example of such complicated tasks, as it involves processes like information extraction, information specification, and domain adaptation. However, little is known about the design principles of such hybrid models leveraging external lexical knowledge. To fill this gap, we define anterior, parallel, and posterior knowledge integration and propose incorporating multiple lexical knowledge sources strategically into the fine-tuning process of pre-trained transformer models for TABFSA. Experiments on the Financial Opinion mining and Question Answering challenge (FiQA) Task 1 and SemEval 2017 Task 5 datasets show that the knowledge-enabled models systematically improve upon their plain deep learning counterparts, and some outperform state-of-the-art results reported in terms of aspect sentiment analysis error. We discover that parallel knowledge integration is the most effective and domain-specific lexical knowledge is more important according to our ablation analysis. Collapse

Topics: microblog domain-specific language Twitter social media knowledge graph

Methods: sentiment analysis RoBERTa transformer model BERT XLNet

Multi-Channel Convolutional Neural Network for the Identification of Eyewitness Tweets of Disaster

2023 | Information Systems Frontiers | Citations: 0

Authors: Kumar, Abhinav; Singh, Jyoti Prakash; Rana, Nripendra P.; Dwivedi, Yogesh K.

Abstract: During a disaster, a large number of disaster-related social media posts are wid ... Expand

Abstract: During a disaster, a large number of disaster-related social media posts are widely disseminated. Only a small percentage of disaster-related information is posted by eyewitnesses. The post of a disaster eyewitness offers an accurate depiction of the disaster. Therefore, the information posted by the eyewitness is preferred over the other source of information as it is more effective at helping organize rescue and relief operations and potentially saving lives. In this work, we propose a multi-channel convolutional neural network (MCNN) that uses three different word-embedding vectors together to classify disaster-related tweets into eyewitness, non-eyewitness, and don’t know classes. We compared the performance of the proposed multi-channel convolutional neural network with several attention-based deep-learning models and conventional machine learning-models such as recurrent neural network, gated recurrent unit, Long-Short-Term-Memory, convolutional neural network, logistic regression, support vector machine, and gradient boosting. The proposed multi-channel convolutional neural network achieved an F1-score of 0.84, 0.88, 0.84, and 0.86 with four disaster-related datasets of floods, earthquakes, hurricanes, and wildfires, respectively. The experimental results show that the training MCNN model with different word embedding together performs better than the conventional machine-learning models and several other deep-learning models. Collapse

Topics: Twitter earthquake hurricane crowdsourcing social media

Methods: machine learning deep learning convolutional neural network experiment GloVe

A deep multi-embedding model for mobile application recommendation

2023 | Decision Support Systems | Citations: 0

Authors: Liu, Yi-Hung; Chen, Yen-Liang; Chang, Po-Ya

Abstract: With the explosion in the number of mobile applications, it becomes difficult an ... Expand

Abstract: With the explosion in the number of mobile applications, it becomes difficult and time-consuming for users to find the most suitable and interesting application from the millions of applications that exist today. Therefore, mobile app recommendation becomes a top priority to help people solve this problem. In previous studies, there are two common methods to solve this problem, one is to apply matrix factorization to the rating data to predict the target user's ratings for other applications, and the other is to predict ratings based on the user's review text. In addition to combining the above two traditional methods, this study also combines deep learning methods to generate better models. For rating data, we use a deep matrix factorization method to obtain latent factor vectors of users and items from the rating matrix to improve the embedding of ratings. For review text, we improve the text representation by using a multi-embedding approach. Most importantly, our model also incorporates deep neural networks with nonlinear transformation capabilities to extract more representative abstract features. We conduct experiments on datasets collected from the real world. Extensive experimental results show that our proposed method has better performance than other existing methods. Collapse

Topics: mobile application mobile application market recommender system digital distribution mobile system

Methods: BERT deep learning experiment sensitivity analysis artificial neural network

Artificial Intelligence for Sign Language Translation – A Design Science Research Study

2023 | Communications of the Association for Information Systems | Citations: 0

Authors: Strobel, Gero; Schoormann, Thorsten; Banh, Leonardo; Möller, Frederik

Abstract: Although our digitalized society is able to foster social inclusion and integra ... Expand

Abstract: Although our digitalized society is able to foster social inclusion and integration, there are still numerous communities suffering from inequality. This is also the case with deaf people. About 750,000 deaf people in the European Union and over 4 million deaf people in the United States face daily challenges in terms of communication and participation. This occurs not only in leisure activities but also, and more importantly, in emergency situations. To provide equal environments and allow people with hearing handicaps to communicate in their native language, this paper presents an AI-based sign language translator. We adopted a transformer neural network capable of analyzing over 500 data points from a person's gestures and face to translate sign language into text. We have designed a machine learning pipeline that enables the translator to evolve, build new datasets, and train sign language recognition models. As proof of concept, we instantiated a sign language interpreter for an emergency call with over 200 phrases. The overall goal is to support people with hearing inabilities by enabling them to participate in economic, social, political, and cultural life. Collapse

Topics: artificial intelligence communication service infrastructure smartphone information system use innovation management

Methods: machine learning design artifact machine translation artificial neural network qualitative interview

Theories: socio technical theory

Don’t Need All Eggs in One Basket: Reconstructing Composite Embeddings of Customers from Individual-Domain Embeddings

2023 | ACM Transactions on Management Information Systems | Citations: 0

Authors: Unger, Moshe; Li, Pan; Sen, Sahana (Shahana); Tuzhilin, Alexander

Abstract: Although building a 360-degree comprehensive view of a customer has been a long- ... Expand

Abstract: Although building a 360-degree comprehensive view of a customer has been a long-standing goal in marketing, this challenge has not been successfully addressed in many marketing applications because fractured customer data stored across different “silos” are hard to integrate under “one roof” for several reasons. Instead of integrating customer data, in this article we propose to integrate several domain-specific partial customer views into one consolidated or composite customer profile using a Deep Learning-based method that is theoretically grounded in Kolmogorov’s Mapping Neural Network Existence Theorem. Furthermore, our method needs to securely access domain-specific or siloed customer data only once for building the initial customer embeddings. We conduct extensive studies on two industrial applications to demonstrate that our method effectively reconstructs stable composite customer embeddings that constitute strong approximations of the ground-truth composite embeddings obtained from integrating the siloed raw customer data. Moreover, we show that these data-security preserving reconstructed composite embeddings not only perform as well as the original ground-truth embeddings but significantly outperform partial embeddings and state-of-the-art baselines in recommendation and consumer preference prediction tasks. Collapse

Topics: recommender system data security collaborative filtering marketing management Google+

Methods: BERT artificial neural network survey experiment machine learning

Unlocking the Power of Voice for Financial Risk Prediction: A Theory-Driven Deep Learning Design Approach

2023 | Management Information Systems Quarterly | Citations: 0

Authors: Yang, Yi; Qin, Yu; Fan, Yangyang; Zhang, Zhongju

Abstract: Unstructured multimedia data (text and audio) provides unprecedented opportunit ... Expand

Abstract: Unstructured multimedia data (text and audio) provides unprecedented opportunities to derive actionable decision-making in the financial industry, in areas such as portfolio and risk management. However, due to formidable methodological challenges, the promise of business value from unstructured multimedia data has not materialized. In this study, we use a design science approach to develop DeepVoice, a novel nonverbal predictive analysis system for financial risk prediction, in the setting of quarterly earnings conference calls. DeepVoice forecasts financial risk by leveraging not only what managers say (verbal linguistic cues) but also how managers say it (vocal cues) during the earnings conference calls. The design of DeepVoice addresses several challenges associated with the analysis of nonverbal communication. We also propose a two-stage deep learning model to effectively integrate managers' sequential vocal and verbal cues. Using a unique dataset of 6,047 earnings call samples (audio recordings and textual transcripts) of S&P 500 firms across four years, we show that DeepVoice yields remarkably lower risk forecast errors than that achieved by previous efforts. The improvement can also translate into nontrivial economic gains in options trading. The theoretical and practical implications of analyzing vocal cues are discussed.Challenge 1: Time variation. Managers' vocal features can vary substantially during conference calls. For example, during the presentation section, when the manager is reading a forward-looking statement, the voice pitch may be quite stable. However, individuals' vocal cues may change as a result of being nervous, uncertain, or excited (Scherer et al., 1991). As such, in the question-answer section, when the manager is being scrutinized by analysts, the voice pitch may change. Mayew et al. (2020) found that the voice pitch of managers during dialogues with analysts is lower by about 4 Hz, on average, compared to their voice pitch during presentation sessions. Therefore, the time-varying patterns of managers' vocal cues may contain useful information. However, prior studies have ignored the time-varying pattern of managers' vocal cues in earnings conference calls (Mayew et al., 2020;, and it is unclear how a predictive model could effectively summarize time-varying vocal sequences.Challenge 2: Vocal-verbal integration. According to nonverbal communication theory (Mehrabian, 1972), vocal cues play a key role in nonverbal communication, as they can either affirm or discredit a message. For example, if a manager claims the company has strong future growth but uses a pessimistic tone of voice, market participants listening to the conference call may spot the irregularity and interpret the verbal message differently. This implies that to accurately gauge the information contained in managers' communication, one should integrate verbal language with the accompanying vocal cues. However, existing computational models (Poria et al., 2017) are limited in terms of integrating these two different communication channels (i.e., vocal and verbal).Challenge 3: Vocal complexity. Identifying the specific acoustic constructs from voice is a complex process, and the literature is still evolving without any consensus on an appropriate model for measuring affective states, emotions, and trust (Mayew & Venkatachalam, 2013). There is also little agreement about how a speaker's vocal cues should be combined to identify emotional states (Schuller, 2010)-e.g., which vocal cues to investigate and measure and how much weight to place on each vocal cue. As a result, attempts to develop constructs from vocal cues may suffer from high measurement errors. Collapse

Topics: database system enterprise information system website accounting social media

Methods: deep learning experiment meta design machine learning design artifact

Theories: transaction cost economics communication theory

sDTM: A Supervised Bayesian Deep Topic Model for Text Analytics

2023 | Information Systems Research | Citations: 0

Authors: Yang, Yi; Zhang, Kunpeng; Fan, Yangyang

Abstract: Topic modeling methods such as latent Dirichlet allocation (LDA) are powerful to ... Expand

Abstract: Topic modeling methods such as latent Dirichlet allocation (LDA) are powerful tools for analyzing massive amounts of textual data. They have been used extensively in information systems (IS) and business discipline research to identify latent topics for data exploration and as a feature engineering mechanism to derive new variables for analyses. However, existing topic modeling approaches are mostly unsupervised and only leverage textual data, while ignoring additional useful metadata often associated with text, such as star ratings in customer reviews or categories of posts in online forums. As a result, the identified topics and variables derived based on the learned topic model may not be accurate, which could lead to incorrect estimations that affect subsequent empirical analysis and to inferior performance on predictive tasks. In this study, we propose a novel supervised deep topic modeling approach called sDTM, which combines a neural variational autoencoder model and a recurrent neural network. sDTM leverages the auxiliary data associated with text to enhance the topic modeling capability. We conduct empirical case studies and predictive analytics on an online consumer review data set and an online knowledge community data set. Experimental results show that in comparison with benchmark methods, sDTM can enhance both the empirical estimation and predictive performance. sDTM makes methodological contributions to the IS literature and has direct relevance for research using text analytics.History: Alok Gupta, Senior Editor; Huimin Zhao, Associate Editor.Supplemental Material: The online appendix is available at https://doi.org/10.1287/isre.2022.1124. Collapse

Topics: online review open source user-generated content Wikipedia accounting

Methods: topic model deep learning machine learning BERT data transformation

Unintended Emotional Effects of Online Health Communities: A Text Mining-Supported Empirical Study

2023 | Management Information Systems Quarterly | Citations: 0

Authors: Zhou, Jiaqi; Zhang, Qingpeng; Zhou, Sijia; Li, Xin; Zhang, Xiaoquan (Michael)

Abstract: Online health communities (OHCs) play an important role in enabling patients to ... Expand

Abstract: Online health communities (OHCs) play an important role in enabling patients to exchange information and obtain social support from each other. However, do OHC interactions always benefit patients? In this research, we investigate different mechanisms by which OHC content may affect patients' emotions. Specifically, we notice users can read not only emotional support intended to help them but also emotional support targeting other persons or posts that are not intended to generate any emotional support (auxiliary content). Drawing from emotional contagion theories, we argue that even though emotional support may benefit targeted support seekers, it could have a negative impact on the emotions of other support seekers. Our empirical study on an OHC for depression patients supports these arguments. Our findings are new to the literature and have critical practical implications since they suggest that we should carefully manage OHC-based interventions for depression patients to avoid unintended consequences. We design a novel deep learning model to differentiate emotional support from auxiliary content. Such differentiation is critical for identifying the negative effect of emotional support on unintended recipients. We also discuss options to alter the intervention volume, length, and frequency to tackle the challenge of the negative effect. Collapse

Topics: web forum website application programming interface web page online community

Methods: machine learning deep learning robustness check instrumental variables estimation statistical hypothesis test

Theories: social contagion theory appraisal theory

TRG-DAtt: The Target Relational Graph and Double Attention Network Based Sentiment Analysis and Prediction for Supporting Decision Making

2022 | ACM Transactions on Management Information Systems | Citations: 0

Authors: Chen, Fan; Xia, Jiaoxiong; Gao, Honghao; Xu, Huahu; Wei, Wei

Abstract: The management of public opinion and the use of big data monitoring to accuratel ... Expand

Abstract: The management of public opinion and the use of big data monitoring to accurately judge and verify all kinds of information are valuable aspects in the enterprise management decision-making process. The sentiment analysis of reviews is a key decision-making tool for e-commerce development. Most existing review sentiment analysis methods involve sequential modeling but do not focus on the semantic relationships. However, Chinese semantics are different from English semantics in terms of the sentence structure. Irrelevant contextual words may be incorrectly identified as cues for sentiment prediction. The influence of the target words in reviews must be considered. Thus, this paper proposes the TRG-DAtt model for sentiment analysis based on target relational graph (TRG) and double attention network (DAtt) to analyze the emotional information to support decision making. First, dependency tree-based TRG is introduced to independently and fully mine the semantic relationships. We redefine and constrain the dependency and use it as the edges to connect the target and context words. Second, we design dependency graph attention network (DGAT) and interactive attention network (IAT) to form the DAtt and obtain the emotional features of the target words and reviews. DGAT models the dependency of the TRG by aggregating the semantic information. Next, the target emotional enhancement features obtained by the DGAT are input to the IAT. The influence of each target word on the review can be obtained through the interaction. Finally, the target emotional enhancement features are weighted by the impact factor to generate the review's emotional features. In this study, extensive experiments were conducted on the car and Meituan review data sets, which contain consumer reviews on cars and stores, respectively. The results demonstrate that the proposed model outperforms the existing models. Collapse

Topics: decision making

Methods: experiment computational algorithm BERT word embedding machine learning

Cross-Lingual Cybersecurity Analytics in the International Dark Web with Adversarial Deep Representation Learning

2022 | Management Information Systems Quarterly | Citations: 0

Authors: Ebrahimi, Mohammadreza; Chai, Yidong; Samtani, Sagar; Chen, Hsinchun

Abstract: International dark web platforms operating within multiple geopolitical regions ... Expand

Abstract: International dark web platforms operating within multiple geopolitical regions and languages host a myriad of hacker assets such as malware, hacking tools, hacking tutorials, and malicious source code. Cybersecurity analytics organizations employ machine learning models trained on human-labeled data to automatically detect these assets and bolster their situational awareness. However, the lack of human-labeled training data is prohibitive when analyzing foreign-language dark web content. In this research note, we adopt the computational design science paradigm to develop a novel IT artifact for cross-lingual hacker asset detection (CLHAD). CLHAD automatically leverages the knowledge learned from English content to detect hacker assets in non-English dark web platforms. CLHAD encompasses a novel Adversarial deep representation learning (ADREL) method, which generates multilingual text representations using generative adversarial networks (GANs). Drawing upon the state of the art in cross-lingual knowledge transfer, ADREL is a novel approach to automatically extract transferable text representations and facilitate the analysis of multilingual content. We evaluate CLHAD on Russian, French, and Italian dark web platforms and demonstrate its practical utility in hacker asset profiling, and conduct a proof-of-concept case study. Our analysis suggests that cybersecurity managers may benefit more from focusing on Russian to identify sophisticated hacking assets. In contrast, financial hacker assets are scattered among several dominant dark web languages. Managerial insights for security managers are discussed at operational and strategic levels. Collapse

Topics: darknet IT security knowledge sharing shapley additive explanation spoofing

Methods: machine learning design science design artifact case study deep learning