Sentiment Bias Research Articles

Generative AI Search Engines as Arbiters of Public Knowledge: An Audit of Bias and Authority

ABSTRACTThis paper reports on an audit study of generative AI systems (ChatGPT, Bing Chat, and Perplexity), which investigates how these new search engines construct responses and establish authority for topics of public importance. We collected system responses using a set of 48 authentic queries for 4 topics over 7 days, and analyzed the data using sentiment analysis, inductive coding, and source classification. Results provide an overview of the nature of system responses across these systems, and provide evidence of sentiment bias based on the queries and topics, and commercial and geographic biases in sources. The quality of sources used to support claims is uneven, relying heavily on News and Media, Business and Digital Media websites. Implications for system users emphasize the need to critically examine Generative AI system outputs when making decisions related to public interest and personal well‐being.

An aspect sentiment analysis model with Aspect Gated Convolution and Dual-Feature Filtering layers

Aspect level sentiment analysis is a basic task to determine the sentiment bias based on the contextual information near the aspect words. Some sentences contain many confusing feature words due to incomplete structure or high complexity in sentiment prediction, which can easily lead to the problem that the model pays too much attention to unimportant features. Furthermore, the role of aspect-related sentiment features is not significant in the feature extraction process. The capsule network, due to its complex network structure, leads to capsule detachment when too much information is found in the dataset. Too much computational resources are easily consumed during dynamic routing, which in turn affects the model's judgement of the polarity of aspect-related sentiment. To address these issues, we propose a capsule network (AGCDFF-Caps) with aspect-gated convolution and Dual-Feature Filtering layers for aspect-level sentiment analysis. A pre-trained BERT model is used to generate a serialised representation of the text, and the semantic representation in the contextual text is enhanced by the self-attention mechanism and bi-GRU. An aspect-gating based convolutional neural network is constructed to selectively extract contextual sentiment information about aspects and discard irrelevant information. A capsule network is then used to learn the multispatial semantic features of the aspect and the text. In particular, we incorporate a Dual-Feature Filtering network structure into the capsule network structure to strengthen the interaction between the particular aspect and the context from global and local perspectives, filtering the redundant information in local semantics and global semantics. The opinion feature representations that can more accurately express the emotional tendency of the aspect are obtained. Experimental results on SemEval2014 and Twitter datasets show that the proposed hybrid network structure has superior classification performance compared to 12 advanced baseline methods.

Evaluating the Impact of Internet Users' Emotional Tendencies on Opinion Dynamics through Computer-Assisted Sentiment Analysis of Texts

- Internet communication is characterised by "non-linear flow" and "unorganised aggregation," and computer-assisted text sentiment analysis is an important means of evaluating the impact of netizens' emotions on the dynamics of public opinion. The uncertainty, anonymity, disorder, and blindness inherent in fast-gathering and fast-dispersing Internet communication drive the continuous generation of texts, which has an important impact on public opinion. In order to study this issue in depth, this paper chooses two text types of "popular Weibo" and "popular topics" in Sina Weibo as the object of study and analyses the issue in detail from the two dimensions of agenda construction and agenda setting. In the process of the study, the required text data were collected through Python, and the theoretical assumptions were verified with text analysis methods to explore the influence of text sentiment bias on opinion diffusion, with a view to providing new insights and references for the application of the theory of opinion diffusion and the field of computational diffusion.

Intelligent Calibration for Bias Reduction in Sentiment Corpora Annotation Process

This paper focuses in the inherent anchoring bias present in sequential reviews-sentiment corpora annotation processes. It proposes employing a limited subset of meticulously chosen reviews at the outset of the process, as a means of calibration, effectively mitigating the phenomenon. Through extensive experimentation we validate the phenomenon of sentiment bias in the annotation process and show that its magnitude can be influenced by pre-calibration. Furthermore, we show that the choice of the calibration set matters, hence the need for effective guidelines for choosing the reviews to be included in it. A comparison of annotators performance with the proposed calibration to annotation processes that do not use calibration or use a randomly-picked calibration set, reveals that indeed the calibration set picked is highly effective---it manages to substantially reduce the average absolute error compared to the other cases. Furthermore, the proposed selection guidelines are found to be highly robust in picking an effective calibration set also for domains different than the one based on which these rules were extracted.

The impact of economic policy uncertainty on manager sentiment: evidence from Chinese non-financial listed firms

ABSTRACT In this study, we investigate the link between economic policy uncertainty (EPU) and manager sentiment. Analysing a comprehensive dataset of Chinese non-financial listed firms spanning from 2007 to 2021, the findings indicate a negative relationship between EPU and manager sentiment expressed in annual reports. Additionally, we observed a significant reduction in manager sentiment bias in the presence of high EPU. The impact of EPU on manager sentiment is moderated by firms benefiting from government industrial policy and those exhibiting stronger risk-taking capability, as the negative effect is less pronounced in these cases. Our channel analysis reveals that EPU diminishes manager sentiment by impairing operating performance and capital market performance.

The politicisation of historical memory on Twitter: “Positive antisemitism” in the Holocaust debate in Poland

ABSTRACT Political controversies concerning the memory of historical events have adapted to the digital specificity of Twitter and are currently growing in significance. This study investigates the sentiment bias of a memory war between two major Polish political parties, the currently ruling conservative Law and Justice (PiS) party and Civic Platform (PO), the main opposition party, concerning Polish-Jewish relations, and especially the Holocaust. Totally, 23699 PO Tweets and 29,441 PiS Tweets from 2015 to 2019 were collected, out of which, 241 concerning memory war were sampled with a lexicon-based approach. Sentiment was analysed using tweet-based consensual coding with an ordinal five-point scale and automated word-based coding with a three-point scale. PiS tweets had more positive sentiment regarding the memory of the Holocaust and Jewish legacy in Poland. The final hermeneutical analysis showed that the rulling party narratives were aimed appropriation of collective memory and “Polonizing” the Holocaust, and that the positive sentiment was a function of ‘positive anti-Semitism’.

Reducing Sentiment Bias in Pre-trained Sentiment Classification via Adaptive Gumbel Attack

Pre-trained language models (PLMs) have recently enabled rapid progress on sentiment classification under the pre-train and fine-tune paradigm, where the fine-tuning phase aims to transfer the factual knowledge learned by PLMs to sentiment classification. However, current fine-tuning methods ignore the risk that PLMs cause the problem of sentiment bias, that is, PLMs tend to inject positive or negative sentiment from the contextual information of certain entities (or aspects) into their word embeddings, leading them to establish spurious correlations with labels. In this paper, we propose an adaptive Gumbel-attacked classifier that immunes sentiment bias from an adversarial-attack perspective. Due to the complexity and diversity of sentiment bias, we construct multiple Gumbel-attack expert networks to generate various noises from mixed Gumbel distribution constrained by mutual information minimization, and design an adaptive training framework to synthesize complex noise by confidence-guided controlling the number of expert networks. Finally, we capture these noises that effectively simulate sentiment bias based on the feedback of the classifier, and then propose a multi-channel parameter updating algorithm to strengthen the classifier to recognize these noises by fusing the parameters between the classifier and each expert network. Experimental results illustrate that our method significantly reduced sentiment bias and improved the performance of sentiment classification.

Removing AI’s sentiment manipulation of personalized news delivery

Artificial intelligence (AI) is empowering personalized online news delivery to accommodate people’s information needs and combat information overload. However, AI models learned from user data are inheriting and amplifying some underlying human prejudice such as the sentiment bias of news reading, which may lead to potential negative societal effects and ethical concerns. Here, substantial evidence shows that AI is manipulating the sentiment orientation of news displayed to users by promoting the presence chance of negative news, even if there is no human interference. To mitigate this manipulation, a sentiment-debiasing method based on a decomposed adversarial learning framework is proposed, which can reduce 97.3% of sentiment bias with only 2.9% accuracy sacrifice. Our work provides the potential in improving AI’s responsibility in many human-centered applications such as online journalism and information spread.

Engineers, Aware! Commercial Tools Disagree on Social Media Sentiment: Analyzing the Sentiment Bias of Four Major Tools

Large commercial sentiment analysis tools are often deployed in software engineering due to their ease of use. However, it is not known how accurate these tools are, and whether the sentiment ratings given by one tool agree with those given by another tool. We use two datasets - (1) NEWS consisting of 5,880 news stories and 60K comments from four social media platforms: Twitter, Instagram, YouTube, and Facebook; and (2) IMDB consisting of 7,500 positive and 7,500 negative movie reviews - to investigate the agreement and bias of four widely used sentiment analysis (SA) tools: Microsoft Azure (MS), IBM Watson, Google Cloud, and Amazon Web Services (AWS). We find that the four tools assign the same sentiment on less than half (48.1%) of the analyzed content. We also find that AWS exhibits neutrality bias in both datasets, Google exhibits bi-polarity bias in the NEWS dataset but neutrality bias in the IMDB dataset, and IBM and MS exhibit no clear bias in the NEWS dataset but have bi-polarity bias in the IMDB dataset. Overall, IBM has the highest accuracy relative to the known ground truth in the IMDB dataset. Findings indicate that psycholinguistic features - especially affect, tone, and use of adjectives - explain why the tools disagree. Engineers are urged caution when implementing SA tools for applications, as the tool selection affects the obtained sentiment labels.

A study of emotion setting based on event evolutionary graph-Take microblog users’ expression of emotions on news reports related to the Beijing Winter Olympics as an example

With the prominence of the “emotion setting” function in media agenda setting, the sentiment bias of netizens has gradually become a noteworthy part of the communication of public opinion events. In this paper, we use the BERT model to analyze the sentiment of news headlines and netizens’ comments from January 2021 to July 2021, and extract news headlines and their relationships with the help of LTP to construct a theory map of news and comments for each month in Neo4j, and then explore the deep logic of netizens’ sentiment changes. The research results show that there is a correlation and lag between the emotional direction of Internet users and the emotional direction of media news reports in online public opinion dissemination. In addition, the public’s attention to information on international current affairs needs to be strengthened.

Empirical evidence of bias in public acceptance of autonomous vehicles

This study synthesizes 91 peer-reviewed survey studies examining the public acceptance of Autonomous Vehicles (AVs). The framework of the study is informed by three questions: (1) How well do the collected samples represent the acceptance of the general population? (2) How often does bias exist in measuring public acceptance in AV’s questionnaires? (3) How much bias persists in reporting public acceptance of AV’s research? The findings indicate that (1) people with disabilities and racial minorities are only included in 10% and 20% of the studies, respectively (2) 50% of the studies present their questionnaire, and most are perceived to be biased as a result of systematic errors such as leading questions, missing questions, and suggestive information, and (3) 72% of the studies suffer from the sentiment bias, where the positive tone in the title and abstract is more significant than in the result. This leads to imprecise findings and unrealistic depictions of acceptance of autonomous vehicles by the public. The analysis alerts researchers and practitioners to empirical evidence of bias in public acceptance of autonomous vehicles and recommends preventive actions.

Language and Domain Specificity: A Chinese Financial Sentiment Dictionary

Abstract We use Word2vec to develop a financial sentiment dictionary from 3.1 million Chinese-language financial news articles. Our dictionary maps semantically similar words to a subset of human-expert generated financial sentiment words. In validation tests, our dictionary scores the sentiment of articles consistently with human reading of full articles. In return association tests, our dictionary outperforms and subsumes previous Chinese financial sentiment dictionaries, such as direct translations of Loughran and McDonald’s (2011, Journal of Finance, 66, 35–65) English-language financial dictionary. We also generate a list of politically related positive words that is unique to China; we find that this list has a weaker association with returns than does the list of other positive words. We demonstrate that state media uses more politically related positive and fewer negative words, and exhibits a sentiment bias. This bias renders the state media’s sentiment as less return-informative. Our findings demonstrate that dictionary-based sentiment analysis exhibits strong language and domain specificity.

Modeling sentimental bias and temporal dynamics for adaptive deep recommendation system

Recommendation systems rely on the historic data of users' purchases and their feedbacks to profile their preferences and make future recommendations. Most of these systems usually employ Collaborative Filtering (CF) models to analyze users’ ratings and infer the latent factors which represent the user and item features in k-dimensional latent space. However, the historical rating data used for recommendations are usually sparsed and unbalanced. Various approaches have been used to resolve these issues by combining the user’s ratings and reviews to better capture the user’s sentiments and make accurate recommendations. Other challenges comprise changes in users’ preferences and items’ perceptions over time. Therefore, this paper presents a new Sentiment Scoring Model (SSM) based on Long-/Short-Term Memory and a combination function that catches the sentiment bias between user rating and review to relieve the sparsity and unbalanced dataset. Next, we proposed an Adaptive LSTM (ALSTM) method that can model the drifting of user and item features to improve the recommendation accuracy. We show the performance of our model on the three real-world rating datasets from Amazon reviews, which comprises Fine Food, Baby, and Cell-phone & Accessories categories. The result shows the superiority of our proposed model over the existing static and dynamic models. The statistical test shows that all the performance gains are significant at p < 0.05.

Sentiment Classification for Film Reviews by Reducing Additional Introduced Sentiment Bias

Film business and its individual reviews cannot be separated and film review sites such as IMDb is a credible source of reviews posted in public forums. With IMDb site reviews being unstructured and bias-heavy, classification methods by reducing additional sentiment bias is needed to create a balanced classification with lower polarity bias. Elimination of additional sentiment bias will improve the model as polarity is defined by non-bias method, resulting in models correctly defined which sequences of words is either positive or negative. This research limits the dataset by 50.000 rows of randomly extracted reviews from the IMDb website using dataset preparation methods such as Preprocessing, POS-Tagging, and Word Embeddings. Then preprocessed data is used in classification methods such as ANN, SWN, and SO-Cal. This paper also used bias processing methods such as Hyperparameter Tuning and BPM, with outputs evaluated using Accuracy and PBR metrics. This research yields 77.39 % for ANN, 66.32% for BPM, 75.6% for SO-Cal, and 76.26% for Hybrid classification. Best PBR resulted in two lexicon-based methods on 0.0009 for BPM, and 0.00006 for SO-Cal. More advanced model configuration in ANN can improve the model, and much complex lexicon models will be a future in the research topic.&#x0D;

Share pledge transactions as an investor sentiment indicator - Evidence from China

This paper examines how share pledge transactions signal changes in investor sentiment. Based on the investor sentiment model proposed by Gervais and Odean (2001), we develop an investor sentiment model by incorporating the impact of share pledge transactions on stock prices by linking the signal transmission of the share pledges with investor decision making. Using data from the China stock market from 2014 to 2018, we find that share pledge transactions can lead to emotional bias and consequently affect investment decisions. Specifically, stock market volatility increases with changes in both optimistic and pessimistic investor sentiment bias regarding share pledge transactions. From the perspective of behavioral finance, with the Vector Autoregression (VAR) model and the impulse response test, we study the relationship between investor sentiment indicators, including share pledge factors and stock market volatility. Among the investor sentiment indicator components, we find that the composition weights of the share pledge volume and the share pledge ratio are 22.54 %, and 21.49 % respectively. Share pledge transactions have an obvious impact on investor sentiment. The paper also shows that share pledge transactions have an accelerating effect on stock price cyclical fluctuations through the investor sentiment effect. During a stock index crash, such transactions may constitute a potential market factor of instability.

Manager Sentiment Bias and Stock Returns: Evidence from China

ABSTRACT Using textual analysis of Chinese listed firms from 2004 to 2017, we examine the role that managers’ sentiment plays in financial disclosures and its impact on firms’ future stock returns. We distinguish manager sentiment as either signal or noise according to its consistency with firm earnings. We find that good signal sentiment positively impacts stock returns, whereas bad signal sentiment and noisy sentiment that reflects overconfidence negatively impact stock returns. Further analysis shows that external supervision, internal control, and managers’ expertise contribute to improving the information quality of both signal and noisy sentiments, and the enactment of earnings forecast policy helps strengthen the signaling impact of manager sentiment.

Light a lamp and look at the stock market

In this study, we investigate the impact of the light-a-lamp event that occurred in India during the COVID-19 lockdown. This event happened across the country, and millions of people participated in it. We link this event to the stock market through investor sentiment and misattribution bias. We find a 9% hike in the market return on the post-event day. The effect is heterogeneous in terms of beta, downside risk, volatility, and financial distress. We also find an increase (decrease) in long-term bond yields (price), which together suggests that market participants demanded risky assets in the post-event day.

No magic bullet: estimating anti-immigrant sentiment and social desirability bias with the item-count technique

Extant scholarship on attitudes toward immigration and immigrants relies mostly on direct survey items. Thus, little is known about the scope of social desirability bias, and even less about its covariates. In this paper, we use probability-based mixed-modes panel data collected in the Southern Spanish region of Andalusia to estimate anti-immigrant sentiment with both the item-count technique, also known as list experiment, and a direct question. Based on these measures, we gauge the size of social desirability bias, compute predictor models for both estimators of anti-immigrant sentiment, and pinpoint covariates of bias. For most respondent profiles, the item-count technique produces higher estimates of anti-immigrant sentiment than the direct question, suggesting that self-presentational concerns are far more ubiquitous than previously assumed. However, we also find evidence that among people keen to position themselves as all-out xenophiles, social desirability pressures persist in the list-experiment: the full scope of anti-immigrant sentiment remains elusive even in non-obtrusive measurement.

Parimutuel betting on the eSports duels: Evidence of the reverse favourite-longshot bias

Abstract We analyse betting behaviour patterns of the visitors of the specialized betting website dedicated to the popular eSports game Counter-Strike: Global Offensive. The reverse favourite-longshot bias is found, which is rather unusual for parimutuel betting markets, where favourite-longshot bias is more common. We define simple betting strategies based on the bets on underdogs and show that these strategies make a sufficiently large positive profit both on the in-sample and out-of-sample datasets, which is a sign of market inefficiency. In matches with popular underdogs, the sentiment bias decreases the market inefficiency caused by the reverse favourite-longshot bias. At the same time, the role of the geographical location of teams in the reverse favourite-longshot bias is only marginal.

Casual bettors and sentiment bias in NBA and NFL betting

ABSTRACT Sentiment bias, defined as investment decisions made for reasons unrelated to fundamentals and related to popularity, represents a common research topic in finance and economics. Empirical research on sentiment bias frequently uses data from sports betting markets. While research generally finds evidence of sentiment bias, in particular, due to team popularity, the prevalence of this bias among fans and bettors in many sports remains unclear and identification of markets with relatively larger numbers of bettors with sentiment bias poses empirical challenges. We analyse outcomes in betting markets for games played in two North American team sports leagues, the National Basketball Association (NBA) and National Football League (NFL) from the 2012–13 to 2016–17 seasons and investigate how game timing, in terms of games played on weekdays and weekends, may affect the presence of bettors with popularity-based sentiment bias. Probit model estimates explaining the probability that a bet on a team wins indicate patterns consistent with the presence of bettors with sentiment bias in the markets for betting on games in the NFL but not in the NBA, and the presence of more such bettors for NFL games played on weekends.