CHAPTER 2: LITERATURE REVIEW
2.4 Concepts Related to Credibility, Their Criteria, and Operationalization
2.4.2 Operationalization of Credibility Criteria
To facilitate the communication of credible online information, previous studies have applied automatic methods for assessing the credibility of information (e.g., Castillo et al., 2011; Ciampaglia et al., 2015; Sondhi et al., 2012). These studies operationalized credibility factors as features for machine-learned models. The final goal of the dissertation study is to build and evaluate computational credibility models. In this section, I review credibility models for health information and social media. Some models are descriptive, and others are predictive.
Furthermore, I consider studies that focused on textual and non-textual sources of evidence. I describe how prior work operationalized credibility factors as features for machine learning according to three types (content, source, and others).
Operationalization related to information content
Mendoza et al. (2010) conducted a preliminary study to examine the dissemination of false rumors and true news through Twitter. According to their analysis, the dissemination of false rumors had a different pattern from true news. Twitter users questioned false rumors more than they did the true news and acted as a collaborative filter. True cases were only denied in 0.3% of cases and questioned in 2.5% of cases. In case of false rumors, about 50% of tweets were denied, and 12.8% of tweets were questioned. Thus, it will be helpful to create features that can operationalize patterns of discussion such as acceptance, questioning, and denial among users by analyzing the linguistic patterns of these interactions.
Benevuto et al. (2010) developed models to detect spammers on Twitter. Results found several trends. First, spammers were found to publish tweets with more URLs, spam words, and hashtags. Second, spammers were found to reference multiple trending topics in the same tweet. O’Donovan et al. (2012) found that credible tweets are more likely to contain language
phenomena (e.g., question, exclamation, smile, and sentiment) that is rare. In particular, credible tweets had more “news” feature that computes similarity between tweets and data archived at Twibe (a service archiving credible tweets and users).
Yamamoto and Tanaka (2011) selected five factors to judge the credibility of Web search results: accuracy, objectivity, authority, currency, and coverage. They operationalized those five factors by making several assumptions. First, they assumed that accurate Web pages are
referenced by other Web pages and assigned the scores of PageRank provided by Google (Agichtein, Castillo, Donato, Gionis, & Mishne, 2008) to accuracy factor. They defined social reputation as one criterion to determine the authority of Web pages and used Hatena Bookmark service (http://b.hatena.ne.jp/) to get the number of social bookmarks for Web pages. They also considered Web pages that are similar to other Web pages as objective and so used LexRank algorithm to calculate content typicality (Erkan & Radev, 2004). To determine the level of coverage in technical topics, they used the method of searching Wikipedia for technical terms used by Nakatani et al. (2009). Regarding currency, updated dates of Web pages were retrieved by using the WayBack Machine of Internet Archive (http://archive.org/web/). They developed a system to calculate scores of Web search results based on those five operational components of credibility. They conducted an experimental study to evaluate the search efficiency of their system by comparing the original Google interface and their system + Google interface. The
mean number of credible Web pages found by their system + Google was significantly larger than when only using Google (p <0.01).
Ciampaglia et al. (2015) utilized the shortest path between concept nodes in the semantic proximity metrics on the knowledge graph to evaluate the veracity of information. They
constructed the semantic proximity metrics by using a public knowledge graph extracted from Wikipedia. They extracted the factual statements of Wikipedia’s “infoboxes” from the DBpedia database (Auer et al., 2007). For example, the statement “Socrates is a person” can be
represented by a subject-predicate-object triple as “Socrates,” “is a,” “person.” These triples constitute a knowledge graph (KG) in which nodes denote entities (e.g., subject or object of the statement) and edges denote predicates. Given a statement, if a short path linking its subject to an object exists within the KG, the statement can be considered true. For instance, their best
algorithm (which used k-NN), achieved 96 in Area under Receiver Operating Characteristic (ROC). They exploited the implicit information from the topology of the Wikipedia Knowledge Graph (WKG) and operationalized it with the semantic proximity metrics. In the domain of health informatics, the Semantic Knowledge Representation based on the UMLS has been produced and maintained by the national library of medicine (NLM) (Aronson & Rindflesch, 1998). The NLM also developed the Semantic Knowledge Representation (SemRep) program that extracts semantic predications from text. These resources from the authoritative organization have potential usefulness for enriching knowledge graphs in improving decision making for credibility judgments.
Griffiths et al. (2005) developed an automated quality assessment procedure (AQA) to help users to find high-quality depression websites based on evidence-based medicine. They primarily used word frequency to generate quality queries consisting of terms that frequently
appear in high-quality websites but rarely appear in low-quality ones. Those quality queries could be expanded and filtered by applying the relevance feedback algorithm. They produced a golden dataset by using statements in the Oxford University Centre for Evidence-Based Mental Health’s guidelines for treating depression. The performance of AQA outperformed the Google PageRank algorithm regarding correlation with a golden dataset. It should be noted that authors applied this automated filtering approach only to depression websites.
Saglam and Temizel (2015) proposed a framework that automatically identifies and classifies diabetes websites according to relevance and information quality. This was one of a few studies that handled both relevance and information quality (Sağlam & Temizel, 2015; Seidman, Steinwachs, & Rubin, 2003). They adopted the relevance feedback approach introduced by Griffiths et al. (2005) and added sentiment analysis. The authors first filtered biased and untrusted websites out by using the subjectivity detection approach that was proposed by Denecke (2009) and based on SentiWordNet (http://sentiwordnet.isti.cnr.it/). They assumed that a sensational writing style (e.g., excessive exclamation marks) and the frequency of positive and negative words could reveal whether a site was biased or not. Then, the quality queries were expanded by using the relevance feedback algorithm. Several studies have demonstrated the effectiveness of simple word frequency, and these quality queries, which frequently appear in high-quality information, could be extended through relevance feedback algorithm.
Aphinyanaphongs et al. (2005) applied text categorization models to automatically identify high-quality articles in internal medicine. They utilized a term frequency-based text model and considered articles published in the ACP journal club which are peer-reviewed as the gold standard of high-quality articles. Compared to the previous Boolean-based PubMed clinical query filters by Haynes et al. (1994), authors’ approaches showed comparable results.
Aphinyanaphongs and Aliferis (2007) applied text categorization models to another topic. This study aimed to identify the quality of health information about unproven cancer treatments. Their approach was unique in that they did not identify high-quality pages, but low-quality pages. Thus, they adopted the approach of information avoidance (Sweeny, Melnyk, Miller, & Shepperd, 2010) as email service providers apply spam filter for their users. They used gold standard on unproven cancer treatments identified by experts at the Quackwatch. Their model outperformed the result (AUC=0.63) by Quacko meter (http://quackometer.net) that applied alternative medication terms for their modeling and also outperformed another baseline approach based on the PageRank algorithm (Brin & Page, 1998). This dissertation study also applies the information avoidance approach by focusing on classifying not_credible information, because it is a minority class and filtering spam information makes more sense in real life.
Operationalization related to information source
With the assumption that a user with more authority would register more tweets, Mendoza et al. (2010) investigated the number of followers and followees on Twitter. For top users, the number of followers is two times larger than the number of followees. The number of followers decreases when the number of tweets that a user created decreases. They also found that top users tended to be connected to one another by friendships. This implies that we can give more credentials to users who are socially connected to other users. Network analysis was
originated from sociology to find meaningful patterns in people's social networks and can produce discriminative features for the credibility models.
Benevuto et al. (2010), however, found that spammers had a higher ratio of followers-to- followees, a shorter history, and fewer mentions in tweets from followees. The contexts of two studies were different. The context of the study by Mendoza et al. (2010) was related to crisis
while the context of the study by Benevenuto et al. is related to a trending hot topic such as Michael Jackson’s death. Therefore, we need to more carefully examine the relationship between the number of followers and the number of followees for credibility judgment. The explanation by Mendoza et al. (2010) about spammers’ dual behavior, usually sharing non-spam tweets to pretend to be a non-spammer, but spreading spam tweets occasionally, partly explains the high ratio of followers per followees.
Gupta et al. (2013) collected user profiles that are suspended by Twitter during the event of the Boston Marathon Blasts and analyzed their characteristics. There were several interesting findings. First, many malicious accounts were made just after the bomb blast, and it took about 8 hours for fake tweets written by them to become viral. The source accounts of fake content had a low social status (e.g., a small number of followers) and unconfirmed identity. However, once other users with high social status (e.g., celebrities and politicians) retweeted those fake contents, the spread of the fake contents showed steep growth. The effect of the identity fraud might be related to previous studies in which the social status did not contribute to predicting credibility. Second, they found each group who tweeted true or fake tweets were unique to each other. In other words, fake users cooperated and formed a closed-community by retweeting each other’s tweets. Their analysis indicated that tracing back to the original source of information and observing their mutual interactions will be critical in identifying fake content.
Yardi et al. (2009) created a predictive model to detect spam tweets using features such as URLs, username pattern match, and social network metrics. In this study, legitimate users were likely to be found at the center of the Twitter users’ network graph rather than the edges. Spammers probably would be on the more isolated parts of the social network. They found a significant difference in the number of tweets per day between spammers and legitimate users
(spam user: 8.66 tweets per day, legitimate user: 6.7 tweets per day). The total number of followers and friends for spammers was three times larger than that of legitimate users. These results did not support the finding by Gupta et al. (2013). The generalizability of their findings to other contexts should be carefully examined.
Operationalization related to information content, source, and other features
Price and Hersh (1999) were the first pioneers who applied an empirical approach to find high-quality consumer health information on the Web. They developed a program to identify some quality criteria such as authorship information and Health On the Net Foundation’s Code of Conduct (HONcode) (Boyer, Selby, Scherrer, & Appel, 1998) acceptance, amount of content, currency, and average link score and return quality scores. They did not make clear the accuracy and precision of the study in the report on their experiment, and their results were too
preliminary to be considered conclusive, it is meaningful in that they opened the door of automatic assessment of quality in the domain of online health information.
Sondhi et al. (2012) developed a framework to automatically predict the reliability of medical information from the Web. The reliability judgments were made based on the quality criteria defined by the Health on Net Foundation. Authors produced two major features: link- based features (e.g., number of internal and external links, presence of contact, count of commercial links) and content-based features (word frequency, frequency of commercial keywords, and writing style). To construct a test set, they used medical websites that had been approved by the HON staff for the positive set and used medical websites that were criticized by Quackwatch for the negative set. Interestingly, they used their approach to re-rank Google’s results for 22 medical queries. The mean average precision (MAP) of re-ranked documents by their approach was significantly better than the MAP by Google’s search results.
Abbasi et al. (2012) utilized an adaptive learning algorithm called recursive trust labeling (RTL) that consists of content and graph-based classifiers. Content-based features comprising of fraud cues were extracted from the Web pages’ body text, source code, anchor text, and URLs, and word n-grams were used to represent body text features. To reduce redundancy in fraud cues, the authors utilized the Unified Medical Language System (UMLS) MetaThesaurus (https://www.nlm.nih.gov/research/umls/) to aggregate synonymous concept words. For graph- based classifier, they adopted dual class (legitimate and fake sites) with two ways (in-link and out-link), as the assumption of the PageRank principles that fake websites can point to legitimate websites, but cannot force legitimate websites to point to them (Page, Brin, Motwani, &
Winograd, 1999) may not be valid for fake medical websites. By comparing the result by RTL classifier with 19 other classifiers, they demonstrated that the recursive labeling mechanism enhanced detection of fake medical websites by effectively leveraging the information provided by these two classifiers.
Castillo et al. (2011) investigated how people judge the credibility of tweets concerned with trending events (e.g., earthquakes). Results found that both content- and source- related features were closely related to assessors’ credibility judgments. Content-related features were whether the user cited a source, if the user questioned the information, and if the user used a lot of opinionated statements in the tweet. Source-related features were account activity (e.g., frequency of posting), age of account, and number of followers. Also, tweets that have been re- tweeted many times tended to be more credible.
Another study by Castillo et al. (2013) used the crowdsourcing tool Mechanical Turk to generate ground truth data. To increase the quality of labeling from crowdsourcing, the
was given to seven assessors. If more than five assessors agreed then the label was utilized, otherwise ignored. They posited that there are several factors which are useful to predict
credibility of information: the reactions like positive and negative sentiment which are generated by certain topic, the level of certainty expressed by users who propagate the information,
external sources cited (e.g., top 100 popular domain), and characteristics of users (e.g., the number of followers). According to the feature selection study based on the correlation-based feature selection (CFS) subset evaluation, there were 16 most useful features such as the average number of tweets posted by the author in the topic in the past, the average number of followees, and the maximum depth of the propagation trees.
Xia et al. (2012) conducted an experiment with supervised learning to predict credibility of tweets. They collected 350 tweets and asked five experts to label them as credible or non- credible. Four categories of features were utilized: source-related feature (e.g., time interval of last 2 tweets and total number of tweets per day), content-related feature (length of the tweet and number of the reply comments), topic-related feature (e.g., the URL fraction of the tweet and the hashtags fraction of the tweet), and diffusion-related feature (e.g., time of the tweet cited and time of the original tweet was cited in case of retweet). Their approach achieved equivalent performance (precision: 0.61 and recall: 0.64) to the state-of-the-art SVM algorithm.
The study by Kakol et al. (2017) derived a comprehensive set of factors that affect credibility judgments using a huge dataset comprising of more than 5,500 Web pages. They used a crowdsourcing platform, Amazon Mechanical Turk (AMT), to collect credibility annotations of those Web pages and workers’ justifications of their credibility judgments. The Latent Semantic Analysis (LSA) and expectation-maximization clustering algorithm were applied to workers’ justifications to create a code book. Then, researchers extended this primitive code book and
made a comprehensive set of credibility factors. Based on the final code book, they ran the second round of crowdsourcing to collect credibility factors of previous workers’ justifications. They also created a predictive model of Web content credibility using these crowd workers’ credibility evaluations and credibility factors. In the linear regression model, the root mean square error (RMSE) reached 0.709 which was better than the random and constant value model. However, the features in their study were not automatically generated, but annotated by AMT workers. Thus, this approach has a limitation in terms of automation.
Most of the features described in the studies above focus on attributes of the content, author, format, etc. All details of the features for machine learning cannot be fully described here but features that are most likely to be discriminative in predicting the credibility of online social question & answer (Q & A) and reviews are summarized in Table 2.
Table 2. Promising Features in Predicting Credibility of Information and Source Type of
feature Feature Reference
Word shared between question and answer
Bian, Liu, Zhou, Agichtein, & Zha, 2009
Non-standard use of grammar and punctuation
Olteanu, Peshterliev, Liu, & Aberer, 2013; Yang, 2007
Part-of-Speech (e.g., number of
adverbs, adjectives, and nouns) Juffinger, Granitzer, & Lex, 2009 Readability (e.g., SMOG score) Eysenbach, Powell, Kuss, & Sa, 2002;
Olteanu et al., 2013
Text complexity (document entropy) X. Liu, Nielek, Wierzbicki, & Aberer, 2013
Text informative (e.g., tf-idf) X. Liu et al., 2013
Quantity of information (uncertainty reduction theory)
Adamic, Zhang, Bakshy, &
Ackerman, 2008; Agichtein, Castillo, Donato, Gionis, & Mishne, 2008; Daft & Lengel, 1986
Content-based
Vividness, descriptiveness, and strength of message (e.g., inclusion of personal experience, number of suggestions)
Adamic et al., 2008; Agichtein et al., 2008; Daft & Lengel, 1986
Perceived expertise (ratio of jargon) Jeon & Rieh, 2014 Frequency of modal and inferential
conjunction
Mukherjee, Weikum, & Danescu- Niculescu-Mizil, 2014
Sentiment (e.g., frequency of
affective words, attitude, tone) Kim, 2010; Mukherjee et al., 2014 Content similarity with validated
source Juffinger et al., 2009
Redundancy of the content Drost & Scheffer, 2005 Frequency of emoticons Weerkamp & Rijke, 2008 Fraction of content containing spam
words
Benevenuto, Magno, Rodrigues, & Almeida, 2010
Authorship (authors and their affiliation) and attribution
(references and sources of content)
Lederman, Fan, Smith, & Chang, 2014; Silberg, Lundberg, & Musacchio, 1997
Number of witnesses (e.g., number
of positive replies) Don Fallis, 2004; Hume, 1977
Link-based
PageRank Brin & Page, 1998
HITS Kleinberg, 1999
Network metrics (e.g., degree, eigenvector, betweenness)
Mui, Mohtashemi, & Halberstadt, 2002
Ratio of external/internal inbound
links to outbound links Liu, Wei, Zhang, & Zhou, 2013 Number of internal/external inbound
links Liu et al., 2013
Ratio of number of replies to
number of questions Liu et al., 2013 Maximum depth of the propagation
User-based
Entropy of the categories of user Liu et al., 2013