دانلود رایگان ترجمه مقاله شناسایی شایعات و تحقیقات اعتقادی در توییتر – ۲۰۱۶

شناسایی شایعات و تحقیقات مربوط به باور در توییتر

Rumor Identification and Belief Investigation on Twitter

Abstract

Social media users spend several hours a day to read, post and search for news on microblogging platforms. Social media is becoming a key means for discovering news. However, verifying the trustworthiness of this information is becoming even more challenging. In this study, we attempt to address the problem of rumor detection and belief investigation on Twitter. Our definition of rumor is an unverifiable statement, which spreads misinformation or disinformation. We adopt a supervised rumors classification task using the standard dataset. By employing the Tweet Latent Vector (TLV) feature, which creates a 100-d vector representative of each tweet, we increased the rumor retrieval task precision up to 0.972. We also introduce the belief score and study the belief change among the rumor posters between 2010 and 2016.

۱ Introduction

Traditionally television, radio channels, and newspapers were the only news sources available. They are still the top trusted news sources but there is a large new trend toward digital sources. A considerable ratio of newspaper readers now read them digitally and the number of people relying on social media as a news source doubled since 2010. Social media helps you post your news online by a single click, this feasibility leads novel breaking news to show up first on micro blogs. Twitter is one of the most popular microblogging platforms with more than 250 million users. Accessibility, speed and ease-of-use have made Twitter a valuable platform to read and share information. However, the same features which make Twitter or any microblogging platform a great resource, but combined with lack of supervision make them fertile grounds for malicious or accidental misinformation in social media. Accordingly, this can lead to harmful incidences especially in sensitive circumstances, which then could cause damaging effects on individuals and society. There are many information seekers who do not rely on a single source to get information, but this is not always a good solution since even other news outlets sometime rely on social media when it comes to novel breaking news. Smart phones enable everyone to capture and tweet every single moment hours before TV cameras arrive. Considering that, social media is an appealing option for those who crave novel tempting news but on the other hand, could deceive anyone by well-structured and formatted rumors. In this study we work on a standard dataset of rumors collected by Qazvinian et al. (Qazvinian et al., 2011). In their work, the definition of rumor is defined as a statement whose truth value is unverifiable or deliberately false. We are using the same definition and not investigating the stimulus behind rumors creation. We investigate the problem of detecting rumors in Twitter data. We start with the motivation behind this research, and then the history of similar studies about rumors is overviewed. Then the overall pipeline is exposed, in which we adopt a supervised machine learning framework, and then we investigate the belief change for president Obama rumors in three years, and finally, we compare our results to the current state of the art performance on the task. We prove that our approach yields superior results in comparison to other works to date. 2 Related Work There is an extension body of related works on trustworthiness and misinformation detection. In this section we only focus on closely related works on the Natural Language Processing field that concentrate on information propagation and trustworthiness on social media, and specially on Twitter. 2.1 Social media and Trustworthiness After the earthquake and tsunami occurred in Japan on March 11th 2012, Takahashi and Igata, (Takahashi and Igata, 2012) targeted two sets of related rumor tweets about the earthquake. They create the model to detect other candidate rumor tweets relying on a sequence of processes. Takahashi and Igata detect the target rumor list using the entities and then the re-tweet ratio for target rumors is calculated, and finally the clue keywords get extracted by analyzing the scoring of each content word w, using the ratio of word occurrence in correction tweets (num in correction(w)) over rumor tweets (num in rumor(w)). In a similar study, Soroush, (Vosoughi, 2015) proposes his two step rumor detection and verification model on the Boston Marathon bombing tweets. The Hierarchical-clustering model is applied for rumor detection, and after the feature engineering process, which contains linguistic, user identity, and pragmatic features, he adopts the Hidden Markov model to find the veracity of each rumor. Soroush also analyses the sentiment classification of tweets using the contextual Information, which shows how tweets in different spatial, temporal, and authorial contexts have, on average, different sentiments. Sina is the popular Chinese microbloging platform like Twitter. Yang et al. (Yang et al., 2012) studied the rumors classification problem on both Twitter and Sina. He extended his primary features including content, client, account, location, and propagation by adding client-based features, which refers to a program that is being used to post on a microblog and also the location-based feature, which is a binary feature, that indicates being inside or outside of China. Yang et al. cover a significant range of meta-data features and fewer sentiment and con textual features in the aforementioned work. The most relevant related works to ours are Qazvinian et al. (Qazvinian et al., 2011)(V11) which use three sets of features, including content-based, networkbased, and Twitter specific meme features. For content-based features, they extract lexical and partof-speech patterns. For network-based features, they build two features to capture four types of networkbased properties utilizing the log likelihood of retweet and reply properties in Tweets, and finally, the Twitter specific meme features include hashtags and URLs. In our previous work (Hamidiain and Diab, 2015)(S15) we used the V11 data set with a new set of features, more labels, different machine learning, and an experimental approach. We proposed Rumor Detection and Classification (RDC) within the context of microblogging social media and suggested Single-step and Two-step models (SRDC and TRDC) in a supervised manner and investigate the effectiveness of the proposed list of features and various preprocessing tasks.