Tweet Text Research Articles

Location reference identification from tweets during emergencies: A deep learning approach

Twitter is recently being used during crises to communicate with officials and provide rescue and relief operation in real time. The geographical location information of the event, as well as users, are vitally important in such scenarios. The identification of geographic location is one of the challenging tasks as the location information fields, such as user location and place name of tweets are not reliable. The extraction of location information from tweet text is difficult as it contains a lot of non-standard English, grammatical errors, spelling mistakes, non-standard abbreviations, and so on. This research aims to extract location words used in the tweet using a Convolutional Neural Network (CNN) based model. We achieved the exact matching score of 0.929, Hamming loss of 0.002, and F1-score of 0.96 for the tweets related to the earthquake. Our model was able to extract even three- to four-word long location references which is also evident from the exact matching score of over 92%. The findings of this paper can help in early event localization, emergency situations, real-time road traffic management, localized advertisement, and in various location-based services.

Deep neural networks for bot detection

The problem of detecting bots, automated social media accounts governed by software but disguising as human users, has strong implications. For example, bots have been used to sway political elections by distorting online discourse, to manipulate the stock market, or to push anti-vaccine conspiracy theories that may have caused health epidemics. Most techniques proposed to date detect bots at the account level, by processing large amounts of social media posts, and leveraging information from network structure, temporal dynamics, sentiment analysis, etc. In this paper, we propose a deep neural network based on contextual long short-term memory (LSTM) architecture that exploits both content and metadata to detect bots at the tweet level: contextual features are extracted from user metadata and fed as auxiliary input to LSTM deep nets processing the tweet text. Another contribution that we make is proposing a technique based on synthetic minority oversampling to generate a large labeled dataset, suitable for deep nets training, from a minimal amount of labeled data (roughly 3000 examples of sophisticated Twitter bots). We demonstrate that, from just one single tweet, our architecture can achieve high classification accuracy (AUC > 96%) in separating bots from humans. We apply the same architecture to account-level bot detection, achieving nearly perfect classification accuracy (AUC > 99%). Our system outperforms previous state of the art while leveraging a small and interpretable set of features, yet requiring minimal training data.

Identifying tweets of personal health experience through word embedding and LSTM neural network

BackgroundAs Twitter has become an active data source for health surveillance research, it is important that efficient and effective methods are developed to identify tweets related to personal health experience. Conventional classification algorithms rely on features engineered by human domain experts, and engineering such features is a challenging task and requires much human intelligence. The resultant features may not be optimal for the classification problem, and can make it challenging for conventional classifiers to correctly predict personal experience tweets (PETs) due to the various ways to express and/or describe personal experience in tweets. In this study, we developed a method that combines word embedding and long short-term memory (LSTM) model without the need to engineer any specific features. Through word embedding, tweet texts were represented as dense vectors which in turn were fed to the LSTM neural network as sequences.ResultsStatistical analyses of the results of 10-fold cross-validations of our method and conventional methods indicate that there exist significant differences (p < 0.01) in performance measures of accuracy, precision, recall, F1-score, and ROC/AUC, demonstrating that our approach outperforms the conventional methods in identifying PETs.ConclusionWe presented an efficient and effective method of identifying health-related personal experience tweets by combining word embedding and an LSTM neural network. It is conceivable that our method can help accelerate and scale up analyzing textual data of social media for health surveillance purposes, because of no need for the laborious and costly process of engineering features.

Correlation of Tweets Mentioning Influenza Illness and Traditional Surveillance Data

ObjectiveTo determine if social media data can be used as a surveillance tool for influenza at the local level.IntroductionThe use of social media as a syndromic sentinel for diseases is an emerging field of growing relevance as the public begins to share more online, particularly in the area of influenza. Several applications have been developed to predict or monitor influenza activity using publicly posted or self-reported online data; however, few have prioritized accuracy at the local level. In 2016, the Cook County Department of Public Health (CCDPH) collected localized Twitter information to evaluate its utility as a potential influenza sentinel data source. Tweets from MMWR week 40 through MMWR week 20 indicating influenza-like illness (ILI) in our jurisdiction were collected and analyzed for correlation with traditional surveillance indicators. Social media has the potential to join other sentinels, such as emergency room and outpatient provider data, to create a more accurate and complete picture of influenza in Cook County.MethodsWe developed a JAVA program which included a customized geofence around suburban Cook County to collect tweets from Twitter’s STREAM application programming interface (API) (available at https://github.com/FoodSafeCookCo/TwitterStream-Program). The JAVA program looked for tweets within the geofence or for tweets with a profile location naming a suburban Cook County municipality and copied them to a text file if the tweet contained “flu” or “influenza”. Captured data included the user’s Twitter handle, Tweet text, Tweet time and date, x and y coordinates (if available), and profile location. Tweets were then manually reviewed to determine if they met the following criteria: 1) language indicated the user was recently ill with influenza; 2) user appeared to reside in CCDPH jurisdiction. Tweets meeting these criteria were included in the analysis. Tweets were aggregated by MMWR week and analyzed for correlation, using Pearson methods (data were normal), with two traditional surveillance sources: 1) the percent of visits to all suburban Cook County emergency departments for ILI as reported to the Cook County Electronic Surveillance System for the Early Notification of Community-based Epidemics (ESSENCE), and 2) the percent of laboratory specimens testing positive for influenza at seven local sentinel laboratories. Analysis was performed in Excel 2013 and SAS 9.4.ResultsFrom MMWR week 40-20, 113 tweets indicating influenza-like illness were collected within Cook County’s jurisdiction. Due to technical issues with the program, data were not collected for weeks 52, 2, and 17-19. Correlations were compared for the percent of laboratory specimens testing positive for influenza (LSL) and percent of visits to emergency departments for ILI (EDILI) to the total number of tweets per MMWR week. A strong correlation exists between LSL and EDILI r=0.92 (p-value&lt;0.0001) indicating the traditional sentinels have a strong positive relationship. The correlation between number of tweets and LSL was 0.46 (p-value =0.0138), indicating a moderate positive relationship. Correlation between number of tweets and EDILI was similarly moderate, r=0.52 (p-value=0.0049). Correlations to EDILI stratified by age (0-4, 5-17, 18-64, 65+) also showed a moderate positive relationship (range 0.50 to 0.55, all p-values &lt; 0.01). Twitter use peaked one week before the recorded peak of other surveillance indicators. When Twitter counts were shifted one week to align the peak in tweets with the peak of the influenza season, the correlations were 0.54 for LSL and 0.61 for EDILI (p-value=0.0034 and 0.0007, respectively).ConclusionsOverall, the tweets collected had a moderately positive relationship with the severity of influenza activity in Cook County. When the data were transitioned to match peaks, there was an increase in the correlations’ strength for both LSL and EDILI. This data indicates that publicly shared social media data may be an underutilized source of syndromic data at the local level, potentially capable of predicting seasonal influenza peaks before traditional data sources. Other jurisdictions may consider using the open source program created by CCDPH to determine the relationship of influenza related social media to their own local influenza surveillance data. For the 2017-2018 influenza season, we established a redundant system for tweet collection in case one of the systems goes down. Exploring machine learning (in place of manual review) to detect tweets indicating illness is also a promising avenue to simplify data collection and cleaning. Data will be collected using the same system for the 2017-2018 influenza season and correlations re-evaluated with more complete data.

Dynamic social media affiliations among UK politicians

Inter-personal affiliations and coalitions are an important part of politicians’ behaviour, but are often difficult to observe. Since an increasing amount of political communication now occurs online, data from online interactions may offer a new toolkit to study ties between politicians; however, the methods by which robust insights can be derived from online data require further development, especially around the dynamics of political social networks. We develop a novel method for tracking the evolution of community structures, referred to as ‘multiplex community affiliation clustering’ (MCAC), and use it to study the online social networks of Members of Parliament (MPs) and Members of the European Parliament (MEPs) in the United Kingdom. Social interaction networks are derived from social media (Twitter) communication over an eventful 17-month period spanning the UK General Election in 2015 and the UK Referendum on membership of the European Union in 2016. We find that the social network structure linking MPs and MEPs evolves over time, with distinct communities forming and re-forming, driven by party affiliations and political events. Without including any information about time in our model, we nevertheless find that the evolving social network structure shows multiple persistent and recurring states of affiliation between politicians, which align with content states derived from topic analysis of tweet text. These findings show that the dominant state of partisan segregation can be challenged by major political events, ideology, and intra-party tension that transcend party affiliations.

SNEIT: Salient Named Entity Identification in Tweets

Social media is a rich source of information and opinion, with exponential data growth rate. However social media posts are difficult to analyze since they are brief, un structured and noisy. Interestingly, many social media posts are about an entity or entities. Understanding which entity is central (Salient Entity) toa post, helps better analyze the post. In this paper we propose a model that aids in such analysis by identifying the Salient Entity in a social media post, tweets inparticular. We present a supervised machine-learning model, to identify Salient Entity in a tweet and propose that the tweet is most likely about that particular entity. We have used the premise that, when an image accompanies a text, the text most likely is about theentity in that image, to build a dataset of tweets and salient entities. We trained our model using this dataset. Note that this does not restrict the applicability of our model in any way. We use tweets with images only to obtain objective ground truth data, while features for the model are derived from tweet text. Our experiments show that the model identifies Salient Named Entity with an F-measure of 0.63. We show the effective ness of the proposed model for tweet-filtering and salience identification tasks. We have made the human annotated dataset and the source code of this model publicly available.

ANALYSIS OF THE DIFFUSION PHENOMENON OF THE TOPIC ON SOCIAL MEDIA � CASE STUDY OF JAPANESE TELEVISION DRAMA �

ABSTRACT In recent years, Social Networking Services (SNSs) which people can feel free to express their opinions are getting more and more popular. Since these services have a vast amount of information about various kinds, utilization to business and marketing strategies has a lot of attention. Therefore, there have been a diversity of studies using SNSs (such as the effects of the words of mouth on Twitter for movie sales) and the utility of using SNSs are confirmed. In this study, we aim to evaluate the relationship between topic diffusion by audience on Twitter and audience rating of TV drama. In particular, firstly, we performed the hierarchical cluster analysis by using transitions of number of Tweets to classify weeks for each TV drama episode. Secondly, we extract the text of Tweets which are diffused well and categorized its contents by using natural language processing. Through the results of these analysis, we clarified the relationship between the diffusion phenomenon and audience rating. Keywords Social Networking Service, Twitter, Natural Language Processing, Diffusion Phenomenon

SensePlace3: a geovisual framework to analyze place–time–attribute information in social media

ABSTRACTSensePlace3 (SP3) is a geovisual analytics framework and web application that supports overview + detail analysis of social media, focusing on extracting meaningful information from the Twitterverse. SP3 leverages social media related to crisis events. It differs from most existing systems by enabling an analyst to obtain place-relevant information from tweets that have implicit as well as explicit geography. Specifically, SP3 includes not just the ability to utilize the explicit geography of geolocated tweets but also analyze implicit geography by recognizing and geolocating references in both tweet text, which indicates locations tweeted about, and in Twitter profiles, which indicates locations affiliated with users. Key features of SP3 reported here include flexible search and filtering capabilities to support information foraging; an ingest, processing, and indexing pipeline that produces near real-time access for big streaming data; and a novel strategy for implementing a web-based multi-view visual interface with dynamic linking of entities across views. The SP3 system architecture was designed to support crisis management applications, but its design flexibility makes it easily adaptable to other domains. We also report on a user study that provided input to SP3 interface design and suggests next steps for effective spatiotemporal analytics using social media sources.

‘22 Push-Ups for a Cause’: Depicting the Moral Self via Social Media Campaign #Mission22

In 2016, the online cause #Mission22 went viral on social media. Established to raise awareness about high suicide rates among US military veterans, the campaign involves users posting a video of themselves doing 22 push-ups for 22 days, and on some platforms, to donate and recruit others to do the same. Based on a ‘big data’ analysis of Twitter data (over 225,883 unique tweets) during the height of the campaign, this article uses #Mission22 as a site in which to analyse how people depict, self-represent and self-tell as moral subjects using social media campaigns. In addition to spotlighting how such movements are mobilised to portray moral selves in particular ways, the analysis focuses on how a specific online cause like #Mission22 becomes popularly supported from a plethora of possible causes and how this selection and support is shaped by online networks. We speculate that part of the reason why Mission22 went ‘viral’ in the highly competitive attention economies of social media environments was related to visual depictions of affective bodily, fitness and moral practices.

Machine Learning, Sentiment Analysis, and Tweets: An Examination of Alzheimer's Disease Stigma on Twitter.

Social scientists need practical methods for harnessing large, publicly available datasets that inform the social context of aging. We describe our development of a semi-automated text coding method and use a content analysis of Alzheimer's disease (AD) and dementia portrayal on Twitter to demonstrate its use. The approach improves feasibility of examining large publicly available datasets. Machine learning techniques modeled stigmatization expressed in 31,150 AD-related tweets collected via Twitter's search API based on 9 AD-related keywords. Two researchers manually coded 311 random tweets on 6 dimensions. This input from 1% of the dataset was used to train a classifier against the tweet text and code the remaining 99% of the dataset. Our automated process identified that 21.13% of the AD-related tweets used AD-related keywords to perpetuate public stigma, which could impact stereotypes and negative expectations for individuals with the disease and increase "excess disability". This technique could be applied to questions in social gerontology related to how social media outlets reflect and shape attitudes bearing on other developmental outcomes. Recommendations for the collection and analysis of large Twitter datasets are discussed.

TWITTER DATA MINING FOR SENTIMENT ANALYSIS ON PEOPLES FEEDBACK AGAINST GOVERNMENT PUBLIC POLICY

Government policies often get positive or negative response from the public. The response from the community feedback can be conveyed through print and electronic media. With the rise of social media today, people have a tendency to convey such feedback through social media such as Facebook, Twitter, Instagram, Path and other social media. Thus, to determine the public response to this policy that has been implemented, the government needs to know how your feedback from people who come from social media. But because of the feedback, it is difficult to detect how many positive or negative response from the public. Therefore, in this study will develop a system to obtain data in the form of feedback coming from one of the social media that is often used by the public, namely Twitter. Tweet or post on the community will be collected based on the time and place specified. Having obtained a collection tweet, would do next text preprocessing stage. Tweet text already passed the stage of preprocessing, for further processing in sentiment analysis, to determine the positive and negative responses from the public against government policies that have been applied.

Classification of news-related tweets

It is important to obtain public opinion about a news article. Microblogs such as Twitter are popular and an important medium for people to share ideas. An important portion of tweets are related to news or events. Our aim is to find tweets about newspaper reports and measure the popularity of these reports on Twitter. However, it is a challenging task to match informal and very short tweets with formal news reports. In this study, we formulate this problem as a supervised classification task. We propose to form a training set using tweets containing a link to the news and the content of the same news article. We preprocess tweets by removing unnecessary words and symbols and apply stemming by means of morphological analysers. We apply binary classifiers and anomaly detection to this task. We also propose a textual similarity-based approach. We observed that preprocessing of tweets increases accuracy. The textual similarity method obtains results with the highest recognition rate. Success increases in some cases when report text is used with tweets containing a link to the news report within the training set of classification studies. We propose that this study, which is made directly in consideration of tweet texts that measure the trends of national newspaper reports on social media, has a higher significance when compared to Twitter analyses made by using a hashtag. Given the limited number of scientific studies on Turkish tweets, this study makes a contribution to the literature.

Image-sharing in Twitter-based professional conversations

This paper reports on ongoing research into the image-sharing practices of two informal professional networks, one dedicated to midwives and the other dedicated to teachers, on Twitter. Each network is brought together through regular, loosely synchronous Twitter conversations, created through the use of an identifying hashtag. In both cases, the conversations have been initiated by practitioners with the explicit intention of creating a space for sharing ideas, practice, experience and opinions. Community members are relied on to provide facilitation, to promote the conversations and, in one case, to suggest conversations themes. These kinds of informal professional conversation may thus serve as somewhat hidden, but potentially influential, sites of professional learning. As with most social media and mobile communications, Twitter has become increasingly saturated with images. Indeed this may be a particularly strong trend on Twitter because images can be used to convey a great deal more than might easily be said in the 140 characters users are limited to in each tweet. Professional conversations are no exception to this, with images accompanying tweets with increasing frequency. Images, then, may provide a rich alternative to research focusing on the text of tweets. This paper describes research using such images as foci for the study of the flow of information, opinion and affect amongst the two professional groups described above. Starting with a period of extended observation of both conversations, the research proceeded to in-depth interviews with a small number of practitioners who were also active participants, in which images previously published during the Twitter conversations were used as prompts. Methods developed to visualise and make sense of the relations between images and users during the conversations, and ultimately to identify candidate interviewees and prompt images are described. The research makes use of philosopher Gilles Deleuze's concepts of lines of articulation, lines of flight and knots to draw out some of the factors affecting flows through the conversation-spaces. It appears that there are at least three broad factors - the technical affordances for communication provided by Twitter, individuals' notions of online professionalism, and individuals' sense of the purpose of the conversation-space - that create lines of articulation and flight, constraints and uncertainties that accelerate, amplify or impede these flows. These lines twist and knot, resulting in a socio-technical disciplining that conditions the conversations and both reveals and hides aspects of professional life.

Risk analysis and visualization for detecting signs of flood disaster in Twitter

Disasters resulting from climate changes and urbanized complex cities are making it more and more difficult to predict disaster occurrences and their effects. Therefore, detecting signs of disaster can play an important role in minimizing damages and preventing secondary damages through immediate response. Meanwhile, the amount of unstructured data such as SNS data is increasing and efforts to analyze useful information from SNS data are actively being made in various academic areas. In the disaster management area particularly, SNS data can be used as a real-time sensor and premonitory signs showing the possibility of disaster occurrence since the amount of SNS data increases even before the occurrence of disaster. Accordingly, the purpose of this study is to differentiate the risk level of Twitter data and visualize on the map in order to detect signs of disaster as soon as possible. The risk level of Twitter data is a quantified indicator that shows possibility of disaster occurrence and estimated from ‘level of risk expression of tweet text’ and ‘flood vulnerability of tweet location’. As a result, 270,492 tweets that contains signs of flood were extracted and 7090 tweets occurred in Seoul were extracted among 270,492 flooding sign tweets. Second, two indicators of tweet risk level assessment were defined and final risk level rating was assigned based on the risk evaluation matrix. Third, analyzed Twitter data was visualized on time-series map and their applicability was verified based on the flood occurred near Gangnam Station on July 27, 2011. The result shows that tweets about flooding started 2 h prior to when the incident was reported for the first time.

From Twitter to detector: Real-time traffic incident detection using social media data

The effectiveness of traditional incident detection is often limited by sparse sensor coverage, and reporting incidents to emergency response systems is labor-intensive. We propose to mine tweet texts to extract incident information on both highways and arterials as an efficient and cost-effective alternative to existing data sources. This paper presents a methodology to crawl, process and filter tweets that are accessible by the public for free. Tweets are acquired from Twitter using the REST API in real time. The process of adaptive data acquisition establishes a dictionary of important keywords and their combinations that can imply traffic incidents (TI). A tweet is then mapped into a high dimensional binary vector in a feature space formed by the dictionary, and classified into either TI related or not. All the TI tweets are then geocoded to determine their locations, and further classified into one of the five incident categories.We apply the methodology in two regions, the Pittsburgh and Philadelphia Metropolitan Areas. Overall, mining tweets holds great potentials to complement existing traffic incident data in a very cheap way. A small sample of tweets acquired from the Twitter API cover most of the incidents reported in the existing data set, and additional incidents can be identified through analyzing tweets text. Twitter also provides ample additional information with a reasonable coverage on arterials. A tweet that is related to TI and geocodable accounts for approximately 5% of all the acquired tweets. Of those geocodable TI tweets, 60–70% are posted by influential users (IU), namely public Twitter accounts mostly owned by public agencies and media, while the rest is contributed by individual users. There is more incident information provided by Twitter on weekends than on weekdays. Within the same day, both individuals and IUs tend to report incidents more frequently during the day time than at night, especially during traffic peak hours. Individual tweets are more likely to report incidents near the center of a city, and the volume of information significantly decays outwards from the center.

Utilising Wikipedia for Text Mining Applications

The process whereby inferences are made from textual data is broadly referred to as text mining. In order to ensure the quality and effectiveness of the derived inferences, several approaches have been proposed for different text mining applications. Among these applications, classifying a piece of text into pre-defined classes through the utilisation of training data falls into supervised approaches while arranging related documents or terms into clusters falls into unsupervised approaches. In both these approaches, processing is undertaken at the level of documents to make sense of text within those documents. Recent research efforts have begun exploring the role of knowledge bases in solving the various problems that arise in the domain of text mining. Of all the knowledge bases, Wikipedia on account of being one of the largest human-curated, online encyclopaedia has proven to be one of the most valuable resources in dealing with various problems in the domain of text mining. However, previous Wikipedia-based research efforts have not taken both Wikipedia categories and Wikipedia articles together as a source of information. This thesis serves as a first step in eliminating this gap and throughout the contributions made in this thesis, we have shown the effectiveness of Wikipedia category-article structure for various text mining tasks. Wikipedia categories are organized in a taxonomical manner serving as semantic tags for Wikipedia articles and this provides a strong abstraction and expressive mode of knowledge representation. In this thesis, we explore the effectiveness of this mode of Wikipedia's expression (i.e., the category-article structure) via its application in the domains of text classification, subjectivity analysis (via a notion of "perspective" in news search), and keyword extraction. First, we show the effectiveness of exploiting Wikipedia for two classification tasks i.e., 1- classifying the tweets1 being relevant/irrelevant to an entity or brand, 2- classifying the tweets into different topical dimensions such as tweets related with workplace, innovation, etc. To do so, we define the notion of relatedness between the text in tweet and the information embedded within the Wikipedia category-article structure. Then, we present an application in the area of news search by using the same notion of relatedness to show more information related to each search result highlighting the amount perspective or subjective bias in each returned result towards a certain opinion, topical drift, etc. Finally, we present a keyword extraction strategy using community detection over the Wikipedia categories to discover related keywords arranged in different communities. The relationship between Wikipedia categories and articles is explored via a textual phrase matching framework whereby the starting point is textual phrases that match Wikipedia articles' titles/redirects. The Wikipedia articles for which a match occurs are then utilised by extraction of their associated categories, and these Wikipedia categories are used to derive various structural measures such as those relating to taxonomical depth and Wikipedia articles they contain. These measures are utilised in our proposed text classification, subjectivity analysis, and keyword extraction framework and the performance is analysed via extensive experimental evaluations. These experimental evaluations undertake comparisons with standard text mining approaches in the literature and our Wikipedia framework based on its category-article structure outperforms the standard text mining techniques.

Geocoding of Spatial Relationships Contained in Tweets

Human ability to understand approximate references to locations, disambiguated by means of context and reasoning about spatial relationships, is the key to describe spatial environments and to share information about them. In this paper, the authors propose an approach for geocoding that takes advantage of the spatial relationships contained in the text of tweets, using semantic and spatial analyses. Microblog text has special characteristics (e.g. slang, abbreviations, acronyms, etc.) and thus represents a special variation of natural language. The main objective of this work is to associate spatial relationships found in text with a spatial footprint, to determine the location of the event described in the tweet. The feasibility of the proposal is demostrated using a corpus of 200,000 tweets posted in Spanish related with traffic events in Mexico City.

Framework for Participative and Collaborative Governance using Social Media Mining Techniques

Social media and networking sites have broken the barriers in communication and brought about a revolution in information access, dissemination and communication. Given the inclusive and collaborative nature of social media, political leaders and government organizations too have attempted to harness it for building support, gauging popularity and analyzing the opinions of citizens. This paper, proposes a framework which focuses on the application of text mining techniques on social media content for achieving participative, collaborative and inclusive governance. Most of the political entities and government departments have twitter accounts. Though the tweet text is limited to 140 characters, the accompanying metadata and hashtag (#tag) embedded content make it a potent source for extracting entities, concepts and topics. Web Crawling for content pertaining to these entities and compilation of the same, results in a corpus suitable for text mining. Applying Natural Language Processing (NLP) and mining techniques like Part-of-Speech (POS) tagging, classification, clustering to this generated corpus enables categorizing and summarizing: a) sentiments b) queries and c) grievances. Summarized view of the content makes it amenable for decision making and formulation of responses. A Knowledgebase (KB) is created and perpetually updated with the concepts, entities, summaries, queries and responses. The continued process of creation and enhancement contributes to machine learning. The scope of this paper is limited to proposing a conceptual framework for social media mining facilitating informed decision making for better governance.

TASC:Topic-Adaptive Sentiment Classification on Dynamic Tweets

Sentiment classification is a topic-sensitive task, i.e., a classifier trained from one topic will perform worse on another. This is especially a problem for the tweets sentiment analysis. Since the topics in Twitter are very diverse, it is impossible to train a universal classifier for all topics. Moreover, compared to product review, Twitter lacks data labeling and a rating mechanism to acquire sentiment labels. The extremely sparse text of tweets also brings down the performance of a sentiment classifier. In this paper, we propose a semi-supervised topic-adaptive sentiment classification (TASC) model, which starts with a classifier built on common features and mixed labeled data from various topics. It minimizes the hinge loss to adapt to unlabeled data and features including topic-related sentiment words, authors’ sentiments and sentiment connections derived from “@” mentions of tweets, named as topic-adaptive features. Text and non-text features are extracted and naturally split into two views for co-training. The TASC learning algorithm updates topic-adaptive features based on the collaborative selection of unlabeled data, which in turn helps to select more reliable tweets to boost the performance. We also design the adapting model along a timeline (TASC-t) for dynamic tweets. An experiment on 6 topics from published tweet corpuses demonstrates that TASC outperforms other well-known supervised and ensemble classifiers. It also beats those semi-supervised learning methods without feature adaption. Meanwhile, TASC-t can also achieve impressive accuracy and F-score. Finally, with timeline visualization of “river” graph, people can intuitively grasp the ups and downs of sentiments’ evolvement, and the intensity by color gradation.

타임라인데이터를 이용한 트위터 사용자의 거주 지역 유추방법

SNS사용자의 거주 지역을 유추하여 그들이 생성한 데이터에 거주위치를 부여하는 것은 위치희박(location sparsity)과 생태학적 오류문제로 인해 연구결과의 신뢰성이 떨어진다는 평가를 받아온 공간빅데이터 연구에 대안이 될 수 있다. 본 연구에서는 Tweet 사용자의 거주 지역을 유추하는 방법으로 사용자 타임라인데이터 속에서 찾아낸 일상생활활동패턴을 이용하는 방법을 고안하였다. 트윗 사용자의 일상생활활동패턴은 이동궤적과 사용자의 언어(text)에서 확인할 수 있었으며 전자를 활용한 모델을 일상이동패턴모델, 후자를 활용한 모델을 일상 활동장 모델이라 명명하고 각각 모델에 입력될 변수를 선정하였다. 자신의 거주 지역에서 가장 높은 빈도의 트윗 발생 여부와 가장 높은 빈도의 거주행정구역 표현 단어를 사용하는지 아닌지를 종속변수로 한 판별분석을 실시하여 모델을 작성하였으며 설명력은 일상 이동패턴모델, 일상 활동장 모델 각각 67.5%, 57.5%였다. 이 모델을 스트레스 관련 트윗을 작성한 사용자의 타임라인데이터로 구성된 테스트데이터에 입력해본 결과 전체 사용자 48,235명 중 5,301명의 거주 지역을 유추하였고 이를 활용하여 위치 부여된 스트레스 관련 트윗 9,606개를 확보하였다. 본 연구의 유추기법을 통해 기존 SNS데이터 분석연구에서 사용하는 데이터 수집 방법보다 44배 많은 위치 부여 트윗을 확보할 수 있었다. 본 연구방법론은 SNS데이터를 이용한 연구에서 위치 부여된 데이터를 확보하는데 활용 가능할 것으로 판단되며, 각종 지역통계와 상관관계파악을 통해 지역적 현상 분석에도 SNS데이터를 이용할 수 있는 가능성을 높일 것으로 판단된다. If one can infer the residential area of SNS users by analyzing the SNS big data, it can be an alternative by replacing the spatial big data researches which result from the location sparsity and ecological error. In this study, we developed the way of utilizing the daily life activity pattern, which can be found from timeline data of tweet users, to infer the residential areas of tweet users. We recognized the daily life activity pattern of tweet users from user's movement pattern and the regional cognition words that users text in tweet. The models based on user's movement and text are named as the daily movement pattern model and the daily activity field model, respectively. And then we selected the variables which are going to be utilized in each model. We defined the dependent variables as 0, if the residential areas that users tweet mainly are their home location(HL) and as 1, vice versa. According to our results, performed by the discriminant analysis, the hit ratio of the two models was 67.5%, 57.5% respectively. We tested both models by using the timeline data of the stress-related tweets. As a result, we inferred the residential areas of 5,301 users out of 48,235 users and could obtain 9,606 stress-related tweets with residential area. The results shows about 44 times increase by comparing to the geo-tagged tweets counts. We think that the methodology we have used in this study can be used not only to secure more location data in the study of SNS big data, but also to link the SNS big data with regional statistics in order to analyze the regional phenomenon.