More than words-Social networks’ text mining for consumer brand sentiments.pdf

《More than words-Social networks’ text mining for consumer brand sentiments.pdf》由会员分享，可在线阅读，更多相关《More than words-Social networks’ text mining for consumer brand sentiments.pdf（11页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、More than words: Social networks text mining for consumer brand sentiments Mohamed M. Mostafa Instituto Universitrio de Lisboa, Business Research Unit, Avenida das Foras Armadas, Lisbon, Portugal a r t i c l ei n f o Keywords: Consumer behavior Global brands Sentiment analysis Text mining Twitter a

2、b s t r a c t Blogs and social networks have recently become a valuable resource for mining sentiments in fi elds as diverse as customer relationship management, public opinion tracking and text fi ltering. In fact knowl- edge obtained from social networks such as Twitter and Facebook has been shown

3、 to be extremely valu- able to marketing research companies, public opinion organizations and other text mining entities. However, Web texts have been classifi ed as noisy as they represent considerable problems both at the lexical and the syntactic levels. In this research we used a random sample o

4、f 3516 tweets to evaluate con- sumers sentiment towards well-known brands such as Nokia, T-Mobile, IBM, KLM and DHL. We used an expert-predefi ned lexicon including around 6800 seed adjectives with known orientation to conduct the analysis. Our results indicate a generally positive consumer sentimen

5、t towards several famous brands. By using both a qualitative and quantitative methodology to analyze brands tweets, this study adds breadth and depth to the debate over attitudes towards cosmopolitan brands. ? 2013 Elsevier Ltd. All rights reserved. 1. Introduction Opinions expressed in social netwo

6、rks play a major role in infl u- encing public opinions behavior across areas as diverse as buying products, capturing the pulse of stock markets and voting for the president (Bai, 2011; Eirinaki, Pisal, Leong, Lee, and RQ2. Can companies effectively use the blogosphere to redesign their marketing a

7、nd advertising campaigns? This paper is organized as follows. Next section provides a brief literature review on the major areas of SA applications. Section 3 deals with the method used to conduct the analysis. In this section issues related to research design and, sampling and data analysis techniq

8、ues are presented. In Section 4, the results of sentiment analysis are presented. Finally, Section 5 presents research implica- tions and limitations. This section also explores avenues for future research. 2. Literature review SA techniques have been recently utilized in applications such as extrac

9、ting suggestions from consumers product reviews (e.g., Vishwanath Papacharissi (1) for buy; (0) for hold; (?1) for sell; and (?2) for strong sell. In this study the authors also used a weighting scheme to assign weights for each sentiment obtained based on the reputation and previous accuracy of the

10、 poster. Using a simulated environment to mimic real trading, the authors reported around four percent increase in returns over one month based on sentiment analysis. Bollen, Mao, and Zeng (2011) found that the aggregation of mil- lions of tweets posted daily on Twitter can be used to predict stock

11、market over time. The authors used measures such as daily Twitter posts over around ten months to predict the Dow Jones Industrial average closing values. To cross-validate the results, the authors also used the resulting time series of Twitter moods to detect the general publics response towards th

12、e outcome of the US presiden- tial campaign. Other studies investigated the relationship between investors sentiments and other factors such as stock returns fol- lowing a major earthquake (Shan & Gong, 2012), air disasters involving US vs. foreign airlines (Kaplanski & Levy, 2010) and local sports

13、events (Chang, Chen, Chou, & Lin, 2012). 3. Method 3.1. Twitter sampling Twitter is a microblogging service that was launched formally on July 13, 2006. Unlike other social media, Twitter is considered a microblog because its central activity revolves around posting short updates or tweets using the

14、 Web or mobile/cell phones. The maximum size of the blog is 140 characters-roughly the size of a newspaper headline. According to S (2012), a marketing research company, there are now around 500 million active twitters. Fig. 2 shows top ranked countries according to ac- tive tweets in 2012 (S, 2012)

15、. Tweets are available publicly as a default, and are also directly broadcasted to the users followers (Bliss, Klouman, Harris, Danforth, & Dodds, 2012). A recent analysis of Twitter activities found that more than 80% of the users either update their followers on what they actually doing or dissemi

16、nate information regarding their daily experiences (Thelwall, Buckley, & Paltoglou, 2011). Since Twitter is the most large, popular and well-known microblog Web site, it was selected to conduct the analysis reported in this study. The data used repre- sent a random set of Twitter posts from July 18,

17、 2012, to August 17, 2012. The data comprised 3516 tweets for sixteen brands. To guar- antee representativeness, sample selection has been varied by day of the week and hours in the day. Our sample size is comparable in size to Qiu, He, Zhang, Shi, Bu, and Chens (2010) sample, which in- cluded 3783

18、opinion sentences. Table 1 shows the random sample of tweets for each brand included in the study. Following Thelwall et al. (2011), only tweets in English was chosen in order to remove complications that might arise with analyzing multilingual tweets. Table 2 shows a sample of tweets for Air India

19、with a manual classifi cation of customers sentiments. As can be seen from the ta- ble, tweets represent a very noisy environment in which messages posted to virtual audience includes abbreviated words, the and the hashtag (#) characters, and heteroglossia-referring to other voices in the tweets in

20、order to convey interpersonal and ideational Fig. 3. Proximity plot based on Egypt Air tweets. M.M. Mostafa/Expert Systems with Applications 40 (2013) 424142514245 meanings (Bakhtin, 1981). Huang, Thornton, and Efthimiadis (2010) found that the hashtag was invented by Twitter users early in 2008 to

21、help followers fi nd a specifi c tweet or post. As opposed to the hashtag, the character has been introduced to address a tweet to another follower, which allows Twitter to function effec- tively as a collaboration and conversation system (Honeycutt & Herring, 2009). 3.2. Lexicon Categorizing words

22、for SA is a major step in applying the tech- nique. Broadly speaking, there are two widely used methods for sentiment orientation identifi cation: the lexicon-based approach and the corpus-based method (Miao, Li, & Zeng, 2010). However, since the corpus-based method has rarely been used to analyze s

23、entiment orientation, we will focus here on the lexicon-based method. Nevertheless, both methods require a pre-defi ned dictio- nary or corpus of subjective words. The sentiment is determined by comparing tweets against the expert-defi ned entry in the dic- tionary, which makes it easy to determine

24、the polarity of a specifi c sentence. Thus, it is crucial to have an accurate classifi er to be used to construct indicators of sentiment. Previous research has typi- cally incorporated lexicons such as the manually coded General In- quirer (Stone, Dunphy, Smith, & Ogilvie, 1966), which includes ove

25、r 11000 hand-coded word stems in 182 categories, the LIWC dictio- nary (Pennebaker, Mehl, & Niederhoffer, 2003), the SentiWordNet (Baccianella, Esuli, & Sebastiani, 2010), the Q-WordNet (Agerri & Garcia-Serrano, 2010) or the lexicon of subjectivity clues (Wiebe, Wilson, Bruce, Bell, & Martin, 2004).

26、 Automatically-coded lexicons have recently been developed, including the sentiment-based lex- icon (Taboada, Brooke, Tofi loski, Voll, & Stede, 2011). In this study we used the Hu and Liu (2004) lexicon to conduct the analysis because this dictionary has been used successfully in a similar applicat

27、ion (Miner et al., 2012). This lexicon includes around 6800 seed adjectives with known orientation (2006 posi- tive words and 4783 negative words). The lexicon has recently been updated by adding words based on a thorough search in the WordNet. The lexicon size is similar to the recently used Opin-

28、ion Finder lexicon, which included 2718 positive words and 4912 negative words (Bollen et al., 2011). Our lexicon is also comparable to the dictionary used by Hu, Bose, Koh, and Liu (2012), which in- cludes 1635 positive words and 2005 negative words. Modifi ca- tions on this approach include the ha

29、ndling of negation (Das & Chen, 2001) and the weight of enhancer/intensifi er words such as the use of the really or absolutely (Turney, 2002). Other at- tempts have been recently tried to develop lexicons capable of detecting deep sentiments expressed in discussions among several actors in informal

30、 situations (Maks & Vossen, 2012). Similarly, Reyes and Rosso (2012) developed a corpus to detect irony in con- sumers reviews. Fig. 4. A 3-D map with link strengths and base lines shown. 4246M.M. Mostafa/Expert Systems with Applications 40 (2013) 42414251 4. Results 4.1. Exploratory data analysis a

31、nd visualizations To conduct the qualitative part of this study we used QDA Miner 4.0 software package (Provalis Research, 2011) for coding textual data posted on Twitter. This software was selected because of its extensive exploratory tools that can be used to identify hidden pat- terns in textual

32、data. In order to analyze consumer sentiments to- wards brands, we started by generating relative frequency word counts. Table 3 shows the percentage of words in a random set of tweets. From Table 3 we can see that words such as global, fl ight and price have the highest frequency for a brand such a

33、s Egypt Air. However, references were also made in the tweets to countries such as Syria, probably because of the ongoing uprising in that country. Analyzing frequency of appearance or simply the incre- mental count of appearance of particular words or phrases might provide insights into a particula

34、r topic. In fact OLeary (2011) ar- gues that despite the simplicity of such approach, it can be used to predict characteristics of the topic analyzed. Fig. 3 shows a prox- imity plot constructed based on Egypt Air tweets. This fi gure shows visually, on a single axis, the distance from a particular

35、object to all other objects. This graph was constructed to extract huge amount of data based on a distance matrix. From this graph we see that most tweets were concerned with things like price, tickets and monopoly. However, Twitters seem to be also concerned with fi ghting in neighboring Syria. Fig

36、. 4 shows a 3-D concept map constructed based on multidi- mensional scaling (MDS) technique. In this graph the closer the cases, the higher the tendency of co-occurrence and vice versa. The lines on the map represent levels of association among words. From the graph we can reconstruct the most infl

37、uential tweets. For example, for Egypt Air, we can see the following pattern: closeness to home as represented by words such as Egypt, Dubai and Resident, personifi cation as represented by names of people tweeting, and relevance and signifi cance as represented by words such as Dreamliner, Service,

38、 and Flight. This result is in line with a recent study using centering resonance analysis, a computa- tional discourse analysis technique, on a random sample of 9000 Egyptian tweets (Papacharissi & Oliveira, 2012). 4.2. Overall sentiment scores We used the twitteR, the plyr, stringr and the ggplot2

39、 libraries in the R software package version 2.15 to conduct the quantitative sentiment score. Fig. 5 shows the distribution of sentiment scores obtained for Nokia and Pfi zer (similar graphs were constructed for remaining brands). From the graph, we immediately recognize some asymmetry. For example

40、, the bars at +1 are much larger for Nokia compared to the +1 bars for Pfi zer. It is also evident that the bars at ?1 are much larger for Pfi zer compared to the ?1 bars for Nokia. This makes it clear that the overall sentiment score for Nokia is generally better than the sentiment score for Pfi ze

41、r. The visualization of the sentiment distribution in Fig. 5 further underlines the fact that most tweets fall either on the neutral point (0) or within the band of circa ?1/+1, which is an indication that several tweets are not very affective. Although this result is in line with Lindgren (2012), w

42、e can focus only on positive or negative sentiments. Following Miner et al. (2012), we ignored the middle and constructed sentiment scores for a random sample of only po- sitive and negative sentiments. Fig. 6 shows results for only three brands. From Fig. 6, we clearly see that most tweet comments

43、were positive for both Lufthansa and DHL. However, most of the tweets were negative for T-Mobile. Finally, we used the StreamGraph software package (Clark, 2008) to visualize the trend of tweets across a period of time for all brands. Fig. 7 shows tweets trend for three brands: Nokia, Pfi zer and Al

44、-Jazeera English. This graph is useful since it represents mul- tiple time series data stacked one on top of the other (Havre, Het- zler, Whitney, & Nowell, 2002). Since the total frequency of all features represents the height of the curve, each time series data should be read off the fi gure not a

45、s the cumulative height but rather as starting with zero. As can be seen the graph is character- ized by a number of spikes, indicating an increase in tweets fre- quency at those particular times. Interestingly, we see spikes on the top part of the fi gure representing Nokias introduction of the new

46、 smart phone Nokia Lumia brand. 5. Implications, limitations and future research In this study we analyzed sentiment polarity of more than 3500 social media tweets expressing attitudes towards sixteen global brands. Social media users represent 67% of around a bil- lion Internet active users (Eirina

47、ki et al., 2012). Although a single tweet is limited to 140 characters in length, the millions of tweets posted on Twitter almost on a daily basis might provide an unbi- ased representation of consumers sentiment towards services and brands. Capturing consumers opinions and gaining knowl- edge about

48、 consumer preferences has long been a major concern for marketing researchers. However, traditional marketing meth- ods such as focus groups and face-to-face interviewing are both costly and time consuming. In contrast, tweets and blogs are readily available for free. Such consumer generated media a

49、re also free of bias that might be introduced by the interviewer in case of Fig. 5. Sentiment scores for Nokia (top) and Pfi zer (bottom). X-axis represents score distributions, Y-axis represents count/frequencies. M.M. Mostafa/Expert Systems with Applications 40 (2013) 424142514247 personal interviews. Moreover, consumers opinion-based ex- pressed qualitatively may easily be benchmarked against objec- tive measures such as sales data, revenues, or stock price. Thus, companies can utilize such online textual content in an effort to gain insight into consumers opinions regarding available