Fabio Taddei Dalla Torre 15/7/2020
Artificial intelligence (AI) refers to systems that display intelligent behavior by analyzing their environment and taking actions – with some degree of autonomy – to achieve specific goals. This is the definition of artificial intelligence given by the European Commission in 2018 [1].
Many are the possible advantages of this technology one recent study, for example, related to the COVID-19 pandemic, has proved that models built using machine learning shown high sensitivity and high specificity in detecting COVID-19 by analyzing chest CT exams [2]. Moreover must be said that Artificial Intelligence is already well present in everybody daily life. Every smartphone has some type of assistant for example “Siri” or “Google Assistant” that displays intelligence like behavior.
Although this theme leads inevitably to some problem. “Ex Machina”, “Trascendence or”Io Robot“, are only three movie titles where movie industry has shown a World where human life is threatened by self conscious machine engineered by human themselves. Those are irrelevant but in 2014 Stephen Hawking told in an interview by the BBC:”The development of full artificial intelligence could spell the end of the human race.", a clear warning by one of the most intelligent people of this century.
This work is composed by five main sections. The second one deals with the formulation of the research question and will provide a brief literature review on previous studies. Section three will describe the methodology used in order to answer to the research question, how data are collected, cleaned and the techniques used. Section four will concern the analysis of the data, it contains the codes and comments about the process and the results. The final section gives a summary of the results, also presenting the main advantages and disadvantages of the techniques and processes used.
From the invention of the World Wide Web until today, this concept has changed and evolved. Among the many implementations that this technology has allowed, the development of social media platforms capable of fast and free communication on a global scale has definitely changed our lives. It is estimated that, in 2020, the global social penetration reach has reached 49%, with a peak considering Asia and North America that respectively scored a rate of 71% and 69%. Northern Europe fallow with a slightly lower 67% [3]. From those statistics is easy to understand the gold mine that social media represent from an opinion mining point of view.
The aim of this work is to:
- Gain a general understanding about the opinion of people, what they think about artificial intelligence. Therefore it aims to understand which are the feeling and the sentiment of people, are they excited or do they fear this technology?
- Understand whether different conversation topics about this subject are present. Maybe there would be a group of people that promote AI and so they will talk about certain subject while others will be more skeptical writing about something else.
- Investigate what are the most common words in the negative sentiment tweets and in the positive in order to understand if there is a potential correlation between these terms and the emotions.
This analysis is carried out by using Tweets, firstly because it guarantees open data despite other platforms and even because of the large amount of available content. In 2012 it was calculated that registered users produced a total of 340 million tweets per day [4]. People do not use Twitter only for sharing personal content but also for expressing their opinion. Because of this behavior by analyzing this tweets, first with more qualitative technique and then with a sentiment analysis is possible to understand the perception of users to a certain topic [5].
From business and marketing purpose to politics, opinion mining is becoming more and more important in order to understand people thoughts, hence for building better strategies [6]. A pivotal factor in opinion mining is sentiment analysis, in this case the technique tries to understand the sentiment of the writer by judging the document. The object of the analysis is to assign for each document (in this case each tweet) a predefined sentiment (positive, negative…), this classification is based on machine learning and lexicon based approaches [7]. One point that needs attention when working with Twitter data is the difference between standard texts used in building sentiment analysis algorithms and tweets. First of all because of the length of those, 140 character maximum, and second because of their nature tweets tend to have misspelling and slang with a relative high rate. [8]
Many studies have been conducted using sentiment analysis and topic modelling on tweets. Indeed, in [9] S.Yang and H.Zhang have used sentiment analysis and LDA topic modelling in order to understand the potential of these techniques on twitter data. S. Yang et al have proven that LDA allows to easily analyze huge number of tweets and to identify hidden topics. The sentiment analysis was carried on using the syuzhet package and the results have proved that this is a powerful tool for understanding and showing the sentiments and emotions present on tweets. A. Alamsyah [10] et al, have shown that using LDA and sentiment analysis leads to the capability of extracting meaningful insights about a specific topic using large scale datasets. Compared with traditional method they have stated that those techniques allow better in-time processing. Understand people reaction and sentiment about the COVID epidemic was done by another research. Also in this case the analysis was carried out by performing a sentiment analysis, with the syuzhet package, and an LDA topic modelling. [11]
In order to accomplish my task I use the Twitter standard API with the package twitteR. I request 7000 tweets by the hashtag #AI.
Cleaning the tweets is the first step of the analysis. This process is done by working on a database created with the requested tweets. At the end of this procedure I get two different text corpus and a cleaned data-set with three column: one containing the text, one the creation date and one with the unique tweet ID. I have decided to create a data-set for further analysis and for consistency of the results, indeed all the code can be runned by taking the information from the df.ai.csv.
Furthermore, only English tweets are gathered with the exclusion of retweet. This because, from one side the sentiment classifier works only with English vocabulary and on the other side because high frequency retweets could distort the percentages of word occurrence.
As a consequent step I proceed by plotting a word cloud in order to give a visual representation of the words that occur the most among all the tweets. To strengthen this I decided to plot a bar-plot with words that have been registered more than 150 times. This because the bar-plot allows to have a closer idea of the importance of a word.
The previous analysis gives a general idea of the most important words but it gives no insight on how these terms are used together. For this exploration we can use a Word Network, in this case words are linked together according to their presence with respect to one another.
The next steps regard Topic Modelling. The aim of this technique is to present the text as a set of topics and so to find different topics among the tweets. Once topics are found the tweets are labeled according to that, for further analysis. This procedure is done by using LDA technique. LDA topic modelling is based on a probabilistic model approach and in particular it is based on the assumption that documents are a mixture of topic and that each topics is composed by a certain probability distribution of words.
The last step regard a sentiment analysis of tweets. This is done by using the package syuzhet because it allows to label tweets not only as positive or negative but also with a range of other emotions (anger, anticipation, disgust, fear, joy, negative, positive, sadness, surprise, trust). The algorithm works by the NRC emotion-Lexicon that is a list of English words associated with emotions and sentiments. At first the result is displayed cumulatively in order to have a general overview of people’s sentiment and then, after normalizing the results, a grouped bar chart is be produced that allows the comparison of the sentiment according to the topic. Moreover, understanding what makes a tweet negative or positive is done by computing a binomial sentiment analysis and by displaying the most occurrenced words for negative and positive sentiment.
#Library
library(twitteR)
library(wordcloud)
library(topicmodels)
library(dplyr)
library(stringr)
library(tm)
library(ggplot2)
library(tidytext)
library(tidyverse)
library(igraph)
library(ggraph)
library(widyr)
library(quanteda)
library(syuzhet)
library(data.table)# key and token
consumer_key <- XXXXXX
consumer_secret <- XXXXXX
access_token <- XXXXXX
access_secret <- XXXXXXGetting twitter access
# connecting to Twitter
setup_twitter_oauth(consumer_key,
consumer_secret,
access_token,
access_secret)Getting tweets
# requiring the tweets
tweets <- searchTwitter("#AI -filter:retweets", n=7000, lang = "en") # requiring tweets
length(tweets)# data preparation
df.ai <- bind_rows(lapply(tweets, as.data.frame))
df.ai$text <- gsub("((?:\\b\\W*@\\w+)+)", "", df.ai$text) # Remove usernames
df.ai$text <- gsub("http.+ |http.+$", "", df.ai$text) # removing links
df.ai <- df.ai[!duplicated(df.ai[, 'id']),] # removing duplicates
df.ai$text <- gsub("[^[:graph:]]", " ", df.ai$text) # removing graphical
df.ai$text <- gsub("[[:punct:]]", " ", df.ai$text) # remoove punctuation
df.ai$text <- gsub("^ ", "", df.ai$text) # remove spaces at the beginning
df.ai$text <- gsub(" $", "", df.ai$text) # remove spaces at the end
df.ai$text <- tolower(df.ai$text) # transform the whole text to lower
df.ai$text <- removeWords(df.ai$text, stopwords('en')) # removing stopwords
df.ai$text <- gsub("[ |\t]{2,}", " ", df.ai$text) # removing tabs
df.ai$text <- gsub(" +", " ", df.ai$text) # removing spaces
df.ai$text <- iconv(df.ai$text, to = "ASCII", sub = " ") # converting to ascii
df.ai$text <- gsub('[0-9]+', "", df.ai$text) # removing numbers
df.ai$text <- removeWords(df.ai$text, c("9", "10", "18","19", "5", "0", "46", "1", "2", "’s", "-", "amp")) # removing certain charAt the end of this cleaning process the prepared dataset is saved
#write.csv(df.ai,file=paste("df.ai.csv"))For consistency and reproducibility of the results the following manipulation will be done by loading and using the saved dataset
# reading and furhter preparation
df.ai <- read.csv("df.ai.csv", header = T) # reading
df.ai <- df.ai[,c(2,6,9)] # keeping only useful column
#df.ai <- df.ai[,c(1,5,8)] # keeping only useful column
#View(df.ai)
tweets_text_network <- Corpus(VectorSource(df.ai$text)) # creating a text corpus used of the network
df.ai$text <- removeWords(df.ai$text, c("ai", "artificial intelligence", "artificialintelligence", "intelligence")) # removing some other particualar words
df.ai$text <- stripWhitespace(df.ai$text) # removing white spaces
df.ai$text <- gsub(" +", " ", df.ai$text) # removing spaces
df.ai$text <- gsub("\\s+", " ", str_trim(df.ai$text))
df.ai <- df.ai[!(df.ai$text == " "), ] # removing tweet with only spaces
df.ai <- df.ai[!(df.ai$text == ""), ] # removing empty tweet
df.ai <- df.ai[nchar(as.character(df.ai$text)) > 3, ]
#dim(df.ai) # left with
#View(df.ai)
#write.csv(df.ai,file=paste("df.ai.csv"))
tweets_text_corpus <- Corpus(VectorSource(df.ai$text)) # creating another text corpus for the word cloudCreating a wordcloud in order to visualize the most important words
wordcloud(tweets_text_corpus, min.freq = 7, max.words = 200, random.order=FALSE, rot.per=0.35,
colors=brewer.pal(8, "Dark2"))#, scale= c(3,0.25))
From the Word cloud it can be seen that the most used words are the ones related with the technological world, machine, big-data, technology and machine learning. Nevertheless are present words that concern the business area like business, marketing or industry.
Displaying a bar-chart of the words that have an occurrency greater than 150 times.
dtm <- DocumentTermMatrix(tweets_text_corpus)
freq = colSums(as.matrix(dtm)) # let's see the most frequent words
ord.df <- data.frame(words = names(freq), count = freq)
ord.df <- ord.df[order(-ord.df$count),]
ord.df <- ord.df[which(ord.df$count >= 150),]
ord.df$words <- factor(ord.df$words, levels = ord.df$words)
ord.df = ord.df[-3,]
#View(ord.df)ggplot(data = ord.df, aes(x = words, y = count)) +
geom_col(stat = 'identity', width = 0.7, alpha = 0.7, fill = "#009999") + coord_flip() +
labs(title="Barplot of the most used words", y="Occurence", x="Words") +
theme_classic() + theme(panel.border = element_rect(colour = "black", fill=NA))## Warning: Ignoring unknown parameters: stat
This plot allows us to have a better idea of the occurrence of the words. As already said the most important ones are related to technology but we can see that both business and COVID-19 have more than 200 occurrences.
Creating the network text
corp_net <- corpus(tweets_text_network)
tokns <- tokens(corp_net)
tokns<- tokens_remove(tokns, pattern = stopwords('en'))
tokns <- tokens_select(tokns,c("t","s",'"',"'", '-'), selection = "remove")
fcmat <- fcm(tokns, context = "document", tri = FALSE)
feat <- names(topfeatures(fcmat, 30))
fcm_select(fcmat, pattern = feat) %>% textplot_network(min_freq = 0.7)
Because the selection of the Tweets was made by searching for tweets with the hashtag #AI this term is in the center and it is the one that connected with all the others. Thanks to this visualization we can notice how the most used and linked words (the ones with the marked blue lines) are the ones that are most related to the technology where others are less mentioned and less related together.
For what I have stated before I will try a two topic modelling.
#Set parameters for Gibbs sampling
seed <-list(2003,5,63,100001,765)
#Run LDA using Gibbs sampling
ldaOut <-LDA(dtm,2, method="Gibbs", control=list(nstart=5, seed = seed, best=T, burnin = 4000, iter = 2000, thin=500))terms = as.matrix(terms(ldaOut,20))
terms## Topic 1 Topic 2
## [1,] "can" "data"
## [2,] "business" "learning"
## [3,] "covid" "new"
## [4,] "help" "machinelearning"
## [5,] "future" "iot"
## [6,] "market" "machine"
## [7,] "industry" "via"
## [8,] "digital" "technology"
## [9,] "use" "read"
## [10,] "time" "using"
## [11,] "gpt" "tech"
## [12,] "top" "bigdata"
## [13,] "like" "world"
## [14,] "work" "learn"
## [15,] "global" "analytics"
## [16,] "human" "now"
## [17,] "robotics" "best"
## [18,] "robots" "need"
## [19,] "see" "join"
## [20,] "companies" "today"
From the top 20 words per topic it can be stated that topic one is more Technical oriented presenting words like: data, machine learning, big data, analytic and so one. The second topic result to be more general with words like: market, future, global, industry, healthcare, digital. This distinction may suggest that there is a distinction in knowledge about this subject between the people that wrote the tweets.
Creating a column in the dataframe with the topic number
ldaOut.topics <- as.matrix(topics(ldaOut))
df.ai$topics <- ldaOut.topics
df.ai$topics <- as.factor(df.ai$topics)
#View(df.ai)Plotting pie for different topics
ggplot(df.ai, aes(x="", y = topics, fill = topics)) +
geom_bar(stat="identity", width=1) +
coord_polar("y", start=0) +
theme_minimal() +
theme(
axis.title.y = element_blank(),
panel.border = element_blank(),
panel.grid=element_blank(),
axis.ticks = element_blank(),
plot.title=element_text(size=14, face="bold")
) + theme(axis.text.x=element_blank())
This chart shows that the majority of the tweets are about topic two and so there are more “soft” knowledge tweets than deep knowledge ones.
Performing the sentiment analysis on the tweets and adding the result to the dataset
sentiment <- data.frame(get_nrc_sentiment(df.ai$text))
total <- data.frame(sum = colSums(sentiment, na.rm = TRUE)) #create a dataset that contin the sum of the sentyment
df.ai <- cbind(df.ai, sentiment)
# creating a dataset with the same of the sentiment according to the topic
sent.topic <- data.frame(topic_1 = colSums(df.ai[which(df.ai$topics == "1"), 5:14]))
sent.topic <- cbind(sent.topic, data.frame(topic_2 = colSums(df.ai[which(df.ai$topics == "2"), 5:14])))
# scaling in order to allow comparison
sent.topic$topic_1 <- abs(scale(sent.topic$topic_1))
sent.topic$topic_2 <- abs(scale(sent.topic$topic_2))
sent.topic_2 <- data.frame(Sentiment = rep(row.names(sent.topic), 2),
Topic = c( rep("1", length(sent.topic$topic_1)), rep("2", length(sent.topic$topic_2)) ),
Value = c(sent.topic$topic_1, sent.topic$topic_2))
#View(sent.topic_2)Plotting
ggplot (data = total,aes(x = rownames(total), y = sum, fill = rownames(total))) +
geom_bar(stat="identity", width=0.5) +
theme_classic() +
theme(legend.title = element_blank(),
legend.position = "none") +
labs(y = "Total Value", x= "Sentiments",title= "Sentiment value for the tweets") 
The bar-plot shows that the positive sentiment is the one with the highest total value by far and in general people tend to have a positive view of the technology (the other two following sentiment are trust and anticipation). This result could have been empathized by the fact that many machine learning algorithms have been introduced due to COVID-19. Indeed this result is supported by the high frequency in which the word “covid” is detected. Moreover the plot shows that the negative (negative and fear) sentiment are way less detected.
It is Interesting to look if there are differences in the sentiments according to the topic.
ggplot(sent.topic_2, aes(x = Sentiment, y = Value , fill = Topic)) +
theme_classic() +
geom_bar(stat="identity", position = "dodge") +
scale_fill_brewer(palette = "Set1") +
labs(title= "Sentiment value for the tweets according to their topic") 
This chart shows that there is not much difference in the sentiment of tweets depending on the topic classification made before. Hence can be stated that there ate not two clear fronts, pro and con AI.
Lastly I will classify the tweets according to a binary sentiment and then display the most occurrenced words depending on negative or positive tweets.
td <- tidy(dtm)
sentiments <- td %>%
inner_join(get_sentiments("bing"), by = c(term = "word"))
sentiments <- sentiments %>%
count(sentiment, term, wt = count) %>%
ungroup() %>% filter(n >= 5) %>%
mutate(n = ifelse(sentiment == "negative", -n, n)) %>%
mutate(term = reorder(term, n))
sentiments <- sentiments[sentiments$n >= 60 | sentiments$n <= -20,]
ggplot(sentiments, aes(term, n, fill = sentiment)) +
geom_bar(stat = "identity") +
theme_classic() +
labs(title= "Principal word for the contribution to the sentiment", y = "Words", x = "Contribution") + coord_flip()
The plot shows the most used words that characterize negative and positive labeled tweets. As can be seen, regarding the positive ones, most occurrence words recall an idea of beneficial and confidential growth. On the other hand negative words underline a fear of something that could be a problem, a risk.
head(df.ai[df.ai$text %like% "cloud", 1])## [1] "ibm cloud forum highlights hybrid cloud post pandemic businesses data storage asean business"
## [2] "machine learning market may set new growth story aibrain amazon anki cloudminds"
## [3] "machine learning market may set new growth story aibrain amazon anki cloudminds jewi"
## [4] "companies relying analytics cloud computervision blockchain ml driven solutions new norm"
## [5] "former soundcloud founders launch e bike subscription service backed blueyard via mike butcher"
## [6] "digital world everything sensing connected intelligent application g cloud"
By looking at the tweets that contain the therm “cloud” it can be seen that this word appears in negative labeled tweets because many of them concern cyber-security.
head(df.ai[df.ai$text %like% "drones", 1])## [1] "one advanced bird like robots ever made robotics drones uav biotech"
## [2] "machinelearning helps robot swarms coordinate case drones ml drone robotics robots mi"
## [3] "drones battlefield syrian regime troops captured black hornet nano spy drone"
## [4] "amazing advanced drones controlled student s minds via robotics drone robots"
## [5] "researchers improved system even spotting koalas bushfire areas using drones"
## [6] "government agencies coming forward providing avenues drones help"
By checking the tweets that present the words “drone” can be seen that only a couple of them talk about military drones and so that is why it could have a negative connotation. The others seems to have a positive sentiment so this could be a case of missclassificaiton.
Resuming the research questions:
Are there different conversation topics?
Thanks to this analysis it can be stated that tweets on AI go from technical to business tweets covering a wide range of topics.This can be seem by the word-cloud. Element that shows technical words, more business-related ones and even more colloquial ones. Nevertheless it has been shown that, even if borders are not so marked, a topic modelling can be done showing a more technical topic and a less one. As already said this could be a result of the presence of two different classes of users, one more competent than the other.
What people think about this technology, what are the feelings about the topic?
One pivotal aspect of the analysis was the sentimental one. Categorization shows an overall positive feeling with the presence of important value for trust and anticipation. Moreover the categorization in the two different topics seems not to have an influence on the sentiment of tweets.
Which are the words that have the most influence on the sentiment of users?
For completing the sentiment analysis I have displayed the most counted words for positive and negative tweets. This analysis has shown that both sentiment are characterized by the words that are facing the future. This could be a sign of the relative young age of the discipline and the huge steps made in the recent years leading to an undefined future perspective.
Moreover an aspect that needs to be take into account is the high presence of the term “covid”. In the past few month it was widely broadcasted by the news the importance that artificial intelligence algorithms can play in the battle against this pandemic. This aspect could have helped developing the overall positive sentiments found in the analysis.
The different techniques used allowed a fast overview of the topic giving easy and fast readable results. Moreover the previous steps can be scaled using larger amount. Furthermore the previous methods allow to perform a good analysis even with only free standards Twitter API, where for example using a survey would have been by far more expensive.
On the down side the previous analysis has not taken into account any information about the users. With other methodologies, for example by analyzing the results of a survey, more detailed insight could have been obtained. For example by asking personal information such as the level of instruction, age, working experience, living area, perhaps a correlation would have been found between these parameters and the opinion on AI leading to a more detailed analysis. Nevertheless, in 2019 a group of researchers have conducted a survey on 2 791 adult U.S. present on the microblogging platform. This research has highlighted that there exists significance difference between U.S. overall population and the sample’s one. Twitter users are younger, on average they are seven years younger. They are 11% more likely to have, at least, one bachelor’s degree and an income that is 9% higher than the one of the total populations. Twitter users are more likely to be Democrats and to have different opinion regarding migrants, gender, and race. The last aspect concerns the activity of the users, the 10% of them produce the 80% of the total amount of contents. [12] Given these hypotheses it is difficult to generalize the results obtained to the whole population.
Moreover the methodology used has not taken into account any type of historical data, hence, it is not possible to evaluate the any kind of trend in the opinion of people. Furthermore it would have been interesting to analyse historical data in correlation with particular global phenomena and so test the change in the sentiment according to other facts. For example the perception on this theme may have been changed before and after the Cambridge Analytica scandal and it would have been interesting to understand whether this change of perception has been maintained or not.
Another disadvantage of the technique used is for example the potential misclassification of words for the sentiment analysis, as shown for the “drone” word. This could have lead to a not so accurate analysis of the tweets and so to a not complete correct perception of them.
In conclusion it can be stated that the overall opinion about artificial intelligence is positive showing, significant expectations for the future, but always with an eye for possible risks.
[1] For the definition of the concept of Artificial Intelligence: European Commission, Brussels, 25-04-2018 http://www.governo.it/sites/new.governo.it/files/CommunicationArtificialIntelligence.pdf
[2] Lin Li, Lixin Qin, Zeguo Xu, Youbing Yin, Xin Wang, Bin Kong, Junjie Bai, Yi Lu, Zhenghan Fang, Qi Song, Kunlin Cao, Daliang Liu, Guisheng Wang, Qizhong Xu, Xisheng Fang, Shiqin Zhang, Juan Xia, Jun Xia, Artificial Intelligence Distinguishes COVID-19 from Community Acquired Pneumonia on Chest CT, Radiology, 2020, (doi : https://doi.org/10.1148/radiol.2020200905)
[3] J.Clement, Social media - Statistics & Facts, Statista, 2020 , available at: https://www.statista.com/topics/1164/social-networks/
[4] T. K. Das, D. P. Acharjya and M. R. Patra, Opinion mining about a product by analyzing public tweets in Twitter, 2014 International Conference on Computer Communication and Informatics, Coimbatore, 2014, pp. 1-4 (doi: 10.1109/ICCCI.2014.6921727.)
[5] B. Gokulakrishnan, P. Priyanthan, T. Ragavan, N. Prasath and A. Perera, Opinion mining and sentiment analysis on a Twitter data stream, International Conference on Advances in ICT for Emerging Regions (ICTer2012), Colombo, 2012, pp. 182-188 (doi: 10.1109/ICTer.2012.6423033)
[6] R. K. Bakshi, N. Kaur, R. Kaur and G. Kaur, Opinion mining and sentiment analysis, 2016 3rd International Conference on Computing for Sustainable Global Development (INDIACom), New Delhi, 2016, pp. 452-455.
[7] B.Sampath Kumar, D.Bhanu Sree Reddy, An Analysis on Opinion Mining: Techniques and Tools, Indian Journal of Research (2016)
[8] A.Go, R. Bhayani, Twitter Sentiment Classification using Distant Supervision,Stanford: CS224N Project Report , 2009
[9] S.Yang, H.Zhang, Text Mining of Twitter Data Using an LDA Topic Model and Sentiment Analysis, World Academy of Science, Engineering and Technology International Journal of Computer and Information Engineering Vol:12, No:7, 2018
[10] A. Alamsyah, W. Rizkika, D. D. A. Nugroho, F. Renaldi and S. Saadah, Dynamic Large Scale Data on Twitter Using Sentiment Analysis and Topic Modeling, 2018 6th International Conference on Information and Communication Technology (ICoICT), Bandung, 2018, pp. 254-258, doi: 10.1109/ICoICT.2018.8528776.
[11] R. J. Medford, S. N. Saleh, A. Sumarsono, T. M. Perl, C. U. Lehmann, An “Infodemic”: Leveraging High-Volume Twitter Data to Understand Public Sentiment for the COVID-19 Outbreak, 2020 (doi: https://doi.org/10.1101/2020.04.03.20052936) (Preprints and early-stage research may not have been peer reviewed yet.)
[12] S. Wojcik, A. Hughes, Sizing Up Twitter Users, Pew Research Center, 2019 Twitter Users”