Tourism in Perú has become very important, since there is a growing number of tourists arriving each year. This paper focus in understand what do speakingenglish tourists have in consideration when they visit Perú. We obtained all the tweets published in english during the year 2016, filtered by touristic places visited. In total, more than 192 thousand tweets were collected. We performed different analysis to describe the data, including correspondence analysis, a statistical technique which is normally applied to categorical data. The goal was to understand the sentiments and opinions expressed in those tweets.
Tourism is an important source of economic
growth in Perú. In 2015, 3.5 million of tourists
visited Perú, leaving more than USD 4,151
million that correspond to the 3.75% of the gross
domestic product (
In this context, every initiative oriented to understand the tourist preferences and sentiments is very valuable, because it allows to discover and develop the main attributes of touristic places in Perú, considering not only the place itself, but all the ecosystem: hotels, touristic agencies, handicraft stores, transportation, public services, etc.
For that reason, we downloaded 192,525
tweets, published during 2016, considering only
english language. After a careful data preparation
and feature extraction, we applied some sentiment
analysis techniques to tag each tweet using the
following eight sentiments: anger, fear, sadness,
disgust, surprise, anticipation, trust and joy,
proposed by Plutchik
In addition, we used other text mining techniques to determine some general concepts mentioned on tweets and identify the association between the places visited and these concepts. 2
Text mining has become a useful tool to deal
with the data deluge. With few and simple tools,
we can now extract valuable insights from a big
amount of unstructured data available today in the
internet. For example, one of the first studies of
sentiment analysis in Twitter was developed in
2011 by Agarwal et al. They built a formula to
calculate the polarity of a comment, if the tweet is
positive or negative. They developed a tree kernel
approach to improve the classification
In the same year, Dodds et al analyzed a huge
list of tweets, around 4.6 billion, posted by over
63 million individual users and extracted in a
33month span
Oku et al, analyzed 4.5 million tweets to find
the location where a tourist posted a photo. They
used only geotagging information from the
pictures, leaving text data out of the analysis
In addition, Bassolas performed an interesting
analysis about the 20 most interesting touristic
places around the world, including Machu Picchu.
They tried to measure the “attractiveness” of each
place, considering two metrics: a) Radius, defined
as the average distance between the places of
residence and the touristic site and b) Coverage, the
area covered by the users’ places of residence
computed as the number of distinct zones (or
countries) of residence
All these studies were considered to develop a different approach, focused on finding the emotions expressed by the tourists in the visited places. 2.1
This research was focused in five goals:
Understand the sentiment expressed in the tweet (positive or negative) about the place. Tag each tweet using the sentiments that the user is expressing in the tweet Identify the main qualitative attributes mentioned in the tweets, related to the touristic place.
Analyze the possible relationships between the places and the identified attributes.
Explore the general activity of the tourists which use twitter (frequent, eventual).
Having those objectives in mind, we determined the limits of our efforts, given the nature of the information source and the tools used. 2.2
There are various limitations to be considered: Being Twitter an internet social network, all the users must be registered and logged to use it. But not every person or tourist in the world has a Twitter account. For this reason, results are not generalizable/representative for the tourists’ population.
We acquired all the tweets directly from Twitter. We didn’t used streaming or scraping methods to download the tweets. But, we used many filters to select tweets referred to touristic places. In consequence, we could have selected a non-representative sample of tweets.
We considered only original posted tweets. We are not considering retweets. But, there are many platforms that post news or copies of tweets, e.g. IFTTT. Therefore, we could find the exact same tweet published by many different users.
Some Tweets (10%) include the location of the user but this information is not always real which makes classification by origin difficult. Besides, the georeferenced information is not representative. 3
Having reviewed the references and settled the scope of the study, we selected some tools and techniques available to analyze the information. 3.1
The population considered for this study was formed by 192,525 tweets downloaded directly from Twitter. The distribution of tweets is shown in the Figure 2. 3.2
Text Mining is the methodological process of
information extraction, where researchers deal
with collections of documents, using specialized
analysis methods
The complete process we have followed is shown in Figure 3. We collected tweets from 12 touristic places from Perú, that were suggested by Promperú, using words related to those places. We used Python with GNIP PowerTrack® API to retrieve and download the information.
The sintax used to obtain tweets was: (lang:en) (machupicchu OR picchu OR sacsayhuaman OR camino inca OR cusco OR cuzco OR plaza Armas cusco OR plaza armas cuzco OR ollataytambo OR salinas maras OR qorikancha OR coricancha OR ccoricancha OR pisac OR aguas calientes OR moray OR valle sagrado OR nasca lineas OR nazca lineas OR paracas OR islas ballestas OR manglares tumbes OR puerto pizarro OR parque reserva lima OR larcomar OR plaza armas lima OR parque kennedy OR catedral lima OR pachacamac OR monasterio santa catalina OR cañon colca OR colca OR uros OR taquile OR sillustani OR cathedral cusco OR cathedral cuzco OR main square lima OR nazca lines OR nasca lines OR ballestas island OR main square cusco OR square kennedy OR santa catalina monastery OR colca canyon OR sacred valley OR inca trail OR inKa trail OR cordillera blanca OR huaraz OR llanganuco OR amazon river peru OR mancora OR vichayito OR madre dios OR manu park OR iquitos OR tarapoto OR kuelap OR moche OR titicaca OR lake 69 OR laguna 69 OR señor sipan OR lord sipan OR pullana OR misti volcan OR misti volcano OR yanahuara OR huascaran OR churup OR chavin OR mochica OR pacaya samiria OR tucume OR cumbemayo OR baños inca OR wari OR punta sal OR amotape OR catacaos OR bahuaja sonene OR cotahuasi OR tingo maria)
During initial pre-processing of the datasets, all tweets where labelled with the cities and touristic places, by matching words related to the places. For example, if the text contained the words “Sacsayhuaman” and “Chan-Chan”, we put 1 in the columns “Cusco” and “La Libertad”. It means, the tourist is talking about attractions in two different regions. And 0 in all other columns, corresponding to other Peruvian regions.
Not every tweet is clear and correctly written, we had to perform a cleaning process, considering the next steps: a) Drop Retweets, Hashtags and mentions. b) Drop strange symbols c) Drop webpages d) Drop punctuation signs e) Drop numbers f) Drop tabs or multiple spaces
We frequently found spelling mistakes in the
tweets. For that reason, we used a spelling
corrector. E.g. we replaced ‘speling’ by ‘spelling’ with
an algorithm in Python
After the cleaning process, we tried to understand the dataset, analyzing the comments, using different R packages (qdap, koRpus, tm). SAS Text Miner was helpful for Topic Modeling and Term Relationship diagrams. Finally, R was used for Sentiment Classification (library syuzhet) and Correspondence Analysis (function corresp).
We have pre-defined some “concepts”, using
the Text Profile Node of SAS Text Mining
In our case, we also identified, all these words, considering verbs, to represent the activities that tourists do in the places and adjectives, to represent the opinions or value perceived by the tourist.
- Wordclouds - Correspondence Analysis - Perceptual Maps - Term relationships - Topic Modeling - Tagging/Classification - Ortographical Review - Terms filtering - Cleaning - Download tweets - Format - Preparation
The first sentiment analysis performed with text
data, were published around 2002. As Agarwal
stated, most of the studies were focused in text
classification, positive or negative categories
One of the first teams performing sentiment
analysis with Twitter information, was leaded by
O’Connor. They also considered only the
classification of 1 billion Twitter messages, but looking
the changes over time
Mohammad developed a historic review of
Sentiment Analysis, but considering Emotions
In fact, some tools include a process which use
these concepts to classify texts, for example the
Library Syuzhet in R. This library calls some
others, to have a great performance when scoring
sentiments and calculating polarity. With this
library, we can use the function get_sentiment with
two arguments (text and method, the default is
‘syuzhet’) in order to calculate the score for the
text in each one of the eight main sentiment
categories listed in Figure 4. The greatest score
indicates the more prevalence of the sentiment. The
default method uses a sentiment lexicon,
developed in the Nebraska Literary Lab under the
direction of Matthew L. Jockers
Maybe, the most representative researcher in
Correspondence Analysis made by Benzécri. He
has not invented this technique, but consolidated
the methods involved (
Correspondence Analysis was created as a qualitative statistical dimension reduction technique. Is very useful for everyone to understand the numeric associations between two categorical factors. After performing a matrix decomposition of the analyzed crosstab, you can represent the original frequencies in a bi-dimensional space.
As stated in the book of Venables and Ripley, after applying the method Singular Value Decomposition (SVD) is easy to calculate the row and column Scores.
Suppose we have an r x c table N of counts. Correspondence analysis seeks ‘scores’ f and g for the rows and columns which are maximally correlated. Clearly the maximum correlation is one, attained by constant scores, so we seek the largest non-trivial solution. Let R and C be matrices of the group indicators of the rows and columns, so RTC = N. Consider the singular value decomposition (SVD) of their correlation matrix: = ⁄ −( .⁄ )( . ⁄ ) √( .⁄ )( . ⁄ ) = − (1) √
Where = .⁄ and = . ⁄ are the
proportions in each row and column. Let Dr and Dc
be the diagonal matrices of r and c.
Correspondence Analysis corresponds to selecting the first
singular value and the left and right singular
vectors of Xij and rescaling by Dr-1/2 and Dc-1/2,
respectively
In our case, see Table 1 and 2, the first input matrix is:
anger anticipationdisgust fear joy sadness surprise trust MachuPicchu 7,679 31,455 6,596 11,297 21,302 10,791 17,428 22,911 CaminoInca 799 4,763 351 1,290 2,992 988 1,879 3,164 ValleSagrado 461 2,153 331 733 2,095 529 962 1,830 AguasCalientes 114 1,269 57 104 253 74 1,022 278 Sacsayhuaman 55 780 44 136 708 99 749 943 L.Titicaca 1,521 2,889 1,434 3,399 2,553 2,290 2,339 2,547 Table 1: Distribution of tweets by touristic places and sentiments expressed.
The other matrix is: amazing beautiful best excellent good love bad old MachuPicchu 6,421 2,037 2,521 1,154 1,701 1,482 244 1,880 CaminoInca 695 153 493 368 405 152 97 275 ValleSagrado 464 361 143 111 109 106 8 335 AguasCalientes 48 20 73 10 24 12 6 4 Sacsayhuaman 47 10 10 15 5 11 3 58 L.Titicaca 334 230 66 64 127 127 32 177 Table 2: Distribution of tweets by touristic places and concepts. 3.5
After applying SVD, in each matrix, we get the row and columns scores. We can use them to generate the Correspondence Plot. It could be done putting in a scatterplot the scores calculated for each matrix.
Some researchers have used this kind of charts
in text analysis before, but having different
approaches. For example, Balbi et al developed two
studies
Interestingly they used first all the verbs and adjectives of the analyzed texts to find those words called “Concepts”, which express key characteristics for them. 4
Having downloaded more than 500 thousand tweets, pre-filtering touristic places and content, we needed to filter the data more carefully, because it contained a lot of tweets talking about other subjects. In consequence, we discarded: a) Retweets b) Tweets without text, only pictures c) Tweets in other languages.
After this process, we got 192,525 tweets, but published by only 70,161 unique users. Even twitter is a popular social network, users have different patterns of activity. We considered “Active users” those who published at least 12 tweets in all the span analyzed, from January to December 2016. At least one tweet per month, in average. We discovered that only 2.5% of the total were active users shown in Figure 6.
192,525
70,161
After that, we identified the most active users, in order to discover specialized users, bots (robots or automatic processes that publish tweets) and other kind of strange or interesting patterns. We found in the top some touristic companies, users who love to travel, accounts for travelers, etc. Shown in Table 3.
N
User
Kondor Path Tours (@kondorpathtours) is a travel agency, very active in Twitter, publishing five times tweets more than the next one in the list. L. J. Blake (@lookeastwest) is a travel blogger, who travels around the world and post photos, places reviews and advices to other travelers.
After performing the twitter cleaning process, explained in section 3.2, we obtained the most frequent words in the corpus. In English tweets, Machu Picchu is most popular than Perú or any other Peruvian place or word. And it appears on the first places in the list again as “Machupicchu” in Table 4.
If we use wordclouds, the visualization is not very helpful, because the most frequent words hide all the other ones in Figure 7.
We filtered some words to see the picture most clear, in our case we decided filter machu picchu, machupicchu, peru, cusco, travel, inca and trail in Figure 8.
Attitude Valuation: After cleaning processes,
we used syuzhet package, to classify each tweet in
Positive or Negative, depending on the polarity
expressed. This method is not 100% accurate, but
it gives a general idea. Other limitation is that
tweets have only 140 characters, a limited way to
express complex feelings
Ica (8,442)
62% 16% 22% 32% 23% 18% 11% 6% Negativa Neutral Positiva
Negativa Neutral Positiva Negativa Neutral Positiva
Negativa Neutral Positiva Negativa Neutral Positiva
Negativa Neutral Positiva
We analyzed the opinions of all tweets related to touristic places in Perú, and after that, each region was considered in the analysis separately, with the count of positive, negative and neutral tags. According to the general results, Cusco is the region with the highest percentage of positive tweets and Lima is leading the group of neutral opinions shown in Figure 9.
We could see that, in general, there are many tweets classified as neutral. It is common in text classification, due to the nature of social media. The most frequent case is a user talking about news or things that are happening. For example: “I’m in Cusco”, or “The truth about Paracas Skulls”. In our context, trying to understand the attitude of the tourist, we could simplify the classification, from three levels to two. Only negative opinions or non-negative opinions, grouping positive and neutral ones.
Puno has the highest rate of negative tweets, because in early October, ten thousand of giant toads were found dead in the Titicaca Lake. According to Journalists, pollution in the Coata River would be blame for the deaths. It triggered a lot of claims and campaigns in social networks. Some protesting activists took around 100 dead frogs to Cusco (160,193)
55%
the central square in the regional capital, Puno. A
BBC article describing the disaster was shared a
lot of times in Twitter
Touristic Places: Considering the most visited places mentioned by Promperú, we looked for negative and non-negative opinions in Table 5.
The place with more negative tweets is Titicaca Lake. The places with better performance are Valle Sagrado, Aguas Calientes, and Caminos del Inca, in Cusco. Machu Picchu has a good performance too, but some incidents affected the reputation of them in Twitter.
Parque Kennedy and Larcomar, in Lima, have good image, too. The district of Miraflores is very touristic. They realized the great potential of tourism in Perú and many good hotels and restaurants are in the zone.
Concept Identification: Using Hierarchical
Bayesian Methods, with the option ‘Text Rules’ in
SAS Enterprise Miner, we identified the words
which most appeared in negative tweets. It
considers rules based on words co-occurrence to
perform a simple classification
We focused the analysis in each touristic place analyzed, in order to find specific words associated to the place. Some examples are shown in Figure 10.
Those words are meaningful, because reflect some incidents occurred in Machu Picchu. A German tourist died in October, while trying take a picture of himself. He fell off a cliff in the mountains. In addition, we found tweets about “andean astronomers” or “shamans” who perform a ritual known as “Ayahuasca”, which consists in drinking an Amazonian plant mixture, that is capable of inducing altered states of consciousness, usually lasting between 4 to 8 hours after ingestion. They could be related to bad or strange experiences, at least.
In the other direction, we also selected other words that appeared in non-negative tweets. Most them are related to describe good experiences with the landscapes, trekking routes and people in Figure 11.
In general, tourists enjoy ‘beautiful’ natural landscapes, ‘spectacular’ ancient buildings, ruins, food and services, ‘colorful’ handicrafts and traditional clothes. They liked to go hiking across the valley, too. All those words were collected to do the next task, which is grouping most relevant expressions in concepts, using co-ocurrence and synonyms.
Relationships between terms: To have a
better understanding of the words and concepts
discovered in the previous step, we performed a Link
Analysis, between terms A and B, using a single
metric, called Strength
= = ∑ = ( )~
( ) ( , ) ) (2) (3) (4) n: Number of documents that contain term B k: Number of documents containing A and B Then, we can build some graphs, to represent the strongest relationships: 3,368 3,266
3,501 joy anticipation anger disgust sadness surprise fear trust 825 371
1,028
Lago Titicaca 1,584 1,514 2,368 2,487 1,982 1,370 3,988 joy anticipation anger disgust sadness surprise fear trust 509
Parque Kennedy about the negative experiences, the thickness of the lines represents the degree of relationship. We can build the same graphs for Non-Negative terms to find things that tourists enjoy or like to do, shown in Figure 13.
Emotions discovery: The next step is dealing
with sentiments. In order to classify all the tweets,
we use the syuzhet package
Machu Picchu 25,165 26,437 8,015 6,737
18,655 11,630 12,439 joy anticipation anger disgust sadness surprise fear trust 339 202 120 141 228 194 joy anticipation anger disgust sadness surprise fear trust
We can notice the differences between places most accepted, like Caminos del Inca or Parque Kennedy. The percentage of tweets that contains bad feelings like anger, disgust, sadness or fear is really low, against the other positive ones. In contrast, Lago Titicaca has fewer differences between the sentiments expressed in the tweets. The most frequent is fear.
Correspondence Analysis: We now have the key elements to build the crosstabs, perform the correspondence analysis and draw the plots. First, we can observe the distribution of sentiments per place visited: 3,658 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0%
MachuPicchu CaminoInca ValleSagrado AguasCalientes Sacsayhuaman L.Titicaca joy anticipation anger disgust sadness surprise fear trust Iquitos R.Amazonas R.Paracas Lin.Nasca Pq.Kennedy Larcomar joy anticipation anger disgust sadness surprise fear trust
We have divided the visualization in two parts,
shown in Figure 15, because is hard to represent
twelve places in one shot. Having the complete
picture, we can see the predominant emotions by
place. For example, Aguas Calientes has a lot of
tweets expressing surprise and anticipation. Lago
Titicaca is associated with fear and sadness
expressions. Lineas de Nasca has many words
related to fear, like “misterious” or “unknown”. Other
similar reactions are caused by accidents
occurring to aircrafts with an alarming frequency
The next step is to reduce dimensionality, calculating the row and column scores, from the tables shown in Table 1 and 2. With the scores, we built the correspondence plots. Like the bar plots in Figure 14, we divided the representation in two blocks, to keep the results visible. We can see the sentiments associated to each place in the plots. 2 1 2 n o i 0 s n e m i D -1 -2 -3 2 1 2 n o i 0 s n e m i D -1 -2
Sacsayhuaman AguasCalientes
Lin.Nasca fear
ValleSagrado trust joy CaminoInca
MachuPicchu anticipation surprise anger disgust sadness
fear anger disgust
L.Titicaca 2 joy
Pq.Kennedy R.Paracas Larcomar trust Iquitos sadness anticipation R.Amazonas surprise -3 -2 -1
The first plot in figure 16 is very clear about Lago Titicaca. All the negative emotions (‘disgust’, ‘fear’, ‘sad’ and ‘anger’) are close to this place, which is interpreted as a relationship. The words ‘trust’ and ‘joy’ are more related to places such Valle Sagrado and Camino Inca. We have explained the situation with pollution and death of animals in Puno, the relationship could be confirmed. In the case of Cusco, there local government and authorities are aware about the importance of providing the best experiences to the visitors, so they and the touristic business (operators, hotels, restaurants, guides, etc.) work together to bring good services and hospitality.
In the second plot in figure 16, we can see the proximity between Lineas de Nasca and the sentiment ‘fear’, Rio Amazonas with ‘surprise’ and Paracas, Parque Kennedy and Larcomar associated with the sentiments ‘joy’ and ‘trust’. In the case of Nasca, many tourists commented about the insecurity to take a flight in old and out-of-dated airplanes. 22 Sacsayhuaman iiseonnDm01 ValleSagrado beautifulenjoolyd happyfun love MachuPicchu best amazing wait -3 -2
small
L.Titicaca luxury -1
AguasCalientes
bad CaminoInca goodexcellent
Pq.Kennedy happy
bad wait Iquitos
small love good best beautiful
fun excellent amazing
Larcomar enjoy
R.Amazonas old
Lin.Nasca R.Paracas luxury -3 -3 -2 -2 0 1 1
2
D-i1mension-11
Considering the concepts, we have selected some Non-Negative ones: Amazing, Best, Beautiful, Enjoy, Excellent, Fun, Good, Happy, Love and Luxury. The other, Negative Concepts selected were: Bad, Old, Small, Wait. All of them were found as common and frequently appeared in the tweets, see Figure 17. They are used to describe or express valuations about their experiences in the visited places. They resume other common expressions, to avoid synonym issues in the interpretation of texts, and were grouped using linking formula (1).
Geolocation: In addition, we represented all the geo-tagged tweets in a map. Only 17 thousand tweets had Longitude and Latitude coordinates. Notice that coordinates correspond to the user location in the moment when the tweet was published. For that reason, 92% of the tweets were posted from any place in Perú, shown in Figure 18.
Negative Neutral Positive This study was a good way to explore the opinions about Perú, but we have to keep in mind some limitations, like representability or social network popularity in our country. Twitter activity is the highest in July, during the northern hemisphere countries summer vacation time. And in October, when some incidents occurred in the most visited place, Cusco. Acknowledgments We would like to express our special thanks to Market research team in Promperú, who gave us the opportunity to develop this project.