In this paper is presented a propose solution to the Sentiment Analysis challenge presents in the Recommendation System for Text Mexican Tourism task during the Iberian Languages Evaluation Forum 2021. The task consists of predicting the polarity of an opinion issued by a tourist who traveled to the most representative places of Guanajuato, Mexico. Thus, our approach is based K-Nearest Neighbors by using a distance based on the Jaccard coe cient concept. In the training stage, by using the training data, our approach rst clusters every word from every opinion (review) by the respective class. Then, the stop words from each cluster are deleted. After, the normalized frequency of each word in a class is computed. In this way, the set of words (trained words) with their normalized frequency (NF) are used as class feature vector. In the classi cation stage, when a new opinion is given, each word from it is intersect with the trained words for each class and the NF of the intersected words are summed (dissimilarity value). The predicted class is assigned to the class with the most high dissimilarity value. The performance on the testing data were of 1.26 MAE and 0.22 of F-measure. We think that the obtained results are because the data is unbalanced and our approach does not deal with that issue.
In 2018, the World Economic Forum reports that the travel & tourism
industry generated 10.4% of the world GDP and supported over 319 million jobs [6].
In the last year, global tourism has been impacted strongly due to COVID-19
pandemic and in the last decade tourism has also been in uenced by numerous
technological advances and tools such as digitization, information and
communication technology, machine learning, robotics, and arti cial intelligence (AI)
[
Most of international travelers plan their trips by digital means, and a big
part of their decisions rely on other travelers shared online information, e.g.
online touristic reviews [5]. To synthesize large amounts of reviews, it is essential
to use algorithms from the Arti cial Intelligence eld, speci cally the area of the
Natural Language Processing (NLP). his sub- eld of the arti cial intelligence
aims to achieve human-like processing capabilities of the language for diverse
scopes [
One task of the Recommendation System for Text Mexican Tourism task
during the Iberian Languages Evaluation Forum 2021 [
This work is organized as follows: { Section 2 describes the task to solve. { Section 3 shows in details the proposal followed in this work. { In section 4 the results are presented.
{ Finally, section 5 presents the conclusions and limitations of our proposal. 2
The subtask is a classi cation task where the participating system can predict the polarity of an opinion issued by a tourist who traveled to the most representative places of Guanajuato, Mexico. Guanajuato city is a well-known destination for domestic tourists and it has gained a progressive notoriety in the international arena since the last quarter of the previous century. Apart from famous international destinations within the Mexican territory such as the cases of Cancun and Mexico City, Guanajuato ranks number six among the most visited cities for tourism purposes1 . Thus, this Sentiment Analysis problem is de ned as follows: { "Given an opinion about a Mexican tourist place, the goal is to determine the polarity, between 1 and 5, of the text." Where 1 indicates most negative and 5 most positive. 2.1
Data set
This collection was obtained from the tourists who shared their opinion on
TripAdvisor between 2002 and 2020. Each opinion's class (review polarity) is an
integer between [
Training data set consists of 5197 instances. Table 1 shows the distribution of the review polarities for the training data set. It is important to mention that the training data set the classes are unbalanced. The test data set contained 2216 instances (the distribution polarity is unknown). 3
Our proposal consists in two main stages: training stage and classi cation stage. We describe each stage below. 3.1
Training stage In this stage, we use the using the training data to extract features of each class. Thus, our approach rst clusters every word from every opinion (review) by the respective class. Then, the stop words from each cluster are deleted. We call to the result set of words fro class c, trained words c. After, the normalized frequency, !i;c, of i-th word in the class, c, is computed. In this way, the sets c with their normalized frequency !i;c (for c 2 1; 2; 3; 4; 5 and i = 1; 2; :::; Nc where Nc = j cj) are used as class feature vector.
2https://sites.google.com/cicese.edu.mx/rest-mex-2021 3.2
classi cation stage In the classi cation stage, when a new opinion/review is given, the stop words from it are deleted. Then the resulting set of words for that opinion is called . After, each word in is intersect with the set c (trained words of class c). Then, the NF of the intersected words are added. This can be represented by equation 1:
Sc =
X k2 c\ !k;c Note that equation 1 is based on the concept of the Jaccard Distance [7]. Thus, the predicted class for the review , C( ), is assigned to the class with the most high similarity value Sc:
C( ) = arg maxfScg 8 c 2 f1; 2; 3; 4; 5g c (1) (2) Equation 2 is the representation of the K-NN method, when K= 1. 4
Figure 1 show the wordclouds of the sets c weighed with their corresponding !k;c for all classes. Note that although there are many words repeated in all classes words as "museo" and "momias" are more frequent in classes 1 and 2, and word as "guanajuato" and "historia" are more frequent in classes 3 to 5.
The o cial results for our proposal are as follows: { Accuracy: 36.95 { F-measure: 0.22 { MAE: 1.27 In this sense our approach obtained the last place according to MAE value (15th place). According to accuracy, we obtained the 12th place. Finally, we obtained the 9th place with accordance to F-measure. 5
In this work, we presented a simple solution based on the concept of the Jaccar Distance to classify the sentiment analysis problem presented on Recommendation System for Text Mexican Tourism task during the Iberian Languages Evaluation Forum 2021. Although, our proposal is based in a simple idea it showed potential. The most signi cant disadvantage of our approach was the unbalance training data set. Additionally, our proposal could be improved by removing the representative words as subjects and only work with qualifying adverbs.
(a) Label/Class 1 (c) Label/Class 3 (d) Label/Class 4 (e) Label/Class 5 5. Calderon, F.A.C., Blanco, M.V.V.: Impacto de internet en el sector tur stico. Revista UNIANDES Episteme 4(4), 477{490 (2017) 6. Calderwood, L.U., Soshkin, M.: The travel and tourism competitiveness report 2019 (Sep 2019) 7. Alvarez Carmona, M.A., Franco-Salvador, M., Villatoro-Tello, E., Montes-y Gomez, M., Rosso, P., Villasen~or-Pineda, L.: Semantically-informed distance and similarity measures for paraphrase plagiarism identi cation. Journal of Intelligent & Fuzzy Systems 34(5), 2983{2990 (2018). https://doi.org/10.3233/JIFS-169483, publisher: IOS Press 8. Chowdhury, G.G.: Natural language processing. Annual review of information science and technology 37(1), 51{89 (2003) 9. Gossling, S., Scott, D., Hall, C.M.: Pandemics, tourism and global change: a rapid assessment of covid-19. Journal of Sustainable Tourism 29(1), 1{20 (2021). https://doi.org/10.1080/09669582.2020.1758708, https://doi.org/10.1080/09669582.2020.1758708 10. Guerra-Montenegro, J., Sanchez-Medina, J., Lana, I., SanchezRodriguez, D., Alonso-Gonzalez, I., Del Ser, J.: Computational intelligence in the hospitality industry: A systematic literature review and a prospect of challenges. Applied Soft Computing 102, 107082 (2021). https://doi.org/https://doi.org/10.1016/j.asoc.2021.107082, https://www.sciencedirect.com/science/article/pii/S1568494621000053 11. Nadkarni, P.M., Ohno-Machado, L., Chapman, W.W.: Natural language processing: an introduction. Journal of the American Medical Informatics Association 18(5), 544{551 (2011) 12. Qiu, R.T., Park, J., Li, S., Song, H.: Social costs of tourism during the covid-19 pandemic. Annals of Tourism Research 84, 102994 (2020). https://doi.org/https://doi.org/10.1016/j.annals.2020.102994, https://www.sciencedirect.com/science/article/pii/S0160738320301389