Social networks have become the most representative tools of Web 2.0 (the Internet), allowing millions of users to post and share information in a fast and seamless way, permitting a continuous flow of information. According to figures from different studies [1], the latest data from 2022 state that more than 9 out of 10 Internet users already use social media every month, and the percentage of social media users amounts to 75% of the world’s population. In this scenario, 17 social media platforms accounted for at least 300 million active users in January 2022. Facebook is the most used, with over 2.91 billion monthly active users, followed by YouTube with 2.562 billion users, and WhatsApp with 2 billion users. Twitter, which combines the features of blogging, social networking, and instant messaging, has about 436 million active users worldwide, standing out not only for its popularity but also for its great monetisation and business potential. As a matter of fact, Twitter seems to be the seventh-favourite social network among Internet users in the 16–64 years age group [1].

In any case, due to the impressively growing amount of information on the Internet and social networks, as well as the large number of sources and the high number of opinions about any given content, it is essential to have automatic methods that allow us to classify and analyse information quickly and efficiently; all this is used to feed any decision-making system. Natural language processing (NLP) seems to be a more rational choice to tackle complexity in the area of artificial intelligence (AI) [2]. The well-known specific fields required in this area are opinion mining (OM) and sentiment analysis (SA). These fields combine NLP and computational linguistics, which involve uncovering words and contexts to understand the opinions they reveal. More precisely, SA deals with determining the emotional tone behind a series of words in order to classify them via their polarity (positive, negative, or neutral) or emotion (sadness, joy, disappointment, etc.). Additionally, contents on social networks such as Twitter have different natures depending on the language employed, which has its own peculiarities, making it different from other uses of language [3]. In fact, some important problematic features of tweets are related to their short length (280 characters), the data sparsity, the lack of context, the low concern for grammar, and the use of an informal linguistic style (idioms, slang, and abbreviations). These characteristics make it difficult to integrate fully effective SA systems into social networks such as Twitter.

Moreover, the application of SA in social networks is very useful in many cases, such as measuring the impact of social media actions, helping understand what consumers think about brands/companies, knowing what users think about certain topics, making better decisions on marketing strategies for product/service development, following trends in real-time, or even assisting in predicting the behaviour of users. For these reasons, the application of NLP strategies and SA to social networks is an active area of applied and basic research in order to make social networks and the web more usable and profitable.

2. Sentiment Analysis Based on Lexical Resources

As for techniques based on Spanish lexical resources, there are several opinion lexicons that encompass words associated with a sentiment or an opinion value (polarity). However, the amount of Spanish lexicons is quite limited. As far as polarity lexicons are concerned, it can consider the opinion lexicon developed by ^[4][9] or the iSOL ^[5][11], based on one of the most important English lexicons for the classification of polarity ^[6][12]. In addition, the Spanish adaptation of the Affective Norms for English Words (ANEW) is presented in ^[7][13]. In another work, the authors of ^[8][14] developed a lemma-level sentiment lexicon for several languages, including Spanish, generated from an improved version of SentiWordNet, a popular lexicon with positive and negative words. The authors of ^[9][15] have developed Sentitext, a sentiment analysis software for Spanish, based on knowledge and supported by several word databases. In contrast, other lexicons are based on emotions rather than polarity, such as the one developed by ^[10][16], which considers six emotions. However, using only opinion lexicons is not entirely effective for all domains, as each domain usually has a specific way of expressing a positive or negative opinion. Moreover, these lexical resources alone are not able to cope with the detection of negation, sarcasm, or ambiguity in opinions. Consequently, many authors focus on the processing of negation in Spanish. Thus, the authors in ^[11][17] adapted the SO-CAL sentiment analysis tool from English to Spanish ^[12][18]. This proposal consists of a lexical dictionary with positive and negative words together with the integration of rules and intensifiers to detect the degree of negation. Furthermore, the authors in ^[13][19] incorporated dependency-based techniques for negation detection to establish the scope of intensifiers and negation cues. This system achieves an accuracy of around 78% in detecting only two polarities (negative or positive), but it is applied in the context of product and service reviews, not on Twitter, where the language characteristics are different. In this sense, the authors of ^[14][20] have proposed rules based on dependency trees to identify the scope of the most important negation cues, defined by the Royal Spanish Academy, together with a lexicon-based sentiment analysis system for polarity classification. This system was applied to study the scope of negation in Twitter ^[15][21], reaching accuracies of around 62% when detecting three levels of polarity (positive, negative, or neutral). In the same line of work, Miranda et al. ^[16][22] developed an opinion mining system based on the ANEW Spanish lexicon applied to hotel reviews using the negation cues studied in ^[14][20]. Although the accuracy increases above 90%, it only focuses on detecting two polarities (negative and positive) and does not integrate the more difficult-to-identify neutral polarity. Furthermore, the system was developed in the context of hotel reviews, not in the more complex context of social networks such as Twitter. Amores et al. ^[17][23] combined different methods to deal with negation by applying effects of negation, modifiers, jargon, abbreviations, and emoticons in sentiment analysis, reaching an accuracy on a Twitter corpus between 83 and 87% for the detection of two polarity classes: positive and negative. Other approaches take into account not only sentiment but also subject matter. Thus, Anta et al. ^[18][24] have proposed a classification system for Spanish tweets by evaluating the use of stemmers and lemmatisers, n-grams, word types, negations, valence shifters, link processing, search engines, special Twitter semantics (hashtags), and different classification methods, where the highest accuracies are around 58% for topics and 42% for Twitter polarity detection. Furthermore, the authors in ^[19][25] implemented a naïve-Bayes classifier to detect the polarity of Spanish tweets, identifying different levels of polarity along with unigrams of lemmas and multiwords based on PoS tag patterns to detect the scope of negation. The accuracy of the system was 66% for four polarity levels and 55% for six polarity levels. Finally, the authors of ^[20][26] describe a Transformer-based approach to detect negation in a corpus of Spanish product reviews, achieving accuracies between 80% and 90%, although this system has not been applied in social network contexts or on Twitter. However, other negation strategies are based on annotated corpora with negation in contrast to negation cues ^[14][20].  Finally, the most recent Spanish negation corpus, the SFU ReviewSP-NEG18, also considers discontinuous negation markers and is related to products and services reviews. Therefore, it can be concluded that lexical approaches require the use of sentiment word dictionaries combined with strong negation detection strategies. However, in addition, much of these works have not been applied in social network contexts such as Twitter, where language features are more complex (sparsity, lack of context, little concern for grammar usage, informal linguistic style, etc.). Moreover, these approaches also need to integrate strategies to detect other language features, such as sarcasm or irony, which are very typical in social networks.

3. Sentiment Analysis based on Supervised Machine Learning

However, techniques based on supervised machine learning do not require lexical resources but a set of previously tagged examples of opinions. However, obtaining labelled examples is a costly task, as these examples are usually labelled manually; in addition, it is important to have examples for the classification domain in which it is applied.  In ^[21][27], Myska et al. focus on sentiment analysis in text documents by applying a support vector machine (SVM) classifier. For training and testing, a dataset with positive and negative valence texts was used, based on the analysis from web pages with product ratings. Based on the rating of the text, it was resolved whether the text is positive or negative. The texts were divided into three groups: positive texts (P), negative texts (N), and texts with no defined valence NEU (neutral, not used). The recognition system was validated in four different languages, including Spanish. Regarding the results in the Spanish language, the classification accuracy was 93.23% when predicting P and N texts, not without considering the neutral class. Although this work shows interesting results, it was carried out with posts on web pages that have a different nature than Twitter posts as, for example, the size of the latter are limited to 280 characters. Thus, the use of language on Twitter has its own characteristics that differentiate it from other uses of language [3]. In addition, the samples database was categorised as P, N, and NEU based on starts and some thresholds. In contrast, theour database is manually labelled via human effort. All these differences make it impossible to generalise their results to the SA of tweets in Spanish and justify the need to explore other alternatives for this purpose by applying Deep Learning. In a similar context, the authors of ^[21][27] proposed a linguistically independent text classifier based on convolutional recurrent neural networks. The classifier works at the character level instead of higher structures (words and sentences). The models were tested on the Yelp dataset (reviews and user data) and on a privately collected multilingual dataset, and the classification accuracy for Spanish texts was only 67.33%. Focusing on sentiment analysis of Spanish tweets, some research can be found in the literature. In ^[22][28], SA was carried out in the context of the Colombian presidential election of 2014. In this case, the corpus consisted of 1030 tweets tagged via humans with positive, negative, or neutral polarity towards each of the candidates. This process included a feature extraction preceded by a normalisation phase, after which a logistic regression classifier was used to assign a label class to each tweet. The results showed some difficulties in inferring the vote based on sentiment analysis of the tweets. In the conclusions, the authors argued that the obtained results showed that inference methods based on Twitter data are not consistent and that more work is needed to deal with the characteristics of the language. In this sense, theour proposal tries to avoid the effects of feature extraction with the use of deep learning techniques. Deep learning makes problem solving much easier because it completely automates what used to be the most crucial step in a machine learning workflow, which is feature engineering ^[23][29], as they rely on the use of unsupervised pre-training features, the most commonly found of which are word embedding vectors ^[24][30]. More recently, the authors of ^[24][30] built the first Spanish corpus of sexist expressions on Twitter and applied different techniques, such as SVM, random forest (RF), long short-term memory (LSTM) networks and Transformer Bidirectional Encoder Representations from Transformers (BERTs), to construct a novel Transformer architecture that aims to provide very promising results in NLP. The proposed methods achieved an accuracy between 61 and 74% in the automatic detection of sexist behaviours. However, this corpus is not oriented towards polarity detection as it focuses on the detection of sexist behaviour on Twitter. In a similar context, research can be found in ^[25][31], where the authors developed a proposal to detect hate speech in Spanish tweets, using BERT. The results allow distinguishing between non-aggressive and aggressive tweets (two classes) with an accuracy between 79 and 86% on different datasets. However, the authors of ^[26][32] have proposed an approach for annotators to reach a consensus in the process of annotating comments on Spanish social networks. They built a corpus with 3259 Spanish comments (P, N, and NEU) and applied several classifiers for sentiment analysis detection, achieving 70% as the average F1-Score with multilayer perceptron. In the context of the Workshop on Sentiment Analysis at SEPLN (TASS), studies focusing on SA in Spanish tweets can be found ^[4][27][6,9] using the TASS corpus. For example, in ^[28][33], a set of classifiers based on SVM, convolutional neural networks (CNNs), and LSTM were used over three variants of Spanish. The achieved results show the best accuracies around 59%, 59.2%, and 54.9% for the SVM, CNN, and LSTM algorithms, respectively. Similar results were reported in ^[29][30][34,35]. Another approach using SVM was implemented in ^[31][36] for SA on Spanish tweets, achieving an accuracy of 62.88% when predicting five levels of polarity (P+, P, N, N, N+, and NEU) and 70.25% for three levels (P, N, and NEU). Furthermore, the authors of ^[27][6] have proposed two deep neural network models (CNNs and dense neural networks) integrating a Gaussian noise layer for tweet polarity classification in a Spanish Twitter corpus, achieving an accuracy of only 57%. The current state of sentiment analysis systems on Twitter continues to advance with the integration of hybrid deep learning architectures, a trend supported by recent research. In ^[32][37], Shazly et al. propose a hybrid architecture that combines the power of bidirectional recurrent neural networks (Bi-RNNs) and other techniques to enhance the efficiency and accuracy of sentiment analysis on Twitter data in an Arabic benchmark dataset. Bi-RNNs are employed to capture contextual information and sequential dependencies within tweets, which is crucial for understanding the nuances of sentiment in short, text-based social media content. In a 2022 study by Li and Shujuanl ^[33][38], a hybrid model combining a bidirectional long short-term memory (BiLSTM) neural network and a convolutional neural network (CNN) showed remarkable improvements in accurately analysing text emotion compared with a single CNN model. Furthermore, the impact of Transformer-based models in sentiment analysis, as highlighted in a study by Mewada et al., in 2023 ^[34][39], remains significant. BERT (Bidirectional Encoder Representations from Transformers) and its variants continue to be influential in leveraging contextual embeddings. The incorporation of ensemble strategies, such as those discussed by Shah et al. ^[35][40], play an essential role in these contemporary hybrid frameworks by combining predictions from multiple models to improve sentiment classification accuracy. These recent articles collectively highlight the dynamic and innovative landscape of hybrid deep learning approaches in tweets sentiment analysis.

4. Ethical Considerations

In the field of sentiment analysis, ethical considerations play a key role in ensuring the responsible and fair application of this technology. One of the main concerns revolves around the potential biases present in the datasets used for training sentiment analysis models. These biases can stem from the data collection process, where certain demographics or viewpoints may be over-represented or under-represented, leading to biased results. Addressing these biases is crucial to prevent reinforcing existing stereotypes or perpetuating discrimination. In addition, the application of sentiment analysis in decision-making processes raises important ethical questions. Depending on the context, relying solely on sentiment analysis can have profound implications. Decisions based on sentiment analysis might inadvertently prioritise popular opinions over minority voices or fail to take into account the nuances and complexities of human emotions. Striking a balance between the perspectives offered by sentiment analysis and ethical considerations of fairness, transparency, and inclusivity is essential for ensuring that this technology is used responsibly and effectively. Engaging in these discussions is essential for a more comprehensive analysis of the ethical landscape surrounding sentiment analysis.