This paper describes the participation of the ITST team in the FinancES 2023 shared task on financial targeted sentiment analysis in Spanish. This paper proposes a sentiment analysis approach based on psycholinguistic features which are obtained through the LIWC tool. Since the features provided by LIWC are many, the use of a feature selection technique based on Rough Set Theory and Information Gain is proposed to eliminate irrelevant features and thus improve the performance of the generated model. With respect to feature selection, a significant difference can be seen in terms of the LIWC categories selected for determining the sentiment polarity of each news headline towards both consumers and companies.
Sentiment analysis has become a relevant technology because it allows us to understand people's
opinions regarding various topics such as product preferences, marketing campaigns, politics, among
others. Today, there is a wealth of information published on the web related to the financial domain.
This represents a great opportunity to understand public opinion and generate solutions based on
sentiment analysis to support decision making. However, sentiment analysis in the financial context
represents a major challenge because the language used in this context is inherently complex since
financial terms refer to an underlying social, economic, and legal context
This paper concerns our participation at FinancES 2023 Task
Next section describes the developed strategies for identifying the main economic target from news headlines as well as for determining the sentiment polarity of each news headline towards both companies and consumers. Finally, under final remarks, we discuss our results and a proposal for future work.
2023 Copyright for this paper by its authors.
The process described above was implemented for both this task (Determining the sentiment polarity) and task 2 (Financial sentiment analysis at document level for companies and consumers). However, for each of the tasks, different categories were selected. With respect to task 1, the result of the feature selection process was 45 selected categories which are shown in Table 1. As can be seen, the dimensions of linguistic process and psychological process contain a greater number of selected categories. This may be because such categories are more related to the understanding of emotions such as anxiety, anger, sadness, positive, negative, among others, which are present in the expression of opinions.
and consumers
As mentioned above, Task 2 employed the process shown in Figure 1 Specifically, for the case of sentiment analysis for companies, the feature selection process resulted in a total of 23 selected LIWC categories (see Table 2). Again, the linguistic process (10) and psychological process (10) dimensions contain the largest number of categories. With respect to the psychological process, the categories of negative emotions (EmoNeg), anger (Enfado) and anxiety (Ansiedad) can be highlighted as the most discriminating characteristics in news headlines towards companies.
Categories Articulo, WPS, WC, Dic, Funct, Numeros, informal, Futuro, Conjunc,
BigWords, PronImp, ElElla, PronPer, TotPron, Cuantif Triste, Tiempo, Ansiedad, Inhib, Relativ, EmoNeg, MecCog
Hogar, Dinero, Trabajo
Comma, AllPunc
Finally, Figure 4 shows the Top 10 features for Task 2 sentiment analysis of news headlines towards consumers. As can be seen, there is a great difference between the characteristics selected for the analysis of news headline sentiments towards companies and consumers. One of the main differences lies in the fact that psycholinguistic features related to negative emotions such as negative emotion (EmoNeg) and anger (Anger) that were selected for the case of companies do not appear in the case of consumers. For the latter case there are a greater number of features belonging to the category of linguistic process with 6.
This paper presented a sentiment polarity detection approach for IBERLEF 2023 Task - FinancES shared task on financial targeted sentiment analysis in Spanish. For polarity detection, an approach was proposed that extracts psycholinguistic features from the opinions through the LIWC tool. In addition, a feature selection process was performed to identify and select the most relevant variables for the construction of the classifier model. As could be seen, the feature selection process produced different results for each of the targets (companies and consumers), which gives some clues as to the difference in the use of language for both cases. For future work, we intend to investigate transformers-based methods for sentiment polarity detection.
We are grateful to the Tecnológico Nacional de Mexico (TecNM, by its Spanish acronym) for supporting this work. This research was also sponsored by Mexico’s National Council of Humanities, Sciences and Technologies (CONAHCYT).