Deep Learning Approach for Aspect-Based Sentiment Analysis of Restaurants Reviews in Spanish

This paper presents a corpus of pre-processed Mexican laws for computational tasks. The main contributions are the proposed JSON structure and the methodology used to achieve the semi-structured corpus with the selected algorithms. Law PDF documents were transformed into plain text, unified by a deconstruction of law–document structure, and labeled with natural language processing techniques considering part of speech (PoS); a process of entity extraction was also performed. The corpus includes the Mexican constitution and the Mexican laws that were collected from the official site in PDF format repealed before 14 October 2021. The collection has 305 documents, including: the Mexican constitution, 289 laws, 8 federal codes, 3 regulations, 2 statutes, 1 decree, and 1 ordinance. The semi-structured database includes the transformation of the set of laws from PDF format to a digital representation in order to facilitate its computational analysis. The documents were migrated to JSON type files to represent internal hierarchical relations. In addition, basic natural language processing techniques were implemented on laws for the identification of part of speech and named entities. The presented data set is mainly useful for text analysis and data science. It could be used for various legislative analysis tasks including: comprehension, interpretation, translation, classification, accessibility, coherence, and searches. Finally, we present some statistic of the identified entities and an example of the usefulness of the corpus for environmental laws.

Evaluating the Performance of Large Language Models for Spanish Language in Undergraduate Admissions Exams

In this article, we propose a method for the automatic retrieval of a set of semantic primitive words from an explanatory dictionary and a novel evaluation procedure for the obtained set of primitives. The approach is based on the representation of the dictionary as a directed graph with a single-objective constrained optimization problem via a genetic algorithm with the PageRank scoring model. The problem is defined as a subset selection. The algorithm is fit to search for the sets of words that should fulfil several requirements: the cardinality of the set should not exceed empirically selected limits and the PageRank word importance score is minimized with cycle prevention thresholding. In the experiments, we used the WordNet dictionary for English. The proposed method is an improvement over the previous state-of-the-art solutions.