Paraphrase detection is a Natural-Language Processing (NLP) task that aims at automatically identifying whether two sentences convey the same meaning (even with different words). For the Portuguese language, most of the works model this task as a machine-learning solution, extracting features and training a classifier. In this paper, following a different line, we explore a graph structure representation and model the paraphrase identification task over a heterogeneous network. We also adopt a back-translation strategy for data augmentation to balance the dataset we use. Our approach, although simple, outperforms the best results reported for the paraphrase detection task in Portuguese, showing that graph structures may capture better the semantic relatedness among sentences.
This paper analyzed the robustness of a state-of-the-art Automated Essay Scoring (AES) model by applying various linguistically motivated perturbations to the Essay-BR corpus. Our findings reveal that the AES model failed to detect these adversarial modifications, often assigning higher scores to the disturbed essays than to the original ones.
29 July 2020 in iSys - Brazilian Journal of Information Systems
This paper presents a new approach to predict the helpfulness of opinions. Usually, researchers in this area use tables of attribute-value to aggregate the features that represent the evaluated texts. Although that representation is common, it considers that the objects are independent. We argue that among the discriminant factors of the helpfulness of opinions, there are dependent factors of the relationship among the opinion-forming elements. Thus, we modeled this task as a network, considering the information of relations among objects in the network (comments, stars, and words). A regularization technique of graphs is used to extract the relevant features of graph structure and, after that, the comments are classified as helpful or unhelpful. We compared our network model with two baselines methods, one based on fuzzy logic and another based on Neural Networks. Our model outperformed the fuzzy logic and Neutal Network methods in 0.17 and 0.19 of F-measure, respectively. The main advantages of our approach are that few data are necessary to helpfulness classification and the relationships may help in the understanding the classification, explaining the reasons for a determinate classification.
The evolution of e-commerce has contributed to the increase of the information available, making the task of analyzing the reviews manually almost impossible. Due to the amount of information, the creation of automatic methods of knowledge extraction and data mining has become necessary. Currently, to facilitate the analysis of reviews, some websites use filters such as votes by the utility or by stars. However, the use of these filters is not a good practice because they may exclude reviews that have recently been submitted to the voting process. One possible solution is to filter the reviews based on their textual descriptions, author information, and other measures. This chapter has a propose of approaches to estimate the importance of reviews about products and services using fuzzy systems and artificial neural networks. The results were encouraging, obtaining better results when detecting the most important reviews, achieving approximately 82% when f-measure is analyzed.