Automatic Identification of Economic Activities in Complaints

Complaint Analysis and Classification for Economic and Food Safety

“…As such, the test set used in these experiments has been drawn from the last 5 years of data only. This also ensures the results for the task of economic activity prediction reported in this paper can be compared to results from ear-lier work (Barbosa et al, 2019), which were obtained using the same test set. A total of roughly 25,000 examples make up this test set, 16% of all available data.…”

“…Based on an earlier work that tackled economic activity prediction on a smaller sample of this dataset (Barbosa et al, 2019), the dataset was preprocessed using the Natural Language Toolkit (NLTK) (Bird et al, 2009) to perform tokenization, lemmatization and remove stop words from Portuguese text. Furthermore, from among the different featurebased representations explored by Barbosa et al (2019), a TF-IDF weighted vector was found to be the most effective method of representing each document. TF-IDF outperformed fastText-based (Joulin et al, 2016) and BERT-based (Devlin et al, 2018) representations, using traditional machine learning approaches, specifically Support Vector Machines (SVM) (Cortes and Vapnik, 1995).…”

“…As mentioned previously, and plainly observable in Table 1, this classification problem is very imbalanced. In previous work (Barbosa et al, 2019), while considering a sample of the dataset with half the time window (and thus with approximately half the size, while maintaining a similar class distribution), we have tried employing both random undersampling and random oversampling (He and Garcia, 2009), in order to improve the overall classification performance and, more specifically, the performance on minority classes. However, such attempts did not succeed, consistently worsening results.…”

Proceedings of the Second Workshop on Economics and Natural Language Processing

“…In order to implement machine learning classifiers based on the textual contents of each complaint, and given their user-generated content nature, a previous preprocessing step was necessary. Based on an earlier work that tackled economic activity prediction on a smaller sample of this dataset (Barbosa et al, 2019), the dataset was preprocessed using the Natural Language Toolkit (NLTK) (Bird et al, 2009) to perform tokenization, lemmatization and remove stop words from Portuguese text. Furthermore, from among the different featurebased representations explored by Barbosa et al (2019), a TF-IDF weighted vector was found to be the most effective method of representing each document.…”

“…Based on an earlier work that tackled economic activity prediction on a smaller sample of this dataset (Barbosa et al, 2019), the dataset was preprocessed using the Natural Language Toolkit (NLTK) (Bird et al, 2009) to perform tokenization, lemmatization and remove stop words from Portuguese text. Furthermore, from among the different featurebased representations explored by Barbosa et al (2019), a TF-IDF weighted vector was found to be the most effective method of representing each document. TF-IDF outperformed fastText-based and BERT-based representations, using traditional machine learning approaches, specifically Support Vector Machines (SVM) (Cortes and Vapnik, 1995).…”