Banca de DEFESA: João Laterza

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : João Laterza
DATE: 10/02/2023
TIME: 08:00
LOCAL: Remoto via Teams
TITLE:

Automated classification of complaints opened by customers and users of the National Financial System

KEY WORDS:

BCB, SFN, BERT, LSTM, text classification, complaints

PAGES: 132
BIG AREA: Ciências Exatas e da Terra
AREA: Ciência da Computação
SUBÁREA: Metodologia e Técnicas da Computação
SUMMARY:

The Central Bank of Brazil (BCB) is responsible for assisting customers and users of the National Financial System (SFN) with complaints against products and services offered by its supervised entities and ensuring that all demands are appropriately addressed. Each concern is manually handled and classified as “proceeding” or “unfounded” based on a preliminary analysis of the facts described by the customer, the alleged entity’s reply, the attached documents, and according to its compliance with current regulations. However, dealing with an increasing demand with the available resources has turned out to be an unprecedented challenge, being currently impossible to handle all issues accordingly. In this context, the BCB has developed an automated solution to filter complaints more likely to be classified as “proceeding”, thus driving the human activities of examination, analysis and judgment. The present study aims to propose a new classifier for this task with better performance than the current model. To this end, deep learning approaches were explored, differing from the traditional method previously employed. As a result, an experimental classifier was tested, based on a hierarchical structure, and combining Bidirectional Encoder Representations from Transformers (BERT) with Bidirectional Long Short-Term Memory (BiLSTM) to generate unique contempt representations of both the citizen’s complaint and the entity’s reply. In cross-validation, this solution reached an average performance of 71.41%, based on the area under the precision-recall curve (PRAUC), exceeding the current BCB model by 0.96 percentage points, and performing 71.92% on the test dataset. If the proposed classifier was adopted to drive the task of handling complaints, we estimate that approximately 90.23% of the proceeding demands could be identified by evaluating only 60% of the total amount. In addition, multimodal strategies were tested to combine the described textual representations with tabular features of the original classifier. However, when compared to the previous proposed solution, the achieved gain with the multimodal strategies did not reach a statistically significant outcome.

BANKING MEMBERS:
Interno - 3064724 - GLAUCO VITOR PEDROSA
Interno - 3128249 - LUIS PAULO FAINA GARCIA
Externa à Instituição - NÁDIA FELIX FELIPE DA SILVA - USP
Presidente - 1821656 - THIAGO DE PAULO FALEIROS