Brazil

In this paper, we show how a Portuguese BERT model can be combined with structured data in order to deploy a chatbot based on a finite state machine to create a conversational AI system that helps a real-estate company to predict its client’s contact motivation. The model achieves human level results in a dataset that containts 235 unbalanced labels. Then, we also show its benefits considering the business impact comparing it against classical NLP methods.

BRACIS 2021

Augmenting Customer Support with an NLP-based Receptionist

applied nlp

customer support

chatbot

**BRACIS** originated from the combination of the two most important scientific events in Brazil in Artificial Intelligence (AI) and Computational Intelligence (CI): the Brazilian Symposium on Artificial Intelligence – **SBIA** (22 editions), and the Brazilian Symposium on Neural Networks – **SBRN** (13 editions). **ENIAC** is a national forum for researchers, practitioners, educators and students to present and discuss innovations, trends, experiences, developments and work in progress in the fields of Artificial and Computational Intelligence. **STIL** is the bi-annual Language Technology event.
All events are supported by the Brazilian Computer Society – Special Interest Groups: **CE-IA**, **CE-IC** and **CE-PLN**.

To access the 10th Brazilian Conference on Intelligent Systems (**BRACIS**), you need to register for the conference.  Please visit the Conference official site to register

[Register now](http://c4ai.inova.usp.br/bracis2021/)

Register now

BRACIS originated from the combination of the two most important scientific events in Brazil in Artificial Intelligence (AI) and Computational Intelligence (CI): the Brazilian Symposium on Artificial Intelligence – SBIA (22 editions), and the Brazilian Symposium on Neural Networks – SBRN (13 editions). 
ENIAC is a national forum for researchers, practitioners, educators and students to present and discuss innovations, trends, experiences, developments and work in progress in the fields of Artificial and Computational Intelligence. 
STIL is the bi-annual Language Technology event.
All events are supported by the Brazilian Computer Society – Special Interest Groups: CE-IA, CE-IC and CE-PLN

technical paper

Este trabalho objetiva apresentar a descrição e a modelagem linguístico-computacional das construções interrogativas QU- do Português brasileiro no Constructicon da FrameNet Brasil.

Modelagem de Construções Interrogativas QU- no Constructicon da FrameNet Brasil

Apresentamos a proposta de descrição e análise de dados verbais do português brasileiro presente no banco de dados VerboWeb. Os verbos foram agrupados e classificados de acordo com o arcabouço teórico-metodológico da Semântica Lexical, que assume que certas propriedades semânticas dos verbos determinam seu comportamento sintático. Mostramos como o banco de dados é estruturado e as generalizações que podemos fazer sobre o léxico verbal de nossa língua.

Banco de dados VerboWeb: um panorama do léxico verbal do PB

In this study, linguistic features were listed for classification of triples,based on a parallel corpus in Galician, Brazilian Portuguese (PT-BR) and European Spanish (EE). In the experiments, at the syntactic level, a relevant per-formance of features that offer greater difficulty to extract valid triples was ob-served, such as the corelated null object/subject and the verb-subject inversion,as well as relational triples that form ungrammatical sub-sentences. At the mor-phological level, it was observed that the grammatical class of the initial wordof each sub-sentence, especially when they are prepositions, were relevant forthe classification of the triples


Engenharia de features linguísticas para classificação de triplas relacionais

This article describes a methodology that was developed to review Emocionário, a lexicon of emotion words in Portugese from Linguateca's AC/DC project. This first portion of the review was carried out in seven of the 24 groups of Emocionário, being them “Desespero”, “Esperança”, “Humildade”, “Pena”, “Satisfação”, “Saudade” and “Surpresa”. The methodology is then illustrated, making use of examples taken from the corpus itself for each of the steps. Some of the main changes are documented and discussed at the end.

Descrição de uma metodologia desenvolvida para revisão semiautomática de um léxico de palavras de emoção

Um estudo exploratório de exposição de participantes às variantes de uma variável linguística saliente do ponto de vista social, a palatalização progressiva, foi realizado com o uso de rastreamento ocular, para examinar o processamento da variação linguística. Os resultados mostram que exposição à variante estigmatizada captou a atenção e aumentou a dilatação da pupila dos participantes, o que pode ser interpretado como evidência de uma resposta emocional.

Respostas emocionais da variação linguística: Análise exploratória de rastreio ocular

This article presents an analysis of the textual complexity of satirical and true news for Brazilian Portuguese. The Fake has been a big problem nowadays. Satirical content is an important point in the automatic detection of false news as its use can cause underlying confusion in the analysis. To carry out this research, the NILC-Metrix tool was applied, and 16 measures were evaluated, including descriptive, syntactic and semantic aspects, noting a greater complexity for real texts.

Complexidade textual em notícias satíricas: uma análise para o português do Brasil

This work describes the syntactic organization of a set of Judgments of the Special Civil Courts, whose documents must be accessible to the lay citizen without the assistance of a lawyer. The phrasal constituents with subject function in 110 Judgments are highlighted in the light of Terminology studies, Corpus Linguistics and research on reading comprehension. The study found that most of these constituents assume forms that tend to demand greater burden in reading and to exhibit traits that hardly coincide with patterns of written daily language, which makes their intelligibility difficult for the lay public.

Constituintes frasais com Função de Sujeito em Sentenças Judiciais

As large amounts of unstructured data are generated on a regular basis, expressing or storing knowledge in a way that is useful remains a challenge. In this context, Relation Extraction (RE) is the task of automatically identifying relationships in unstructured textual data. Thus, we investigated the relation extraction on unstructured e-commerce data from the smartphone domain, using a BERT model fine-tuned for this task. We conducted two experiments to acknowledge how much relational information it is possible to extract from product sheets (structured data) and product titles (unstructured data), and a third experiment to compare both. Analysis shows that extracting relations within a title can retrieve correct relations that are not evident on the related sheet.

Relation extraction in structured and unstructured data:a comparative investigation on smartphone titles in thee-commerce domain

O aprendizado multimodal visa explorar as características das diversas modalidades (texto, imagem, áudio) para gerar modelos computacionais. No comercio eletrônico, devido à grande variedade das características dos produtos e a ausência ou inconsistência de informações, a combinação de informações de modos diferentes vem a ser bastante adequada. Neste trabalho são apresentados alguns experimentos para a classificação multimodal (texto e imagem) de produtos (produtos adultos) que não podem ser vendidos no marketplace da empresa parceira. Nesses experimentos, redes neurais foram usadas para treinar classificadores uni e multimodal. O classificador multimodal atingiu 99% de F1 contra 98% do modelo textual e 94% do visual.

Classificacão multimodal de produtos que não podem ser vendidos no marketplace

Textual similarity deals with determining how similar two pieces of texts are, considering the lexical (surface forms) or semantic (meaning) close-ness. In this paper we applied word embeddings for measuring e-commerce product title similarity in Brazilian Portuguese. We generated some domain-specific word embeddings (using Word2Vec, FastText and GloVe) and compared them with general-domain models (word embeddings and BERT models). We concluded that the cosine similarity calculated using the domain-specific word embeddings was a good approach to distinguish between similar and non-similar products, but the multilingual BERT pre-trained model proved to be thebest one.

Premium content

Augmenting Customer Support with an NLP-based Receptionist

Downloads

Next from BRACIS 2021

Modelagem de Construções Interrogativas QU- no Constructicon da FrameNet Brasil

Similar lecture

A Conversational Agent to Support Hospital Bed Allocation