Brazil

This work presents a new database containing high sampling
rate recordings of a single male speaker reading sentences in Brazilian
Portuguese with neutral voice, along with the corresponding text corpus. Intended for synthesis and other speech-oriented applications, the
dataset contains text scripts extracted from a popular Brazilian news TV
program, read out loud by a trained individual in a controlled environment, resulting in roughly 20 h of audio data. The text was normalized
in the recording process and special textual occurrences (e.g. acronyms,
numbers, foreign names etc.) were replaced by their phonetic translation to a readable text in Portuguese. There are no noticeable accidental
sounds and background noise has been kept to a minimum in all audio
samples. To illustrate the potential benefits of having this data available,
text-to-speech experiments were conducted using state-of-the-art models
for speech synthesis (Tacotron 2 and Waveglow). As a result, we obtained
intelligible and natural sounding voices from as few as 8 minutes of audio samples coming from an unseen target speaker, after having trained
over our data; moreover, by increasing the target recording time to 75
minutes, we have noticeably improved accuracy in pronunciation.


PROPOR 2022

A Corpus of Neutral Voice Speech in Brazilian Portuguese

speech syntesis

brazilian portuguese

audio signal processing

automatic speech recognition

dataset

The International Conference on Computational Processing of Portuguese (PROPOR) is the main event in the area of human language processing that is focused on theoretical and technological issues of written and spoken Portuguese and Galician. The meeting has been a very rich forum for the exchange of ideas and partnerships for the research and industry communities dedicated to the automated processing of Portuguese, promoting the development of methodologies, resources and projects that can be shared among researchers and practitioners in the field.

Find more information and pay the registration fee on [event organizer’s website](https://sites.universidadedefortaleza.com/propor2022/?page_id=401)

The International Conference on Computational Processing of Portuguese (PROPOR) is the main event in the area of human language processing that is focused on theoretical and technological issues of written and spoken Portuguese and Galician.

technical paper

Named entity recognition (NER) tasks imply token-level labels. Annotating documents can be time-consuming, costly, and prone
to human error. In many real-life scenarios, the lack of labeled data has
become the biggest bottleneck preventing NER being effectively used in
some domains and with some natural languages, with negative impacts
on the quality of some tasks. To overcome the barrier of the lack of annotated data for new application domains in some natural languages,
we propose a method that uses the output of an ensemble of NER’s to
automatically annotate the data needed to train a Bidirectional Encoder
Representations from Transformers (BERT) based NER for Portuguese.
The performance was assessed using MiniHAREM dataset with promising results. For domain relevant classes such as LOCAL, F1, Precision
and Recall above 50% were obtained when training only with automatically annotated data.

Named Entity Extractors for New Domains by Transfer Learning with Automatically Annotated Data

Open Information Extraction (OpenIE) is the task of extracting structured information from text. Recent advances in applying Deep Learning to OpenIE tasks have improved the state of the art for the task, although few works have been produced for languages other than English. In this work, we propose PortNOIE, a neural framework for open information extraction for the Portuguese language. We evaluate our method on a manually annotated corpus of Open IE extractions, obtaining better performance than the current state of the art for OpenIE for Portuguese, both based on rule-based approaches or neural methods.

PortNOIE: A neural framework for Open Information Extraction for the Portuguese language

Tax evasion is a problem that affects our society, costing billions of Brazilian reais of public funds every year. Stopping this practice is a complex challenge that involves analyzing a large and diverse volume of data. In this work, we propose an approach to analyze invoices and extract information about measures and units from product descriptions using a neural network with the BiLSTM-CRF architecture. Our method can validate product quantity information to, for instance, check whether any product was bought or sold by a business without issuing an invoice. The results were evaluated according to precision, recall, and f-score. The proposed approach can correctly detect more than 90% of cases of each type of information, showing its feasibility to process invoice data.

Automatic Recognition of Units of Measurement in Product Descriptions from Tax Invoices Using Neural Networks

Automatic book genre classification is a hard task as it requires the whole book's content or a high-quality summary, which is challenging to write automatically. On the other hand, online reviews are an accessible resource for readers to evaluate a book or even get a general sense about it, including its genre. As the amount of book reviews is always increasing, using such information to genre classification needs a robust solution to deal with high volumes of data. In such a context, we introduce a model for automatically classifying book genres by analyzing online text reviews. We build a dataset of compiled texts from online book reviews. Then, we use multiple machine learning algorithms to categorize a book into a specific genre. Such a process enables to compare algorithms and detect the best classifiers. Hence, the most efficient machine learning algorithm completed the task with an accuracy of 96%; i.e., the proposed model is convenient for various information retrieval systems due to its high certainty and efficiency.

Book Genre Classification Based on Reviews of Portuguese-language Literature

The amount of legislative documents produced within the past decade has risen dramatically, making it diﬃcult for law practitioners to consult and update legislation. Named Entity Recognition (NER) systems have the untapped potential to extract information from legal documents, which can improve information retrieval and decision-making processes. We introduce the UlyssesNER-Br, a corpus of Brazilian Legislative Documents for NER with quality baselines. The presented corpus consists of bills and legislative consultations from Brazilian Chamber of Deputies. We implemented Conditional Random Field (CRF) and Hidden Markov Model (HMM) models, and the promising F1-score of 80.8% in the analysis by categories and 81.04% in the analysis by types, was achieved with the CRF model. The entities with the best average F1-score results were “FUNDlei” and “DATA”, and the ones with the worst results were “EVENTO” and “PESSOAgrupoind”. The corpus was also evaluated using a BiLSTM-CRF and Glove architectures provided by the pioneering state-of-the-art paper, achieving F1-score of 76.89% in the analysis by categories and 59.67% in the analysis by types.

UlyssesNER-Br: a corpus of Brazilian Legislative Documents for Named Entity Recognition

We present PetroBERT, a BERT model adapted to the oil and gas domain.

PetroBERT: a domain adaptation language model for oil and gas applications in Portuguese

Dependency parsing can enhance the performance of Named Entity Recognition (NER) models and can be leveraged to boost information extraction. NER tasks are essential to deal with clinical narratives, but models for Brazilian Portuguese dependency parsing are scarce, even less for clinical texts and their specificities. This paper reports on the development of a treebank of clinical narratives in Brazilian Portuguese and the drafting of guidelines. Based on a corpus of 1,000 clinical narratives manually annotated with semantic information, split into 12,711 sentences, we identified some characteristics of these texts that differ from traditional domains and have a deep impact on the annotation process, such as extensive use of acronyms and abbreviations, words not recognized by POS taggers, misspelling, special use of some symbols, different uses for numerals, heterogeneity of sentence sizes, and coordinated phrases without any punctuation. We developed a document to describe the annotation types and to explain how difficult cases should be treated to ensure consistency, including examples that could be found in this kind of text. We created a Tag versus Frequency relation to justify some of the characteristics and challenges of the corpus. The corpus when completely annotated will be made available to the entire scientific community that performs research with clinical texts.

Challenges in annotating a treebank of clinical narratives in Brazilian Portuguese

This paper aims to summarize the key points and results of an analysis of the alignment of gesture and prosodic boundaries collected through manual annotation and automated treatment. For this purpose, we used a multimodal corpus of spontaneous speech, the BGEST corpus (Barros, 2021), part of the C-ORAL-BRASIL language resources (Moneglia and Cresti, 2005; Raso and Mello, 2012). Gestures and prosodic boundaries were manually annotated with ELAN. The gestural an-notation followed partly the Linguistic Annotation System for Gestures guidelines (Bressem et al., 2013), whereas the prosodic boundaries followed the perceptual criterion adopted in other C-ORAL resources. The tabular data outputted by ELAN was then analyzed by a script that compares the alignment of gesture phrases and intonation units. The results point to a larger gestural phrase that encompasses the intonation unit, while the stroke is almost always within the intonation unit. Around 85 percent of all information units align with a gesture. The analysis also points that the time values of alignment are shorter than analyzed by Loehr (2004) both for initial as well for final boundaries and more complex than proposed by Cantalini (2018). This indicates that more research is needed before we can set an approximate value for the overlap of larger units.

A protocol for comparing gesture and prosodic boundaries in multimodal corpora

Palatal consonants in Portuguese are considered complex or marked segments because they are inherently heavy and restricted in terms of their distribution, in relation to other consonants. Moreover, they appear to display differences between themselves, as first language acquisition and creoles' adaptation suggest that /ʎ/ is more complex than /ɲ/. The arguments for complexity are endorsed by some qualitative studies but are still lacking quantitative support. This paper aims at analyzing the phonological restrictiveness of these consonants by comparing their actual frequency in several different corpora, reporting both lexical entries and usage in discourse. In addition to their context-free frequency, we control for their word position and phonetic adjacency. We find that palatals are less frequent than other consonants. However, relative to each other, they do not display proportional lexical and usage frequencies. These results shed new light not only on the representation of /ɲ/ and /ʎ/but also on the relation between frequency and markedness in language studies.

Comparing lexical and usage frequencies of palatal segments in Portuguese

Forced phonetic alignment (FPA) is the task of assessing the time boundaries of phonetic units, i.e., calculating when in the speech utterance a certain phoneme starts and ends. This paper describes experiments on FPA for Brazilian Portuguese using Kaldi toolkit. Based on time-delay neural networks (TDNN), several acoustic models were trained on the top of the combination between hidden Markov models (HMM) and Gaussian mixture models (GMM). The nature of the input features and the topology of the HMMs have been varied in order to analyze each one's influence. Results with respect to the phone boundary metric over a dataset of 385 hand-aligned utterances show that the network is mostly invariant to the input features, while regular HMM topologies do perform better in comparison to a modified version used in chain models. Conversely, the neural network still does not outperform GMM models for phonetic alignment.


Premium content

Downloads

Next from PROPOR 2022

Named Entity Extractors for New Domains by Transfer Learning with Automatically Annotated Data

Similar lecture

Brazilian Portuguese Speech Recognition Using Wav2vec 2.0