Dataset Card for CachacaNER

Dataset Description

Dataset Summary

CachacaNER is a new NER dataset in Portuguese, manually annotated, in the specific domain of the Brazilian drink called Cachaça. It is the first dataset labeled for the NER task in the beverage domain. The data was collected data from 24 e-commerce sites for the cachaça product and a total of 3,381 product web pages. The collected texts were broken by punctuation marks (.,?,!) resulting on 13,618 sentences, about 183K tokens and 24,039 entities, which is comparable to other manually labeled datasets for the Portuguese language.

The dataset is divided into training and test partitions, in order to facilitate the execution of experiments and compare the performance of NER algorithms. It contains 10 partitions, each has 100 documents (all the text corresponding to a product), which allows for cross-validation experiments. Furthermore, the first 7 partitions are marked as part of the training set, and the last 3 as part of the test set, in order to establish a baseline pattern for experiments with a single training and test run.

The dataset has a 17 different named entity classes (Drink Name, Alcoholic Graduation, Drink Classification, Distillation Equipment, Storage Time, Storage Container, Wood Type, Sensory Characteristic - Color, Sensory Characteristic - Aroma, Sensory Characteristic - Flavor, Sensory Characteristic - Consistency, Person Name, Organization Name, Location Name, Volume, Time, Price)

Languages

Portuguese - BR.

Data Fields

-id: id of the sample
-tokens: the tokens of the example text
-ner_tags: the NER tags of each token

The NER tags correspond to this list:

-'O','B-NOME_BEBIDA', 'I-NOME_BEBIDA', 'B-GRADUACAO_ALCOOLICA', 'I-GRADUACAO_ALCOOLICA', 'B-EQUIPAMENTO_DESTILACAO', 'I-EQUIPAMENTO_DESTILACAO', 'B-TEMPO_ARMAZENAMENTO', 'I-TEMPO_ARMAZENAMENTO', 'B-RECIPIENTE_ARMAZENAMENTO', 'I-RECIPIENTE_ARMAZENAMENTO', 'B-TIPO_MADEIRA', 'I-TIPO_MADEIRA', 'B-CARACTERISTICA_SENSORIAL_COR', 'I-CARACTERISTICA_SENSORIAL_COR', 'B-CARACTERISTICA_SENSORIAL_AROMA', 'I-CARACTERISTICA_SENSORIAL_AROMA', 'B-CARACTERISTICA_SENSORIAL_SABOR', 'I-CARACTERISTICA_SENSORIAL_SABOR', 'B-CARACTERISTICA_SENSORIAL_CONSISTÊNCIA', 'I-CARACTERISTICA_SENSORIAL_CONSISTÊNCIA', 'B-NOME_PESSOA', 'I-NOME_PESSOA', 'B-NOME_LOCAL', 'I-NOME_LOCAL', 'B-NOME_ORGANIZACAO', 'I-NOME_ORGANIZACAO', 'B-TEMPO', 'I-TEMPO', 'B-PRECO', 'I-PRECO', 'B-VOLUME', 'I-VOLUME', 'B-CLASSIFICACAO_BEBIDA', 'I-CLASSIFICACAO_BEBIDA',    

Dataset Creation

Source Data

The data was collected data from 24 e-commerce sites for the cachaça product and a total of 3,381 product web pages.

Initial Data Collection and Normalization

To create the CachacaNER dataset, we collected data from 24 e-commerce sites for the cachaça product. To collect and extract data from web pages, we develop web scraping scripts using the Beautiful Soup 2 and Selenium 3 APIs. In total, we collected data from 3,381 web pages in HTML format.

Data from the pages collected were extracted, structured and stored in a database.

The data preparation stage consisted of: (i) preprocessing the data; (ii) completing it with metadata; (iii) dividing the texts into sentences; (iv) selecting a subset of the data for manual labeling.

Annotations

The dataset labeling was done manually by three students (two undergraduates and one graduate), from UFLA, using the Doccano tool.

Annotation process

To guide the process, we have created a labeling guidelines document. The identification of entities related to sensory characteristics are the ones that most depend on the labeler’s perception. The orientation for labeling these entities was based on the Cachaça Sensory Wheel, proposed by Bortoletto (2016), and on the survey of sensory attributes developed in the work of Pinheiro (2010).

We used the Fleiss’ Kappa coefficient of agreement, proposed by Fleiss (1971), to assess the consistency of the labels. This metric measures the agreement among three or more raters when attributing categorical classifications to the same data set. As a result, the overall agreement among our three labelers was 0.857. The annotations in which there was agreement among the three labelers and those in which two of them agreed were directly inserted into the final dataset, using the majority vote strategy. The annotations with disagreement between the three labelers ware jointly reanalyzed and defined the final annotation that was inserted into the dataset.

Citation Information

@Article{Silva2023, author={Silva, Priscilla and Franco, Arthur and Santos, Thiago and Brito, Mozar and Pereira, Denilson}, title={CachacaNER: a dataset for named entity recognition in texts about the cacha{\c{c}}a beverage}, journal={Language Resources and Evaluation}, year={2023}, month={Jun}, day={17}, abstract={Named Entity Recognition (NER) is the task of identifying and classifying tokens in texts corresponding to a set of pre-defined categories, such as names of people, organizations and locations. Datasets labeled for this task are essential for training supervised machine learning models. Although there are many datasets labeled with texts for English, in the Portuguese language they are scarcer. This work contributes to the creation and evaluation of a manually labeled dataset for the NER task, with texts in Brazilian Portuguese, in the specific domain of the beverage called Cacha{\c{c}}a. This is a popular drink in Brazil, and of great economic importance. This is the first NER dataset in the beverage domain, and can be useful for other types of beverages with similar entity categories, such as wine and beer. We describe the process of data collection, creation of the dataset and its experimental evaluation. As a result, we created a dataset containing over 180,000 tokens labeled in 17 entity categories. The labeling obtained an agreement coefficient of 0.857 among the labelers, according to the Fleiss' Kappa metric, which is considered almost perfect. In our experimental evaluation, we obtained a micro-F1 value equal to 0.933 in the test set. The size of the dataset, as well as the result of its experimental evaluation, are comparable to other datasets in the Portuguese language, even though ours has a greater number of entity categories.}, issn={1574-0218}, doi={10.1007/s10579-023-09665-0}, url={https://doi.org/10.1007/s10579-023-09665-0} }