--- license: cc-by-4.0 configs: - config_name: default data_files: - split: train path: data/train-* - split: validation path: data/validation-* - split: test path: data/test-* - config_name: mcqa data_files: - split: train path: mcqa/train-* - split: validation path: mcqa/validation-* - split: test path: mcqa/test-* - config_name: paraphrases data_files: - split: train path: paraphrases/train-* dataset_info: - config_name: default features: - name: id_qa dtype: string - name: corpus dtype: int64 - name: question_en_origin dtype: string - name: question_pt_origin dtype: string - name: question_en_paraphase dtype: string - name: question_pt_paraphase dtype: string - name: answer_en_origin dtype: string - name: answer_pt_origin dtype: string - name: answer_en_validate dtype: string - name: answer_pt_validate dtype: string - name: abstract dtype: string - name: eid_article_scopus dtype: string - name: question_generic dtype: float64 - name: answer_in_text dtype: float64 - name: answer_difficulty dtype: float64 - name: question_meaningful dtype: float64 - name: answer_equivalent dtype: float64 - name: question_type dtype: string - name: abstract_translated_pt dtype: string - name: pt_question_translated_to_en dtype: string - name: at_labels dtype: float64 splits: - name: train num_bytes: 8002269 num_examples: 1806 - name: validation num_bytes: 994524 num_examples: 225 - name: test num_bytes: 940555 num_examples: 227 download_size: 3976683 dataset_size: 9937348 - config_name: mcqa features: - name: id dtype: string - name: text dtype: string - name: question dtype: string - name: A dtype: string - name: B dtype: string - name: C dtype: string - name: D dtype: string - name: E dtype: string - name: correct dtype: string - name: alternative dtype: string splits: - name: train num_bytes: 4327619 num_examples: 1798 - name: validation num_bytes: 582526 num_examples: 225 - name: test num_bytes: 551723 num_examples: 227 download_size: 2148096 dataset_size: 5461868 - config_name: paraphrases features: - name: question_AUT_EN_1 dtype: string - name: question_AUT_EN_2 dtype: string - name: answer_AUT_EN_1 dtype: string - name: answer_AUT_EN_2 dtype: string - name: question_AUT_PT_1 dtype: string - name: question_AUT_PT_2 dtype: string - name: answer_AUT_PT_1 dtype: string - name: answer_AUT_PT_2 dtype: string splits: - name: train num_bytes: 1175020 num_examples: 1806 download_size: 720519 dataset_size: 1175020 --- # Pirá: A Bilingual Portuguese-English Dataset for Question-Answering about the Ocean, the Brazilian coast, and climate change Pirá is a crowdsourced reading comprehension dataset on the ocean, the Brazilian coast, and climate change. QA sets are presented in both Portuguese and English, together with their corresponding textual context. The dataset also contains human and automatic paraphrases for questions and answers, as well as a number of qualitative assessments. The original paper was published at CIKM'21 and can be found [here](https://dl.acm.org/doi/pdf/10.1145/3459637.3482012). As a subsequent project, we have produced a curated version of the dataset, which we refer to as Pirá 2.0. In this step, we have also defined a number of benchmarks and reported the corresponding baselines. This is the version that we make available at HuggingFace. Pirá 2.0's preprint is available in [Arxiv](https://arxiv.org/abs/2309.10945). Pirá is, to the best of our knowledge, the first QA dataset with supporting texts in Portuguese, and, perhaps more importantly, the first bilingual QA dataset that includes Portuguese as one of its languages. Pirá is also the first QA dataset in Portuguese with unanswerable questions so as to allow the study of answer triggering. Finally, it is the first QA dataset that tackles scientific knowledge about the ocean, climate change, and marine biodiversity. More information on the Methodology, Dataset versions, Benchmarks can be found on the project's [Github page](https://github.com/C4AI/Pira/). You can also find there the Multiple-Choice version of Pirá. # Dataset The dataset is split into train, validation, and test sets. | Split | Size | #QAs | |---|---|---| | Training | 80% | 1806 | | Validation | 10% | 225 | | Test | 10% | 227 | | Full dataset | 100% | 2258 | Above is an example of a question-answer set from Pirá: ``` { "id_qa": "B2142", "corpus": 2, "question_en_origin": "What are the proportion of men and women employed in the fishery sector worlwide?", "question_pt_origin": "Qual é a proporção de homens e mulheres empregados no setor pesqueiro em todo o mundo?", "question_en_paraphase": "Which share of the fishery sector workers of the world are women?", "question_pt_paraphase": "Qual parcela dos trabalhadores do setor da pesca no mundo são mulheres?", "answer_en_origin": "85 per cent men and 15 per cent women.", "answer_pt_origin": "85 por cento homens e 15 por cento mulheres.", "answer_en_validate": "It is estimated that more than fifteen per cent of the fishing sector workers are women.", "answer_pt_validate": "Estima-se que mais de quinze por cento dos trabalhadores do setor da pesca são mulheres.", "eid_article_scopus": "", "text_excerpts_un_reports": "Distribution of ocean benefits and disbenefits Developments in employment and income from fisheries and aquaculture The global harvest of marine capture fisheries has expanded rapidly since the early 1950s and is currently estimated to be about 80 million tons a year. That harvest is estimated to have a first (gross) value on the order of 113 billion dollars. Although it is difficult to produce accurate employment statistics, estimates using a fairly narrow definition of employment have put the figure of those employed in fisheries and aquaculture at 58.3 million people (4.4 per cent of the estimated total of economically active people), of which 84 per cent are in Asia and 10 per cent in Africa. Women are estimated to account for more than 15 per cent of people employed in the fishery sector. Other estimates, probably taking into account a wider definition of employment, suggest that capture fisheries provide direct and indirect employment for at least 120 million persons worldwide. Small-scale fisheries employ more than 90 per cent of the world’s capture fishermen and fish workers, about half of whom are women. When all dependants of those taking full- or part-time employment in the full value chain and support industries (boatbuilding, gear construction, etc.) of fisheries and aquaculture are included, one estimate concludes that between 660 and 820 million persons have some economic or livelihood dependence on fish capture and culture and the subsequent direct value chain. No sound information appears to be available on the levels of death and injury of those engaged in capture fishing or aquaculture, but capture fishing is commonly characterized as a dangerous occupation. Over time, a striking shift has occurred in the operation and location of capture fisheries. In the 1950s, capture fisheries were largely undertaken by developed fishing States. Since then, developing countries have increased their share. As a broad illustration, in the 1950s, the southern hemisphere accounted for no more than 8 per cent of landed values. By the last decade, the southern hemisphere’s share had risen to 20 per cent. In 2012, international trade represented 37 per cent of the total fish production in value, with a total export value of 129 billion dollars, of which 70 billion dollars (58 per cent) was exports by developing countries. Aquaculture is responsible for the bulk of the production of seaweeds. Worldwide, reports show that 24.9 million tons was produced in 2012, valued at about 6 billion dollars. In addition, about 1 million tons of wild seaweed were harvested. Few data were found on international trade in seaweeds, but their culture is concentrated in countries where consumption of seaweeds is high.", "question_generic": false, "answer_in_text": true, "answer_difficulty": 1, "question_meaningful": 5, "answer_equivalent": 5, "question_type": "None of the above" } ``` # Automatic Paraphrases As we have only generated automatic paraphrases for questions and answers in the train set, they had to be saved in a different Dataset file. To download the automatic paraphrases, just run: ``` paraphrases = load_dataset("paulopirozelli/pira", "paraphrases") ``` # Multiple Choice Question Answering We have also developed a multiple choice question answering version of Pirá 2.0. To download the automatic paraphrases, just run: ``` mcqa = load_dataset("paulopirozelli/pira", "mcqa") ```