CCNet: Extracting High Quality Monolingual Datasets from Web Crawl Data

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4003–4012 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4003 CCNet: Extracting High Quality Monolingual Datasets from Web Crawl Data Guillaume Wenzek∗, Marie-Anne Lachaux∗, Alexis Conneau, Vishrav Chaudhary, Francisco Guzmán, Armand Joulin, Edouard Grave Facebook AI {guw, malachaux, aconneau, vishrav, fguzman, ajoulin, egrave}@fb.com Abstract Pre-training text representations have led to significant improvements in many areas of natural language processing. The quality of these models benefits greatly from the size of the pretraining corpora as long as its quality is preserved. In this paper, we describe an automatic pipeline to extract massive high-quality monolingual datasets from Common Crawl for a variety of languages. Our pipeline follows the data processing introduced in fastText (Mikolov et al., 2017; Grave et al., 2018), that deduplicates documents and identifies their language. We augment this pipeline with a filtering step to select documents that are close to high quality corpora like Wikipedia. Keywords: Common Crawl, web data 1. Introduction Pre-trained text representations have brought signifi- cant performance gains on many natural language pro- cessing tasks (Peters et al., 2018). Since the intro- duction of Transformers (Vaswani et al., 2017) and BERT (Devlin et al., 2018), we have a seen a steady improvement in the quality of these pre-trained mod- els, mainly driven by increasing the size of the pre- training corpora (Radford et al., 2019; Yang et al., 2019; Lan et al., 2019). Nonetheless, the size only does not guarantee better models and the quality of the data has to be preserved, which has lead to the use of ad-hoc datasets created by concatenating exist- ing high-quality data sources like Wikipedia. Unfortu- nately, such datasets cannot be replicated as easily for low-resources languages, as many have much smaller curated datasets such as Wikipedia. In this paper, we present a data collection pipeline that allows to gather massive monolingual corpora of high quality in a variety of languages, including many low-resource ones. The principles of our pipeline are general and we show the results of its application to data collected by the Common Crawl project.1 Com- mon Crawl is a massive non-curated dataset of web- pages in many languages, mixed together in tempo- ral snapshots of the web. Our pipeline performs stan- dard document deduplication and language identifica- tion similar to Grave et al. (2018), but differs in two ways: first, we preserve the document-level structure to allow for the training of paragraph-level represen- tations like BERT (Devlin et al., 2018) ; second, we add an optional monolingual filtering step that selects documents that are close to high quality sources, like Wikipedia. This is achieved by training a language model on the targeted sources and use the perplexity as a scoring function for documents. Our pipeline can be applied to any number of Common Crawl snapshots and takes 8.5 hours to process per snapshot on 5000 CPU cores. For example, the dataset obtained by pre- processing the February 2019 …

Content truncated to 500 words