Sun, Tony
NL-Augmenter: A Framework for Task-Sensitive Natural Language Augmentation
Dhole, Kaustubh D., Gangal, Varun, Gehrmann, Sebastian, Gupta, Aadesh, Li, Zhenhao, Mahamood, Saad, Mahendiran, Abinaya, Mille, Simon, Srivastava, Ashish, Tan, Samson, Wu, Tongshuang, Sohl-Dickstein, Jascha, Choi, Jinho D., Hovy, Eduard, Dusek, Ondrej, Ruder, Sebastian, Anand, Sajant, Aneja, Nagender, Banjade, Rabin, Barthe, Lisa, Behnke, Hanna, Berlot-Attwell, Ian, Boyle, Connor, Brun, Caroline, Cabezudo, Marco Antonio Sobrevilla, Cahyawijaya, Samuel, Chapuis, Emile, Che, Wanxiang, Choudhary, Mukund, Clauss, Christian, Colombo, Pierre, Cornell, Filip, Dagan, Gautier, Das, Mayukh, Dixit, Tanay, Dopierre, Thomas, Dray, Paul-Alexis, Dubey, Suchitra, Ekeinhor, Tatiana, Di Giovanni, Marco, Gupta, Rishabh, Gupta, Rishabh, Hamla, Louanes, Han, Sang, Harel-Canada, Fabrice, Honore, Antoine, Jindal, Ishan, Joniak, Przemyslaw K., Kleyko, Denis, Kovatchev, Venelin, Krishna, Kalpesh, Kumar, Ashutosh, Langer, Stefan, Lee, Seungjae Ryan, Levinson, Corey James, Liang, Hualou, Liang, Kaizhao, Liu, Zhexiong, Lukyanenko, Andrey, Marivate, Vukosi, de Melo, Gerard, Meoni, Simon, Meyer, Maxime, Mir, Afnan, Moosavi, Nafise Sadat, Muennighoff, Niklas, Mun, Timothy Sum Hon, Murray, Kenton, Namysl, Marcin, Obedkova, Maria, Oli, Priti, Pasricha, Nivranshu, Pfister, Jan, Plant, Richard, Prabhu, Vinay, Pais, Vasile, Qin, Libo, Raji, Shahab, Rajpoot, Pawan Kumar, Raunak, Vikas, Rinberg, Roy, Roberts, Nicolas, Rodriguez, Juan Diego, Roux, Claude, S., Vasconcellos P. H., Sai, Ananya B., Schmidt, Robin M., Scialom, Thomas, Sefara, Tshephisho, Shamsi, Saqib N., Shen, Xudong, Shi, Haoyue, Shi, Yiwen, Shvets, Anna, Siegel, Nick, Sileo, Damien, Simon, Jamie, Singh, Chandan, Sitelew, Roman, Soni, Priyank, Sorensen, Taylor, Soto, William, Srivastava, Aman, Srivatsa, KV Aditya, Sun, Tony, T, Mukund Varma, Tabassum, A, Tan, Fiona Anting, Teehan, Ryan, Tiwari, Mo, Tolkiehn, Marie, Wang, Athena, Wang, Zijian, Wang, Gloria, Wang, Zijie J., Wei, Fuxuan, Wilie, Bryan, Winata, Genta Indra, Wu, Xinyi, Wydmański, Witold, Xie, Tianbao, Yaseen, Usama, Yee, M., Zhang, Jing, Zhang, Yue
Data augmentation is an important component in the robustness evaluation of models in natural language processing (NLP) and in enhancing the diversity of the data they are trained on. In this paper, we present NL-Augmenter, a new participatory Python-based natural language augmentation framework which supports the creation of both transformations (modifications to the data) and filters (data splits according to specific features). We describe the framework and an initial set of 117 transformations and 23 filters for a variety of natural language tasks. We demonstrate the efficacy of NL-Augmenter by using several of its transformations to analyze the robustness of popular natural language models. The infrastructure, datacards and robustness analysis results are available publicly on the NL-Augmenter repository (\url{https://github.com/GEM-benchmark/NL-Augmenter}).
BOLD: Dataset and Metrics for Measuring Biases in Open-Ended Language Generation
Dhamala, Jwala, Sun, Tony, Kumar, Varun, Krishna, Satyapriya, Pruksachatkun, Yada, Chang, Kai-Wei, Gupta, Rahul
Recent advances in deep learning techniques have enabled machines to generate cohesive open-ended text when prompted with a sequence of words as context. While these models now empower many downstream applications from conversation bots to automatic storytelling, they have been shown to generate texts that exhibit social biases. To systematically study and benchmark social biases in open-ended language generation, we introduce the Bias in Open-Ended Language Generation Dataset (BOLD), a large-scale dataset that consists of 23,679 English text generation prompts for bias benchmarking across five domains: profession, gender, race, religion, and political ideology. We also propose new automated metrics for toxicity, psycholinguistic norms, and text gender polarity to measure social biases in open-ended text generation from multiple angles. An examination of text generated from three popular language models reveals that the majority of these models exhibit a larger social bias than human-written Wikipedia text across all domains. With these results we highlight the need to benchmark biases in open-ended language generation and caution users of language generation models on downstream tasks to be cognizant of these embedded prejudices.