WANLI: Worker and AI Collaboration for Natural Language Inference Dataset Creation

A recurring challenge of crowdsourcing NLP datasets at scale is that human writers often rely on repetitive patterns when crafting examples, leading to a lack of linguistic diversity. We introduce a novel approach for dataset creation based on worker and AI collaboration, which brings together the generative strength of language models and the evaluative strength of humans. Starting with an existing dataset, MultiNLI for natural language inference (NLI), our approach uses dataset cartography to automatically identify examples that demonstrate challenging reasoning patterns, and instructs GPT-3 to compose new examples with similar patterns. Machine generated examples are then automatically filtered, and finally revised and labeled by human crowdworkers. The resulting dataset, WANLI, consists of 107,885 NLI examples and presents unique empirical strengths over existing NLI datasets. Remarkably, training a model on WANLI improves performance on eight out-of-domain test sets we consider, including by 11% on HANS and 9% on Adversarial NLI, compared to training on the 4× larger MultiNLI. Moreover, it continues to be more effective than MultiNLI augmented with other NLI datasets. Our results demonstrate the promise of leveraging natural language generation techniques and re-imagining the role of humans in the dataset creation process.

Label & Reasoning P: 5 percent probability that each part will be defect free. H: Each part has a 95 percent chance of having a defect. P: 1 percent of the seats were vacant. H: 99 percent of the seats were occupied.

Entailment Set complements P: The artisans, tradespeople, and providers of entertainment (reputable and not so reputable) lived downtown on the reclaimed marshlands north and east, in the area still known as Shitamachi. H: The only place where artisans, tradespeople and entertainers could live was in the marshlands to the north and east. P: At the time of the Revolution, the old port of Marseille was a great center of shipbuilding and commerce. H: The only place where ships were built was in the old port of Marseille.

Hypothesis introduces exclusivity P: To the south, in the Sea of Marmara, lie the woods and beaches of the Princes' Islands. H: In the north is the Sea of Marmara where there are mountains to climb. P: From the park's southern entrance, follow the avenue south to the Hotel de Ville. H: From the park's northern entrance, follow the avenue north to the Hotel de Ville.

Reversing cardinal directions P: Democrats released documents indicating that Republicans sold big political donors meals with the party's leaders in federal buildings in 1995. H: It is illegal for a party to solicit products to donors. P: In the late 1960s, students at a university in Wisconsin tried to organize a union. H: It was illegal for the students to organize a union.

Illegal things can happen P: She ducked and parried the blow. H: She ducked to miss the blow. P: She stepped on the brake and the car came to a stop. H: She stepped on the brake to stop the car.

Implied intention P: To build a worldclass finance organization and help achieve better business outcomes, each of the organizations we examined set an agenda for transforming the finance organization by defining a shared vision -i.e. H: The transformation was a disaster and the entire organization had to be scrapped. P: In order to help improve customer service, I suggested that they send a representative to our office to discuss our concerns. H: The representative sent to our office did not solve our problems and we lost a lot of business.

Intended goals may not actualize P: Salinger wrote similar letters to other young female writers. H: Other young female writers received similar letters from Salinger as well. P: The three schools have a number of students who are from families with no history of financial difficulties. H: Families with no history of financial difficulties send their children to the three schools.