Natural Language Inference

Understanding that when the sun is shining the weather is nice, and definitely not horrible, is a simple task for humans. To make this kind of reasoning, computers need to specifically model the relationship between sentences. This is what Natural Language Inference (NLI) systems do. They model relationships between real-world states or events, how dependent or irrelevant they are, and whether they contradict each other. This is what makes them an important part of AI solutions aimed for natural language understanding.

See Our Demo


With this demo you can predict the relation between two sentences using the neural network model introduced in Talman et al. 2019. The model has been trained on the training sets from both the Stanford Natural Language Inference (SNLI) corpus and the Multi-Genre Natural Language Inference (MultiNLI) corpus. The model achieves the prediction accuracy of 86.1% in the SNLI benchmark. Read our research on the known limitations of the state-of-the-art NLI models here: Talman & Chatzikyriakidis 2019.

Click below to evaluate random sentence pairs from the MultiNLI and SNLI test sets or type your own sentences in the text boxes. You can try for example some obvious entailments like "The Tesla stopped at the traffic lights while the driver was sleeping." and "The driver was not driving the car." or contradictions like "The sun is shining." and "The weather is horrible.".