This visualization aims to help us develop intuition of (1) how many queries have opposing views, (2) if we detect the opposing views, and (3) what the claims for a given query. In short, are we able to align the claims and are we able to appropriately detect entailment and contradiction between the claims.
The color scales with value. Blue indicates entailment and red indicates contradiction. The highlighted cells indicate that we detect there is a negation (not an NLI model) by using some simple rules. We expect this to only be useful if the distance between the negated pairs is also high.
Hover to see the specific queries & details. Select queries from the dropdown. Taxonomy means the groups are sampled from the query results of a specific claim or its negation. Similarity is based on nearest neighbors using cosine distance with summed ELMo embeddings. Sparse Clustering uses KMEANS with tf-idf. Dependency parsing is groups based off the nsubj of sentences -- currently sentences with the subject "you" are being ignored. View all the claims below the grid.