Cryptic dataset

The Cryptic dataset is constructed using from A Dataset of Cryptic Crossword Clues by George Ho (available at https://cryptics.georgeho.org) This database consists of cryptic clues from multiple different sources, incuding the New York Times, The Guardian, The Times and The Hindu. Each row in the database stores the clue (normally a cryptic sentence or two along with the length of the answer), the answer (normally a word or two), the definition (one or a few words, almost always either the start of the clue or the end). We have extracted each clue that consists of 6 words from the database, and encoded the words with the BERT encoder to 6 word embeddings of dimension 768. The task is to determine whether the cryptic clue contains the definition at the beginning or end of the clue. This makes a binary classification problem. The labels are as follows {beginning: 1, end: 0}. The shape of the input data is (n, 1, 6, 768), and the labels are binary (0 or 1). The extracted data is split into three sets, the test set consists of clues from The Times, the validation set from The Hindu, and the train set from all other sources.