| Title | Learning to generate one-sentence biographies from Wikidata |
| Authors | Andrew Chisholm, Will Radford, Ben Hachey |
| Year | 2017 |
| URL | https://aclweb.org/anthology/E17-1060 |
Chisholm et al. 2017 present a system that generates a one-sentence biography on the basis of Wikidata facts. The backbone is a sequence-to-sequence model that takes as input a flat string representation of structured data from Wikidata, such as "TITLE mathias tuomi SEX OR GENDER male DATE OF BIRTH 1985-09-03", where the slots are ordered by frequency. Its encoder and decoder are both 3-layer GRUs, with tied embeddings, and an attention mechanism to focus generation on specific facts.
Interestingly, Chisholm et al. extend this encoder-decoder architecture to an autoencoder, where a second encoder-decoder component tries to regenerate the Wikidata facts on the basis of the generated sentence. In this way, the system learns to generate only facts it finds in the Wikipedia input. Experiments show this autoencoder setup drastically improves the output of the system beyond the capabilities of a single encoder-decoder architecture. Its BLEU score with Wikipedia biography sentences is .41, and its output is preferred 40% of the time to the Wikipedia alternative by human judges.
The code and data from this paper can be found on Github.