@mhoye @XaiaX interesting thing is that it is probably *much* easier to simultaneously get the information for text completion *and* also what authors were involved in that match (instead of a large pool of authors just the relevant subset). [in the case of these compression-decompression + reference dataset models]