freemanjiang

One thing is that this is trained on an English, character-level representation of kana characters, so it's possible it generates names that are not legal in the Japanese syllabary

We have some heuristics for filtering articles based on the coherence of their content (sentence length, line length).

The idea is also that high quality sources will link to to other high quality sources. Based on the filtered articles we've indexed, we use a TrustRank algorithm to order them.