Experimental progress over time.
| Date | Changes | Accuracy | Acc. reach. | Compare | Commit |
|---|---|---|---|---|---|
| 2018-06-01 10:25:44 | automatic result recording | 0.6645 | 0.9074 | diff | original |
| 2018-06-04 07:51:20 | no oracle in training | 0.7027 | 0.9140 | diff | original |
| 2018-06-04 08:18:19 | predict from 1000 candidates | 0.3037 | 0.3361 | diff | original |
| 2018-06-04 08:31:27 | train and predict with 1000 candidates | 0.1309 | 0.1449 | diff | original |
| 2018-06-05 13:57:01 | represent mentions with subword units | 0.6645 | 0.8645 | diff | original |
| 2018-06-07 14:45:37 | candidate generation: similarity of phrase vectors | 0.5044 | 0.7548 | diff | original |
| 2018-06-07 14:52:36 | subword-unit embeddings for candidate generation and CNN | 0.3062 | 0.6770 | diff | original |
| 2018-06-07 15:05:48 | both skip-gram overlap and phrase-embedding for candidate gen. | 0.5972 | 0.7165 | diff | original |
| 2018-06-08 10:31:56 | candidate generators provide a score for each candidate | 0.6874 | 0.8272 | diff | original |
| 2018-06-08 10:38:22 | switch back to using only skip-gram candidates | 0.6811 | 0.886 | diff | original |
| 2018-06-11 11:01:30 | new candidate generator based on actual cosine similarity | 0.7433 | 0.8616 | diff | original |
| 2018-06-11 11:20:25 | use candidates from cosine similarity only | 0.77 | 0.8951 | diff | original |
| 2018-06-13 09:28:32 | fix evaluation mistake: "|"-separated reference IDs are composite, not alternatives | 0.756 | 0.9043 | diff | original |
| 2018-06-20 10:21:15 | use both word and subword-units | 0.7395 | 0.8845 | diff | original |
| 2018-06-20 10:26:55 | use subword-unit embeddings only | 0.756 | 0.9043 | diff | original |
| 2018-06-20 10:31:31 | trainable subword-unit embeddings | 0.756 | 0.9043 | diff | original |
| 2018-07-28 20:24:58 | more comprehensive word embeddings | 0.7713 | 0.9225 | diff | original |
| 2018-07-31 07:38:04 | use stemmed embeddings for cand. gen. | 0.7827 | 0.9419 | diff | original |
| 2018-07-31 08:07:39 | cand. gen. with stemmed emb. + s-gram cos. | 0.7789 | 0.8975 | diff | original |
| 2018-07-31 10:39:34 | adam optimizer | 0.7891 | 0.8974 | diff | original |
| 2018-07-31 12:08:43 | add an input node for token overlap | 0.8005 | 0.9104 | diff | original |
| 2018-08-06 08:59:47 | hyperonym cand. gen. | 0.8145 | 0.9105 | diff | original |
| 2018-08-06 15:51:28 | abbreviation cand. gen. | 0.8348 | 0.9075 | diff | original |
| 2018-08-07 09:48:01 | omit rank_score, use bare cand-gen scores | 0.8145 | 0.8854 | diff | original |
| 2018-08-09 08:53:52 | allow multiple filter widths in the convolution | 0.8208 | 0.8923 | diff | original |
| 2019-01-27 11:01:45 | classification with vocabulary pretraining | 0.859 | -- | diff | original |