| … | … | @@ -17,7 +17,7 @@ Merges: 100 of 100 loops at 10K and 30K (the loop limit, not the MI cutoff, ende |
| 17 | 17 | |
| 18 | 18 | ## What it says |
| 19 | 19 | |
| 20 | 1. **Under Linas's own call the clusterer covers too little to be a layer.** | |
| 20 | 1. **Under Linas's own call `(in-group-mi-cluster sha 200 100)` the clusterer covers too little to be a layer.** | |
| 21 | 21 | About half the tokens (LV-I) or a quarter (LV-A) get a class; the rest stay as single words, so there are |
| 22 | 22 | ~1,200 labels and the class-pair MI is nearly as sparse as word MI. The re-parse is slightly *worse* than words. |
| 23 | 23 | 2. **No per-occurrence sense effect** (L1): the partial-merge mechanism does not separate senses at this scale. |