| … | … | @@ -0,0 +1,36 @@ |
| 1 | # Linas's clusterer as layer 1 — Test LV. Pre-registration |
| 2 | |
| 3 | 2026-10-18. Written and committed **before any run beyond a timing check** (seed 1, 3K, Linas's setting: 13 s, |
| 4 | 95 merges, stopped by the MI cutoff; no classes were inspected). Follows #15 (recommendation 3) and #17. Linas |
| 5 | clusters word-disjunct vectors, and a merge moves only a word's shared disjuncts into the class, so senses are |
| 6 | factored inside the clustering (`gram-class/README.md:190-194`, `agglo-mi-rank.scm:461-466`). Our port |
| 7 | (`ladder/port/cluster.c`) matches his round 1 exactly; later rounds are unverified (round-2 divergence, parked). |
| 8 | |
| 9 | ## Pipeline (per seed and tier; 10 seeds, tiers 3K/10K/30K) |
| 10 | |
| 11 | 1. **Sections:** the word-level MST of PREREG-U (ungated MI, undirected, spanning). Each training token gives |
| 12 | Section(germ word, ConnectorSeq of its linked words, `-`/`+` by position, sorted by position); counts are |
| 13 | summed over tokens. Held-out tokens get Sections the same way (held-out parse with training MI). |
| 14 | 2. **Clustering:** `cluster <sections> --nrank 200 --loops 100` (Linas's `(in-group-mi-cluster sha 200 100)`, |
| 15 | `agglo-mi-rank.scm:697`), in two arms: **LV-I** `--intended` (Sections + Shapes in the similarity) and **LV-A** |
| 16 | as-run (Shapes only, as Linas's stack runs today, STAGE_E_SPEC §5). A final-state dump lists every Section pair |
| 17 | (row, ConnectorSeq, count) and the word → class memberships. |
| 18 | 3. **Token labels:** a token with Section (w, d) gets the class that holds column d among w's classes (largest |
| 19 | count if several); if none does, w's class with the largest membership count; else w itself (a singleton label). |
| 20 | |
| 21 | ## Measures |
| 22 | |
| 23 | Token labels on held-out text: V-measure, many-to-one, number of labels, share of tokens with a class label; sense |
| 24 | accuracy on polysemous words (gold: a word seen with ≥ 2 gold classes in training) and the same for **LV-COLL** |
| 25 | (each word's most frequent LV label, as PREREG-v2). Re-parse: link F1 of the PREREG-U MST over LV labels; Test A-U |
| 26 | ucomp type recall and precision with that parse. References from the same run: word MST (round 0), L-MST |
| 27 | (exchange), G-MST. |
| 28 | |
| 29 | ## Predictions (at 30K; paired over 10 seeds; "beats" = 95% CI of the paired difference excludes 0) |
| 30 | |
| 31 | - **L1 (senses come from the clustering):** LV-I beats LV-I-COLL on sense accuracy. |
| 32 | - **L2 (classes constrain the parse):** the MST over LV-I labels beats the word MST on link F1. |
| 33 | - **L3 (as good a layer as exchange):** link F1 of LV-I − L-MST has 95% CI lower bound > −0.02. |
| 34 | - **L4 (and at level 2):** LV-I's ucomp type recall beats the word MST's. |
| 35 | |
| 36 | Reported, no prediction: LV-A against LV-I (Linas's stack as run vs as designed); number of classes formed. |