| … | … | @@ -0,0 +1,31 @@ |
| 1 | # Parse with the learned dictionary — Test DICT. Pre-registration |
| 2 | |
| 3 | 2026-10-18. Written and committed **before any code or run**. Follows #17 and #18. #17: classes lift the pair-MI MST |
| 4 | parse to 0.715 link F1, but even gold classes cap it at 0.735: the parser scores pairs, not words' whole |
| 5 | connector sets. Linas's next step parses with the learned **dictionary**: each word's class has disjuncts (ordered |
| 6 | class connectors) with costs, and a parse is a choice of one disjunct per word whose connectors mate. |
| 7 | |
| 8 | ## Method (window counts W = 6, PREREG-U's MST: ungated ordered-pair MI, undirected, spanning) |
| 9 | |
| 10 | Sources of layer-1 classes: **G** (gold) and **RP1** (exchange over the word-MST links, round 1 of PREREG-reparse). |
| 11 | |
| 12 | 1. **Dictionary 0** from the training MST over class-pair MI: each token's disjunct = its links in position order as |
| 13 | connectors (side, class of the linked word), as Linas's Sections after clustering. Counts n(z, d) per class z. |
| 14 | P(d | z) = (n(z, d) + β Π_k Φ_side(y_k | z)) / (n(z) + β), β = 10, Φ from connector counts (add-0.1). |
| 15 | 2. **Parse with the dictionary.** Per sentence, candidates = the MST and 30 MSTs with Gumbel(0, 1) noise (bits) on |
| 16 | every window-edge score (root edges unchanged). Choose the candidate maximising Σ_t log P(d_t | z_t) (first on |
| 17 | ties). Search is approximate (candidates from MI), scoring is the dictionary's. |
| 18 | 3. **Re-estimate** (Viterbi): dictionary i+1 from the training sentences' chosen parses; iterations i = 1, 2, 3. |
| 19 | |
| 20 | **Measures** (held-out, 10 seeds, tiers 3K/10K/30K): undirected link F1 of the chosen parse per iteration; the |
| 21 | **oracle** F1 (best candidate by gold, per sentence) as the search ceiling; references G-MST and RP1-MST (the |
| 22 | MST over the same classes, from this run). |
| 23 | |
| 24 | ## Predictions (at 30K; paired over 10 seeds; "beats" = 95% CI of the paired difference excludes 0) |
| 25 | |
| 26 | - **D1 (the dictionary lifts the pair-MI ceiling):** with G, DICT (iteration 1) beats G-MST. |
| 27 | - **D2 (with learned classes too):** with RP1, DICT (iteration 1) beats RP1-MST. |
| 28 | - **D3 (re-estimation helps):** with RP1, iteration 3 beats iteration 1. |
| 29 | - **D4 (strong):** with RP1, DICT (iteration 3) beats G-MST. |
| 30 | |
| 31 | **Control R0:** without `--dict`, output is unchanged (Test A lines reproduce `results/1792340082`). |