| … | … | @@ -0,0 +1,25 @@ |
| 1 | # Prior work: a discovered level constraining the next search (2026-10-18) |
| 2 | |
| 3 | Research-agent survey; two load-bearing claims re-checked against the sources (marked ✓). |
| 4 | |
| 5 | | line of work | does a discovered level constrain the next? | strongest reported result | source | |
| 6 | |---|---|---|---| |
| 7 | | ADIOS (Solan et al., PNAS 2005) | yes: patterns → context-local classes → patterns over classes | CFG: 100% precision / 99% recall with 28 patterns; ATIS perplexity 11.5 vs ~14 (3-gram) | tau.ac.il/~horn/publications/pnas.pdf | |
| 8 | | Ponvert, Baldridge & Erk 2011 (cascaded chunkers) | yes: chunk → collapse to pseudo-word → chunk again (5–7 levels) | PP recall 20.6 (level 1) → 70.2 (level 2); full-parse F 54.2 vs 51.7 (CCL, WSJ) | aclanthology.org/P11-1108 | |
| 9 | | Spitkovsky et al. 2011 | classes condition the parse | ✓ fully lexicalized DMV "drop[s] in half"; "some degree of categorization is essential to its learnability"; context-dependent induced tags 59.1% > gold tags | aclanthology.org/D11-1118 | |
| 10 | | Christodoulopoulos, Goldwater & Steedman 2012 | loop: classes → parse → classes | undirected-dependency features: +8.5% POS, +7% DMV over 5 iterations; directed: no significant gain | aclanthology.org/W12-1913 | |
| 11 | | Bisk & Hockenmaier 2012/2013 | yes: round t lexicon restricts round t+1 categories | 68.4 WSJ10 (needs gold POS + seeds) | aclanthology.org/Q13-1007 | |
| 12 | | Spitkovsky et al. 2010 (Baby Steps) | curriculum passing a learned model forward | 45.0% WSJ23 (Leapfrog) | aclanthology.org/N10-1116 | |
| 13 | | Headden et al. 2009; Koo et al. 2008 | class backoff / cluster-prefix features (multi-scale) | +3.8 (WSJ10); ~+1 supervised parsing | N09-1012; P08-1068 | |
| 14 | | DreamCoder (Ellis et al. 2021) | yes: learned library shrinks later search | sort: 5 calls with library vs 32 without; libraries up to 4 deep, depth vs solves r = 0.79 | arxiv.org/abs/2006.08381 | |
| 15 | | HVM (Wu et al., ICLR 2025) | yes: chunks + variables built from earlier ones | better likelihood than LZ78/HCM on BabyLM; depth not reported | arxiv.org/abs/2410.21332 | |
| 16 | | OpenCog / SingularityNET ULL (Linas's grammar report, 2019) | no recursion past word classes | ✓ link F1: LG-parse input 0.809; MST-MI 0.457; plain adjacent-word ("sequential") 0.549 | learn/learn-lang-diary/grammar-report/grammar-report.pdf | |
| 17 | | BabyLM curricula (findings, arXiv 2504.08165) | curriculum | "largely unsuccessful, though some showed modest improvements" | arxiv.org/abs/2504.08165 | |
| 18 | |
| 19 | **Takeaways for our work:** (1) pooling by class is the right core (Spitkovsky 2011); (2) every successful system |
| 20 | biases or adds levels, never hard-filters — keep the finer level reachable, classes context-dependent; (3) order |
| 21 | alone doesn't help (BabyLM), passing a learned model forward does (Baby Steps); (4) multi-scale backoff gains are |
| 22 | small in the literature too (+1 to +4), matching our fractal runs; (5) the recursion that works is Ponvert's: |
| 23 | collapse each found unit to one token and re-run the same learner; (6) in real-text language, level-wise systems |
| 24 | plateau around 50–60% — no one shows gains still compounding past 2–3 levels; (7) nobody has done it with |
| 25 | Linas-style connector sections, and ULL found plain adjacent-word parses beat MST-MI parses. |