Restart of English language-learning experiments

Restarted July 2021 ... ongoing, through Oct 2022

Directories:

• run-1 -- attempt to have a perfect run-through of everything. (July 2021)
• run-2 -- work with a flawed copy of tranche-1 only
• run-3 -- Similarity Smackdown - compare MI to overlap, jaccard, etc. (Aug 2021)
• run-4 -- Deep trimming of datasets (Sept 2021)
• run-5 -- Trimming of word-pair files.
• run-6 -- All similarities between top-ranked words.
• run-7 -- Exploring one merge at a time.
• run-8 -- Exploring one merge at a time, w/shapes, and more merges.
• run-9 -- Bringup of production ranked merge.
• run-10 -- Four different merge experiments, crashed around 15 classes
• run-11 -- Precise & imprecise merge, crashed, flawed
• run-12 -- Attempted compute of similarities. Broken.
• run-13 -- Staging area; no experiments.
• run-14 -- Clustering attempts; get about 100-200 merges deep, then blah. (later; revive for some frame debugging work.)
• run-15 -- Deep compute of MI similarities and GOE similarities. (Sept 2022)
• run-16 -- Cluster attempts with GOE similaries.
• run-17 -- Link-Grammar API development & test. (Oct 2022)
• run-18 -- Integrated pair+MST+clustering bring-up. (Nov 2022)

Spindled databases

Archived databases are in

/home2/linas/src/novamente/data/rocks-archive

Databases in the ~/data directory

The assorted run-1-*.rdb databases are "master copies" of the best runs with the properly, correctly applied processing. These took a long time to generate, and need to be archived. They are imperfect: right from the get-go, there's some bug with escaping quotes that needlessly pollutes these files. However ... that bug has not been found yet, has not been fixed yet, and we haven't re-run anything, so the below will do.

• run-1-en_pairs-tranche-1.rdb -- run-1 guten-tranche-1 only.
• run-1-en_pairs-tranche-12.rdb -- run-1 guten-tranche-1 and 2.
• run-1-en_pairs-tranche-12*.rdb -- etc.

The above are large.

• run-1-en_mpg-tranche-1234.rdb -- ?? I guess mpg-parsed ?? Huge. See Diary Part Two

• run-1-marg-tranche-123.rdb -- Described in Diary Part Three page 6,8 ... contains word pairs and also disjuncts, and MMT marginals for disjuncts (and maybe word-pair marginals?)

  This takes 50GB to load word pairs, and 60 GB to do anything with them.

• run-1-t1*-trim-1-1-1.rdb -- MPG-parsed and trimmed to remove words, disjuncts, and sections with a count of 1. Includes MM^T marginals (but for word-disjunct pairs only) and redone pair marginals. This amount of trimming was not enough! See below.

• run-1-t1*-tsup-1-1-1.rdb -- As above, but also removed all words, disjuncts with a support of only 1. See see run-5, run-6 and diary part three. This is the correct amount to trim! Includes MM^T marginals. ... but the MM^T marginals are only on the (w,d) pairs, and NOT on the shapes. The merge work needs shapes!

• run-1-t12-tsup-1-1-1.rdb has ... 7.5K x 7.8K words, total of 7.4M word-pairs; takes 6.0 GB to load word-pairs. has 7.1K x 66K (w,d) matrix, 270K word-disjunct pairs. Needs only additional 0.4 GB to also load (w,d) for total of 6.4 GB

  XXX -- there is an issue: the matrix-summary says 7.4M word-word pairs, but there are only 3.4M atoms. So the summary is pre-trim.

• run-1-t123-tsup-1-1-1.rdb has ... 44K x 44K words, total of 18.5M word-pairs; takes 13.6 GB to load word-pairs. 20 minutes to load. has 11.3K x 136K (w,d) matrix, 560K word-disjunct pairs. Needs only additional 0.9 GB to also load (w,d) for total of 14.5 GB

• run-1-t1*-shape.rdb -- Copy of above, with MM^T marginals on shapes. This is on the fat side, as it still retains the original word-pairs. It also contains the (un-needed) support and MM^T on the shapeless (w,d) pairs.

• r9-sim-200.rdb -- See Diary

• r9-sim-200+entropy.rdb

• r9-sim-200+mi.rdb

• r14-sim200.rdb -- qualiity connector set with top 200 words with MI in them.

Junk Databases

The following were generated in various experiments, but do not need to be archived, they can be deleted.

• r2-en_pairs.rdb -- just pair counts for guten-tranche-1 only. Above is missing 400+ files. There were some crashes. Part of run-2 -- includes MI

• r2-mpg_parse.rdb -- MPG disjunct counts for above aka run-2. This includes the MM^T marginals.

• r3-*rdb -- Assorted Similarilty Smackdown databases.

• r6-similarity-tsup.rdb -- copy of run-1-t1234-tsup-1-1-1.rdb with MI for word-pairs between the top 1200 wordvecs computed.

• r7-merge.rdb -- individual merge experiments.

• r8-merge.rdb -- individual merge experiments.

• r9-merge.rdb -- Bringup of production merge.

...THE END...