Mova — a small, transparent SLM in Rust that can say why

Papers Mova applies

105 works from Mova's own list, by the stage of Mova's pipeline that uses them. measured marks a work whose effect Mova measured in its own experiments — 30 of them, including ideas that did not help; the others are rule sources, designs or background.

Ideas in progress

Every card ends with ideas for Mova. Mova keeps them in a ledger — 5413 ideas over its shelves — and marks each one it tries. These are the 60 already taken up; each links to the card it came from.

Kept 36 · tried, measured and kept in Mova

• keptTurn the morpheme head into a deterministic check: a segmenter's output for a token must concatenate back to the surface string. The paper's character-with-boundary view makes this a free invariant to test on every token.

  Result: consonant check on every candidate (Model::spells): 0 failures on dev, test and the two-fold cv of train; kept as a gate over cut(); commit 5c24bc409

  from AlephBERT: Language Model Pre-training and Evaluation from Sub-Word to Sentence Level I2 · shelf he · 2026-10-18
• keptTest the "of" effect: report PP attachment accuracy for of phrases and for the rest in en parse-eval. The paper says of is almost always a noun attachment, so a rule for of is the baseline the rest must beat. Measure on EWT test.

  Result: of-PPs: parser 87.42 (missed = error) / 97.06 (discarded) vs the rule of -> noun 95.36; commit 35e51ff1b

  from How Bad is the Problem of PP-Attachment? A Comparison of English, German and Swedish I2 · shelf en · 2026-10-18
• keptFloor before model: compute the majority-attachment rate per treebank and report it with every PP score, so that a parser's accuracy is read against it, not against 50%.

  Result: always-noun floor reported with every PP score: 48.67 on EWT dev (of 95.36, other prepositions 37.93); commit 35e51ff1b

  from How Bad is the Problem of PP-Attachment? A Comparison of English, German and Swedish I3 · shelf en · 2026-10-18
• keptMova construction corpus: about 700 sentences of our own across seven constructions (object and subject relatives, reduced relatives, wh-questions, tough adjectives, free relatives), held out from training; report recall per construction for en and gparse on each release candidate (dev).

  Result: en/data/constructions.tsv: 110 own sentences, 11 constructions, 220 arcs; EWT+ESLSpok model 61.8 recall, en.bin 0.4 70.0, after the parser changes 65.5; commit 25d072835

  from Unbounded Dependency Recovery for Parser Evaluation I1 · shelf en · 2026-10-18
• kepten parse-eval as a construction gate: a release fails if any construction falls below its last-release value even when LAS rises; check that the gate goes red on a deliberately broken tagger (the red control).

  Result: en construct-eval gate with per-model floors; red control (tags shifted one word) turns all 11 constructions red (61.8 -> 3.6); commit 25d072835

  from Unbounded Dependency Recovery for Parser Evaluation I2 · shelf en · 2026-10-18
• kepten/src/parse.rs constraint mode: force the gold head and label for one error class at a time during beam decoding, and report direct and cascaded gains per class, as the paper does; measure on EWT dev (test only for the final numbers).

  Result: en parse-diag constrained beam: forcing punct / core args / function words / root fixes 11.7 / 10.6 / 10.3 / 10.3% of the other errors; coordination 3.4%, PP 5.5%; commit 3ba5388e8

  from Identifying Cascading Errors using Constraints in Dependency Parsing I1 · shelf en · 2026-10-18
• kepten beam-curve coverage: report, for each beam width, the share of sentences where the constrained decoder finds no tree; measure it on EWT dev with the class constraints from I1.

  Result: no-tree count per class at beam 8 on EWT dev: clauses 41, PP 23, coordination 14, punct 12, others <= 6 sentences of 2001; commit 3ba5388e8

  from Identifying Cascading Errors using Constraints in Dependency Parsing I4 · shelf en · 2026-10-18
• kepten/src/parse.rs feature template: add conjunct-symmetry features (lemma match, suffix match, capitalization match, UPOS match) for candidate conj attachments in the arc-standard perceptron; metric: conj Rel+Att F1 and LAS on EWT test.

  Result: conjunct-symmetry features, EWT dev our tags, 3 shuffles: 79.20 -> 79.37 (+0.17), gold tags 83.92 -> 84.38; commit 68f0f000b

  from Improving a Strong Neural Parser with Conjunction-Specific Features I1 · shelf en · 2026-10-18
• keptConstrain by a wordlist: in unpointed mode, restrict each token's candidates to forms in Mova's lexicon, then score them. The paper's largest single gain (6 to 11 points) comes from this lookup step, not from a bigger model. It fits the unpointed mode of tools/hbo, with the lex table as the list.

  Result: host-lexicon check on unseen words (hostlex:b): hard filter hurts (names cut into known hosts: pointed dev 97.38 -> 95.77); soft b=0.3 kept for pointed (with partial: dev 97.38 -> 97.58, cv 96.74 -> 97.08, e2e LAS dev 82.14 -> 82.51, test DEU 87.02 -> 87.53, GEN 87.07 -> 87.35); unpointed b=0.7 dev…

  from NLP in the DH pipeline: Transfer-learning to a Chronolect I1 · shelf he · 2026-10-18
• keptConsonant check: the first error category (the predicted consonantal form differs from the surface text) needs no linguistics. Make it a validator on every tools/hbo output, with the input text as the gold consonants; it costs nothing and catches a whole error class.

  Result: consonant check in tools/hbo (Model::spells, no-check as red control): 0 failures on dev, test and cv (278 799 tokens); commit 5c24bc409

  from A Transformer-based parser for Syriac morphology I1 · shelf arc · 2026-10-18
• keptCritical-span ablation as a validity check of course questions: remove the critical sentence; the student should fall to guessing on that question. Questions still answered right are guessable from elsewhere or from the options and should be fixed or marked.

  Result: critical-span ablation per lesson as the «key sentences removed» control (dba7f8c6e)

  from STARC: Structured Annotations for Reading Comprehension I3 · shelf en · 2026-10-18
• kepten/src/parse.rs: add a distance feature for each candidate arc (direction, binned distance, and a count of verbs or punctuation between the two words); measure cv LAS on the train folds, then EWT test once per release.

  Result: verbs/punctuation between the pair, on top of conj+subcat: 79.80 -> 79.90 (+0.10, noise level), gold tags 84.41 -> 84.77 (3 shuffles), construction set 62.7 -> 65.5; all three groups +0.70 dev LAS; commit d80e79ef9

  from Head-Driven Statistical Models for Natural Language Parsing I1 · shelf en · 2026-10-18
• kepten/src/parse.rs: add a subcategorization feature for a verb head, namely which of its obligatory dependents (nsubj, obj, iobj, ccomp, xcomp) are already attached; measure cv LAS and the LAS of verb-headed arcs with en ud-eval.

  Result: subcat bit set of attached core dependents, on top of conj: 79.37 -> 79.80 (+0.43, 3 shuffles), gold tags 84.55; commit 1c729b451

  from Head-Driven Statistical Models for Natural Language Parsing I2 · shelf en · 2026-10-18
• kepten parse-eval: report PP attachment accuracy taken from the predicted tree, not from gold (v, n1, p, n2) tuples, next to LAS on UD EWT test. Count the PP cases the parser misses as a separate number.

  Result: PP attachment from the predicted tree in en parse-diag: 1615 dev PPs, 190 missed; commit 35e51ff1b

  from Prepositional Phrase Attachment without Oracles I1 · shelf en · 2026-10-18
• keptReport three figures for every PP score, as the paper does: missed cases counted as errors, missed cases given the noun attachment, and missed cases discarded. A score that rises only when cases are discarded must not pass the gate.

  Result: three figures, EWT dev: missed = error 72.20, missed -> noun 77.65, discarded 81.82; commit 35e51ff1b

  from Prepositional Phrase Attachment without Oracles I2 · shelf en · 2026-10-18
• keptPartial-input controls for every course, after the paper's ablation: run the student with the lesson removed (question + options only) and with the question removed (lesson + options only). Questions still answered right without the lesson measure prior or option bias, not reading; report them per course and subtract…

  Result: permanent course controls: no lesson, questions shuffled, key lines removed; all red in all seven courses (biology real 2238/4581, no lesson 1208, shuffled 1517, key removed 1501) (dba7f8c6e)

  from Cosmos QA: Machine Reading Comprehension with Contextual Commonsense Reasoning I1 · shelf en · 2026-10-18
• keptReport Mova's bAbI results (world/src/babi.rs) and each course with the paper's three columns: full, question-only (lesson replaced), passage-only (questions shuffled within the lesson), and Δ(min) = full minus the better partial score. A task or course with small Δ(min) does not show reading. Same partial-input…

  Result: question-only and passage-only style controls in every course report (dba7f8c6e)

  from How Much Reading Does Reading Comprehension Require? A Critical Investigation of Popular Benchmarks I1 · shelf en · 2026-10-18
• keptA last-sentence ablation for the courses: keep only the lesson sentence with the highest overlap to the question (or only the last k sentences). If the student scores the same, the questions test sentence lookup, not integration across the lesson; use the result to choose which courses can show gains from coreference…

  Result: key-line ablation (line with the most content words shared with question + right answer) as the third control (dba7f8c6e)

  from How Much Reading Does Reading Comprehension Require? A Critical Investigation of Popular Benchmarks I2 · shelf en · 2026-10-18
• keptKeep the red controls in the course report permanently (not one-off), so every claimed gain is shown next to them, in line with Mova's rule that a gate which cannot show red is not a gate.

  Result: the controls are printed in every course report with a red / NOT RED verdict (dba7f8c6e)

  from How Much Reading Does Reading Comprehension Require? A Critical Investigation of Popular Benchmarks I3 · shelf en · 2026-10-18
• keptImplement Eckert and Strube's context test as the switch in front of the pronoun sieves for it, this, that: an I-incompatible predicate (be true/false/right/wrong/why/the reason/a pity/the problem, happen, mean, object of believe/know/say/forget) sends the pronoun to the abstract list; an A-incompatible one…

  Result: the context test (event subjects: happen, mean, be true/a mistake…) in coref/src/it.rs (2bb818bf8); «it» as object of know/say as an event dropped (−0.02 CoNLL)

  from Anaphora With Non-nominal Antecedents in Computational Linguistics: a Survey I1 · shelf en · 2026-10-18
• keptError map: the workers failed mostly when the PP attached to a verb or an adjective, and rarely to a noun. Report PP head accuracy in en parse-eval split by the gold head's part of speech, on EWT test.

  Result: PP accuracy by gold head, EWT dev: noun 70.23 / verb 74.07 (missed = error), 79.08 / 84.46 among found; commit 35e51ff1b

  from Towards Semi-Automated Annotation for Prepositional Phrase Attachment I2 · shelf en · 2026-10-18
• keptTreat missing points as unspecified, not as absent: change the distance so that a point the text does not show does not count as a disagreement. For partly pointed Hebrew (editions with some points, or rabbinic texts), draw the accuracy-against-point-density curve, as in the paper's Figure 1, and test which point…

  Result: missing points unspecified (partial): pointed lexicon forms whose points include the input's, else the unpointed segmenter for thinly pointed words; dev with 0/25/50/100% points dropped 97.48/81.82/69.28/48.84 -> 97.63/95.37/94.61/94.21; fully pointed dev +0.15, cv +0.08; commit 5dbc900f6

  from Exploiting Arabic Diacritization for High Quality Automatic Annotation I2 · shelf sem · 2026-10-18
• keptClassify every it three ways before the pronoun sieve runs, with rules from the paper's strongest cues, and let only the anaphoric ones reach nominal antecedents. This attacks Mova's worst pronoun (it: 41% precision on gold trees, 30% on SLM trees; the README's first error type). Pleonastic: weather and time…

  Result: coref: «it» pleonastic/event/anaphoric before the pronoun sieve; GUM dev CoNLL gold 75.18/64.30→75.21/64.32, SLM trees 69.70/57.57→69.82/57.66, «it» precision 51.9→53.5 / 42.5→45.3 (2bb818bf8). In the SLM frame reader: courses 7672/2199→7633/2202, dropped there

  from What is it? Disambiguating the different readings of the pronoun 'it' I1 · shelf en · 2026-10-18
• keptError profile by distance to the root in en parse-eval and in gparse (Eisner decoding, gparse/src/eisner.rs): precision and recall of arcs at distance 1, 2, 3–4 and 5 or more, on EWT test; the question is whether the arc-standard parser fragments the tree near the root as the paper suggests for greedy parsers.

  Result: en parse-diag, EWT dev our tags: labelled recall by depth 1/2/3-4/5+ = 87.16/77.33/79.62/78.76 (depth 2 weakest); by arc length 1..7+ 86.8 -> 56.8; commit 3ba5388e8

  from Characterizing the Errors of Data-Driven Dependency Parsing Models I1 · shelf en · 2026-10-18
• keptNon-projectivity split: report LAS separately for sentences with a gold arc of degree 1 or more; the arc-standard parser in en/src/parse.rs cannot build those arcs, so the number is its ceiling on EWT and GUM test (GUM is test only).

  Result: en parse-diag: 33 of 2001 dev sentences non-projective (39 arcs, 0.16%), LAS 72.57 vs 79.67; non-projectivity costs at most 0.16 LAS; commit 3ba5388e8

  from Characterizing the Errors of Data-Driven Dependency Parsing Models I4 · shelf en · 2026-10-18
• keptOracle gap from en beam-curve: report the oracle LAS of the beam against the 1-best LAS at beam sizes 1, 2, 4, 8 and 16 on EWT dev; a large gap, as the paper shows for 50-best, says a reranker has room to help.

  Result: beam oracle EWT dev our tags: width 1/2/4/8/16 1-best 75.27/78.12/78.83/79.39/79.24, oracle 75.27/79.67/82.20/84.14/85.52 (gap 4.74 at 8); commit 3ba5388e8

  from Coarse-to-Fine n-Best Parsing and MaxEnt Discriminative Reranking I2 · shelf en · 2026-10-18
• keptKeep events in non-real modes out of «what happened» answers: negated frames (neg), conditional ones (cond) and the unrealised complements proposed in white-2016-universal-decompositional-semantics-universal-dependencies I1 do not answer «What did the lion do after the meal?» but do answer «What would the lion…

  Result: an «if» clause is irrealis for the student (PRAG_OFF=cond-irrealis): 7702/2200→7704/2206 (35db4e3f7)

  from TORQUE: A Reading Comprehension Dataset of Temporal Ordering Questions I3 · shelf en · 2026-10-18
• keptUse the paper's verb–signature table as the seed of the implication signatures proposed for Mova's frame factuality (see rudinger-2018-neural-models-factuality I1): our own list built from these published class memberships, stored in prag/data with the signature and the complement type (to vs that). Extend…

  Result: the signature table seeded prag/data/signatures.tsv (six of the eight classes, «to» complements only) (35db4e3f7)

  from How well do NLI models capture verb veridicality? I1 · shelf en · 2026-10-18
• keptGive every frame a factuality value, not only neg and cond: happened / did not happen / not asserted (wanted, planned, hoped, asked, ordered) / uncertain (modal, hedge). Compute it by propagating the Nairn implication signatures down the tree in prag/src/frames.rs: manage to X +|−, fail to X −|+, forget to…

  Result: implication signatures (51 verbs, prag/data/signatures.tsv) set neg/irrealis on «to» complements top-down; courses 7672/2199→7704/2206 with I2 (35db4e3f7)

  from Neural Models of Factuality I1 · shelf en · 2026-10-18
• keptLet the student remember only factual frames as facts and keep the rest as wishes or plans (world/src/knowledge.rs): «The fox wanted to eat the grapes» must not answer «Did the fox eat the grapes?» with yes. The Torah diagnosis already shows negation losing to a positive frame; a factuality value is the cleaner fix…

  Result: the student treats an irrealis mismatch as NegDisagree; reader alone 7677/2201, with the student 7704/2206 (35db4e3f7)

  from Neural Models of Factuality I2 · shelf en · 2026-10-18
• keptSplit every wrong prediction into two kinds: "gold analysis missing from the candidate list" and "gold analysis present but not chosen". The paper's 330 of 850 counts are the template. Add this split to the tools/hbo evaluation so that segmenter coverage and ranking are reported apart.

  Result: hbo seg eval splits errors into gold missing from the candidates vs present but not chosen, plus oracle@1/2/3/5: unpointed dev 36 missing / 79 not chosen, oracle@2 97.68 vs 94.21 chosen; pointed dev 10 / 42; test DEU pointed 62 / 175, unpointed 165 / 494; commit 5882d8df1

  from A Pointer Network Architecture for Joint Morphological Segmentation and Tagging I1 · shelf he · 2026-10-18
• keptA small choice model over the segmenter's candidates: the tools/hbo tables already give the candidate splits. A tiny scorer (linear, or a rule with a few context features such as neighbour UPOS and prefix class) picks one split per token. Compare it with the current rules by cross-validation on train, not on test.

  Result: choice over the lexicon's candidates by the neighbouring words (ctx:0.5, the lexicon share smoothed toward the counts with the word before, then the word after): unpointed split right cv 93.31 -> 93.72, dev 94.21 -> 94.56; end-to-end LAS dev 76.56 -> 77.09, Chronicles 78.43 -> 79.73, test DEU 78.96…

  from A Pointer Network Architecture for Joint Morphological Segmentation and Tagging I2 · shelf he · 2026-10-18
• keptStore the candidates per token as an ordered list, so that a choice is a single index. Log the index with the features that decided it. This matches Mova's "say why" rule and fits the world store.

  Result: Model::candidates gives the ordered candidate list; the choice is logged in MISC (Seg= with the neighbour that decided, Lattice=<label>:<gain>); commits 5882d8df1, fc707558a

  from A Pointer Network Architecture for Joint Morphological Segmentation and Tagging I3 · shelf he · 2026-10-18
• keptA realis flag on every frame: did the event happen according to the text? Mova has neg and cond; add «not realised» for the complement of want/plan/try/hope/refuse/promise, for future and modal frames, and for imperatives, so the student does not answer «Did the fox get the cheese?» yes from «the fox wanted to…

  Result: realis flag for complements of want/plan/try… and «if» clauses kept (35db4e3f7); modals (will/would/might/may/shall/should) as irrealis dropped: −38 tests, −2 exams

  from Universal Decompositional Semantics on Universal Dependencies I1 · shelf en · 2026-10-18
• keptRelax number agreement in a controlled way in the pronoun sieve (coref/src/sieve.rs) and in the frame reader's resolver: they/them/their may link to a singular antecedent when that antecedent is a collective noun (government, family, team, crowd, army, enemy, company, people of a town), a quantity head…

  Result: coref: they → singular collective noun (24 nouns) only when no strict plural fits; GUM dev gold 75.21/64.32→75.25/64.36, «they» precision 64.9→66.3, SLM trees unchanged (a7fa2d765). In the SLM resolver: courses 7672/2199→7653/2189, dropped there

  from A Predictive Model for Notional Anaphora in English I1 · shelf en · 2026-10-18
• keptUse the paper's processing finding as an ordering rule: prefer a strictly agreeing candidate nearby, and accept a notional one only at a larger distance or when no agreeing one exists; check on GUM dev that the rule does not take links from correct plural antecedents.

  Result: the ordering rule: strict agreement first, notional only as a fallback (a7fa2d765; test with a strict-plural control)

  from A Predictive Model for Notional Anaphora in English I2 · shelf en · 2026-10-18

Tried 10 · tried; the measurement is in, the decision is pending

• triedNewman's "moderate antiquity" as a strength setting of a delta. Rank an archaic author's overrides by count and apply only the top k. A small k gives thou and hath; a large k gives the rare forms Arnold calls odd. Sweep k on the school registers and measure archaism (I1) against the public score (I2). This gives a…

  Result: translate r6: the archaism dial for KJV → WEB — WEB's k strongest overrides: k=0 250.0, 10 246.4, 50 244.1, 200 240.7, all 239.0 bits a verse on dev (monotone)

  from On Translating Homer; with F. W. Newman's "Homeric Translation in Theory and Practice" and Arnold's "Last Words" I3 · shelf mtrans · 2026-10-18
• triedAlign all English renderings of one verse across the translations in the same way. The shared backbone is the frame-preserving part of the verse, and the slots are the author delta. A slot filled in three or more translations goes to the standard; a slot filled in one goes to the author.

  Result: translate r6-lattice: the four Bibles' renderings of each training verse aligned (LCS of concepts); the standard learned on the shared backbone only (1.20 M of 1.66 M tokens), deltas on all: twelve-pair mean 180.8 → 182.3 bits a verse on dev; not kept (--lattice on)

  from Learning to Paraphrase: An Unsupervised Approach Using Multiple-Sequence Alignment I1 · shelf mtrans · 2026-10-18
• triedSplit each meta token into what is meant and the way of meaning. The concept (thou/you) is what is meant. The way of meaning becomes explicit features: number, person, politeness, definiteness, aspect, word class. Today meta.rs merges members and drops their differences. Benjamin's distinction says the merge is…

  Result: translate: concept + features measured twice on dev — pronoun number/person (r2-feats, 197.0→209.0 bits a verse, twelve-pair mean) and determiners/prepositions as features (r4-fw, 186.4→278.2): both worse, off (--feats, --fw)

  from Die Aufgabe des Übersetzers (The Task of the Translator) I1 · shelf mtrans · 2026-10-18
• triedStructural insertion templates by POS. Learn from verse pairs rules such as NOUN NOUN to NOUN of NOUN, keyed by enum tags, and use them in the encoder when the concept is unseen. This gives the encoder a fallback below the concept-level contexts. Measure the function-word accuracy on unseen concepts.

  Result: translate fwt.rs: per author pair insert/drop table of function words on the meta-world (r5-fwt); kjv→web 249.6→248.6 dev, shuffled-table control red; kept (--fwt on)

  from Syntactic Models for Structural Word Insertion and Deletion during Translation I2 · shelf mtrans · 2026-10-18
• triedDeletion only where it helps. Keep structural deletion off as a standard rule and allow it through the author DELTA, which the paper's direction finding suggests. Compare the standard with and without deletion per author pair, using the rule that a delta must fix at least three more tokens.

  Result: translate fwt.rs: deletion only where it holds ≥3 cases and half the context's (r5-fwt), kept

  from Syntactic Models for Structural Word Insertion and Deletion during Translation I3 · shelf mtrans · 2026-10-18
• triedJoin only where the parents agree. When tools/translate merges two meta tokens from two translations of a verse into one concept (meta.rs), allow the join only if the heads they hang from already joined, and go top-down, stopping at the first disagreement. This is the paper's rule against open problems 1 and 2.…

  Result: translate r6-lattice (with barzilay-2003 I1): backbone by LCS of concepts across the four renderings; worse on dev, not kept

  from Syntax-based Alignment of Multiple Translations: Extracting Paraphrases and Generating New Sentences I1 · shelf mtrans · 2026-10-18
• triedIllusory correspondences inside English. The interlingua lexicon anchors two Bibles on equal lemmas (the longest common subsequence in meta.rs). Equal lemmas across centuries are Postgate's illusory correspondences: KJV let (hinder), prevent (go before), conversation (conduct), quick (alive), meat (food)…

  Result: translate report «False friends»: KJV's let, prevent, conversation, quick, meat, suffer against WEB and JPS: anchored to themselves vs where they differ and what the other writes; meat is never anchored to itself (food, meal), suffer differs from WEB 27 times (allow) but anchors to JPS 26 times…

  from Translation and Translations: Theory and Practice I1 · shelf mtrans · 2026-10-18
• triedA lower-bound gate for lexicon merges (open problem 2): accept a merge, or a chunk choice in the encoder, only when the 95 per cent lower bound of the co-occurrence rate clears a threshold. This is a closed formula over counts that could decide cases like be with have, where today's merges are noise. Test it on…

  Result: translate r6-lb: the judge takes the one-sided 95% lower bound of the measured saving (--lb on); twelve-pair mean 181.0 → 180.8 on dev; kept

  from Statistical Glossing - Language Independent Analysis in Bible Translation I1 · shelf mtrans · 2026-10-18
• triedA count test for lexicon merges. Merge two concepts (or two meta tokens) only when their proportions over the same contexts agree under a dissimilarity like the paper's, computed from counts. This targets open problem 2. Measure the merges kept, including the be/have case, and the reader plus encoder bits on dev.

  Result: translate r6-lb: a count/lower-bound test for merges (with riding-2008 I1); kept

  from Learning synchronous context-free grammars with multiple specialised non-terminals for hierarchical phrase-based translation I1 · shelf mtrans · 2026-10-18
• triedAttach, do not align. In the standard encoder (profile.rs), write a target function word (case marker, article, auxiliary) with the chunk of the content word it binds to, not through the word alignment. Measure the function-word class of the residual in exact mode on dev, run twice.

  Result: translate r4-fw: function words bound to the head concept were much worse (186.4→278.2); the per-pair insert/drop table (r5-fwt) is kept instead

  from Effective Use of Function Words for Rule Generalization in Forest-Based Translation I1 · shelf mtrans · 2026-10-18

Measured, no gain 14 · tried and measured; it did not help, so it was not kept

• measured, no gainJoint segmentation and parsing as a Mova measure: run tools/hbo in a mode where all segmentations from its tables stay as arcs, and compare its one-best pipeline with the lattice reading on a dev split of treebanks/hbo/ptnk. Report both numbers and the share of tokens that change segmentation.

  Result: one-best vs lattice reading measured on dev and Chronicles (hbo lattice/choose): no gain; unpointed dev 77.09 best 77.39 (tuned) but Chronicles -0.10 at that setting; commit fc707558a

  from Joint Hebrew Segmentation and Parsing using a PCFGLA Lattice Parser I1 · shelf he · 2026-10-18
• measured, no gainStore affixes as atomic units plus a combination rule, not as a list of every compound. A Hebrew prefix chain (conjunction, preposition, article, then a stem) can be written as a few atomic prefixes, and the allowed chains follow from a short rule set, in place of an enumerated pre table in tools/hbo.

  Result: prefix chains as a table of atomic prefixes (chain:N in tools/hbo): no gain, dev unchanged, cv +-0.01; commit 73a57de20

  from Arabic Morphological Analyzer with Agglutinative Affix Morphemes and Fusional Concatenation Rules I1 · shelf sem · 2026-10-18
• measured, no gainKeep the lattice: let tools/hbo output every segmentation candidate of a token, with its tag and lemma per arc, and pass the whole lattice to gparse, so that the parser can choose the segmentation that fits the tree. The paper's own error analysis is the reason.

  Result: lattice in tools/hbo (hbo lattice / gparse GP_SCORE / hbo choose): the tree score does not pick the gold split (gold candidate +0.13 mean tree gain vs -0.41 for a wrong one replacing gold); unpointed dev LAS 77.09 -> 76.42..77.39, Chronicles (unseen) 79.73 -> 79.58..79.99, dev-best -0.10; pointed…

  from What's Wrong with Hebrew NLP? And How to Make it Right I1 · shelf hbo · 2026-10-18
• measured, no gainConstrain prefix segmentation by a table of legal proclitic sets keyed on the word's initial letters. The model applies only the valid function sets for a word's first letters at inference. A deterministic pre table in the segmenter can enforce the same rule; measure how many illegal splits it removes on a held-out…

  Result: legal prefix-chain table as a hard filter (chain:N, learned chains): split right dev 97.38 / 94.21 unchanged, cv +-0.01 (pointed, unpointed); the learned prefix table already proposes only seen chains; commit 73a57de20

  from DictaBERT: A State-of-the-Art BERT Suite for Modern Hebrew I1 · shelf he · 2026-10-18
• measured, no gainTurn the semantic predicates already stored in prag/data/verbnet.tsv (the frame rows carry has_possession !has_possession transfer cause …, but nothing reads them) into derived state facts for the student: after a transfer frame, the recipient has the theme and the source does not; after a motion frame the theme…

  Result: result states from VerbNet predicates (possession after transfer, location after motion) remembered as Fact frames: courses 7704/2206→7698/2203; possession only 7678/2207, location only 7688/2208; not committed

  from VerbNet Representations: Subevent Semantics for Transfer Verbs I1 · shelf en · 2026-10-18
• measured, no gainIrrealis for future-having classes (promise, offer, owe): their transfer did not happen yet, so no possession fact is derived. A red control for I1: «The king promised the miller gold» must not yield «the miller has gold».

  Result: no states from classes whose predicates say irrealis (promise, offer) — part of I1, dropped with it

  from VerbNet Representations: Subevent Semantics for Transfer Verbs I4 · shelf en · 2026-10-18
• measured, no gainBlock same-head links when the mention is marked as comparative or contrastive, a cheap attack on the second main error of coref («same text or head, different entities», exact 69.6%, strict_a 54.7%): a mention with other, another, further, different, second/next (or an ordinal not matching the candidate's) is…

  Result: block same-head links under other/another/second/different…: no change on gold trees, one link on SLM trees (+0.01) — strict_a's modifier subset and the indefinite rule already block them; not committed

  from Fine-grained Information Status Classification Using Discourse Context-Aware BERT I1 · shelf en · 2026-10-18
• measured, no gainKeep an abstract-referent list next to the entity list, with the survey's search order: the clause to the left in the same sentence, then the rightmost main clause of the previous sentence (and a subordinate clause to its right), then earlier main clauses. In prag/src/frames.rs this list is just the recent event…

  Result: search order for the clause antecedent used by the deixis sieve (left clause, then previous main clause): 12% precision on GUM dev, −0.8 CoNLL with singletons; sieve kept off (2bb818bf8)

  from Anaphora With Non-nominal Antecedents in Computational Linguistics: a Survey I2 · shelf en · 2026-10-18
• measured, no gainTable 7.1 as data: a template × configuration matrix (e.g. niXYaZ and hitXaYeZ never simple transitive; XuYaZ and huXYaZ passive only; XiYeZ active only) is a hard constraint for a Modern Hebrew parser and for frame building: a nifʿal verb with an obj dependent, or a puʿal verb with an agentive nsubj, is flagged…

  Result: tried on Biblical binyanim in tools/hbo: passive stems (pual, hophal, qal passive, pulal, polal) take obj 4 times in 460 531 training words, and gparse's output breaks it 2 times in 40 657 words (Chronicles e2e), 0 on dev: a hard check has nothing to fix; not built

  from Voice at the Interfaces: The Syntax, Semantics, and Morphology of the Hebrew Verb I1 · shelf he · 2026-10-18
• measured, no gainAdd a bridging pass after coreference in coref/src/sieve.rs, with high-precision rules from the paper's list, only for mentions the sieves left unlinked and only for unmodified definites (the N): building part → the latest building (house, castle, hut … the door, the roof, the window), body part → the latest…

  Result: bridging for unmodified «the N» (building part, relative, body part → latest building/person, ≤3 sentences) scored against GUM Bridge=: 3 links on dev (1 right) of 697 gold; not committed

  from Bridging resolution: Task definition, corpus resources and rule-based experiments I1 · shelf en · 2026-10-18
• measured, no gainMDFirst as a re-ranker: keep the top-B segmentations from the Hebrew segmenter and let the parser rescore each one with its tree score, then take the best joint result. The first target is the segmenter in tools/hbo.

  Result: top-k segmentations rescored by gparse's tree score (hbo lattice/choose): no gain beyond noise on Chronicles (unpointed 79.73 -> 79.58..79.99, pointed 85.24 -> 85.10..85.27); the parser was trained on gold words only; commit fc707558a

  from Joint Transition-Based Models for Morpho-Syntactic Parsing: Parsing Strategies for MRLs and a Case Study from Modern Hebrew I1 · shelf he · 2026-10-18
• measured, no gainen/src/parse.rs: add a dynamic oracle for arc-standard that charges each action by the number of gold arcs it makes unreachable, and train with it; measure cv LAS on the train folds against the static oracle, two runs each.

  Result: approximate arc-standard dynamic oracle with exploration, greedy parser, EWT dev our tags: 77.67 -> 76.40 (variants 76.04-76.49; 20 epochs 77.50 -> 76.47); recorded in d80e79ef9

  from Approximate Dynamic Oracle for Dependency Parsing with Reinforcement Learning I1 · shelf en · 2026-10-18
• measured, no gainTag-conditioned lemma rules in en/src/morph.rs: when the dictionary has several candidates or a suffix rule is ambiguous, choose the rule with the UPOS and FEATS predicted by en/src/ud.rs; measure lemma accuracy on EWT dev, reporting ambiguous tokens (more than one lemma) separately from the rest.

  Result: lemma by predicted UPOS and relation: EWT dev lemmas 96.67 -> 96.68 (3 tokens); ambiguous forms 97.66% vs the rest 95.67% (split kept in ud-eval EN_UD_DEBUG); commit d2b1fb9a8

  from A Simple Joint Model for Improved Contextual Neural Lemmatization I1 · shelf en · 2026-10-18
• measured, no gainImplement the discourse-deixis baseline as a sieve in coref/src/sieve.rs (the disabled demonstratives slot is the place): this/that (and it as subject of happened, was true, means) not followed by a noun → the preceding clause, as an event mention with a VERB head (coref/src/mention.rs). It directly…

  Result: deixis sieve (event «it», bare this/that → previous clause, as an added clause mention): 17 of 141 right on GUM dev, −0.76/−0.09 CoNLL gold; in coref, off by default (2bb818bf8)

  from The CODI-CRAC 2022 Shared Task on Anaphora, Bridging, and Discourse Deixis in Dialogue I1 · shelf en · 2026-10-18