Ideas behind the dictionary

673 ideas from the papers Mova reads that touch dictionaries — lexicons, glosses, word senses — each with the paper it comes from and what came of it when Mova tried it. Those taken up first, then those still to try.

to try — : plural formation by semantic class as a table of counts, not a vector. For each class (fruit, animal, person, vehicle) count which plural ending and which exceptions occur in a lexicon, then predict the ending for an unseen noun by class majority, falling back to -s. Report the accuracy by class against the default -s rule. (the paper)
to try — Frequency-rank control. The decipherment baseline maps words by frequency rank. For the meta-lexicon learned on two Bible translations, a rank-matching control must be beaten before the LCS anchors and IBM steps get credit. (the paper)
to try — Style nodes in the meta-world. Each meta token already has a concept. Mark the tokens whose concept has different chunks in the two authors' lexicons, and rewrite only those, keeping the frame and order fixed. Measure exact-match change per author pair. (the paper)
to try — Keep two rule tiers apart, as the paper does: a hand-written table for constructions that are hard to learn (passive, tense, order) and an acquired table for word substitutions. In Mova the first tier is the job and, the second is the lexicon learner. Count how many tokens each tier fixes on dev, run twice. (the paper)
to try — Lexicon from aligned pairs with a residual check. The transfer database comes from aligned pairs, as the meta-world lexicon does from Bible pairs. The paper's results support problem 2: keep a concept merge only if it lowers the reader and encoder bits. (the paper)
to try — Build a degree lexicon with three ranks (down-toner, neutral high degree, full degree) for the adverbs in the book. Tag the adverbs in UD trees as advmod heads of an adjective, and check the rank order on held-out pairs (for example "quite" below "very" for most adjectives). (the paper)
to try — Label each intensive with a register tag (colloquial, vulgar, neutral) in the lexicon, and let the school student use the tag when a question asks for a formal register. (the paper)
to try — A default-and-override table: default to the relation-based construction, with a list of frozen compounds that must not take it. The frozen list belongs, the default. (the paper)
to try — Write the paper's lexical pass as a deterministic rule list: a Levenshtein distance of at most 15% of the longer word, morphological negation, and the negation rule. These rules live in Rust, are counted in the report, and run before any learned lexicon. (the paper)
to try — and : a deterministic polarity score for Mova's adjectives, built from four counts (how-question neutrality, ratio-phrase counts, frequency, a valence list) and combined as a first principal component, with lexicon order fixed by BTreeMap. Check the licence of the valence and corpus data before use. (the paper)
to try — Let hand-written overrides act with precedence over learned tables, as SYSTRAN's glossary entries do, and measure them as a separate layer: how many tokens each override changes and whether the change is kept by the residual. (the paper)
to try — Add the ten adjectives used in the paper (stupid, smart, wise, fortunate, lucky, brave, polite, mean, foolish, rude) to the lexicon as one entry each, with no per-frame duplicates. (the paper)
to try — Attribute errors to a layer. The paper says it could not tell analysis, lexicon and generation apart. The residual by layer that Mova already measures answers that question, and the paper shows why it is needed. (the paper)
to try — : a symbol class for mathematical signs (for example ≦, ∈) in the lexicon, measured by XPOS accuracy on the GENTLE proof genre (test only). (the paper)
to try — Use the paper's coarsest filter as a merge test. A meta-concept merge should be kept only if its dominant member is at least half of the word's total count. Then measure the reader and encoder bits and without the test, which is the residual-driven lexicon of open problem 2. (the paper)
to try — The three laws as a lexicographic order over the residual. When the encoder in chooses between candidate chunks, it ranks them by world-level agreement first, the author delta's fit second and the stack's fluency third, and never trades a higher level for a lower one. In normal mode, report the three in that order: world agreement, delta bits, stack bits. (the paper)
to try — Tytler's split between general and particular idioms is the split between the standard encoder and the lexicon. General idioms are rules of the target language: order, aux + participle, *to* + infinitive. They belong and, the same for every author. Particular idioms are multiword concepts. Test it this way: an override in an author delta that is only a general idiom of the language should move into the standard. (the paper)
to try — Enhanced++ partitive and multiword-preposition rules, with the expression lists; measure the share of nmod and det arcs in EWT test that change, and the triple precision on 100 sampled sentences checked by hand. (the paper)
to try — en/src/morph.rs: test a suffix guesser for words missing from the dictionary (analyses per fixed-length suffix, with the shortest lemma rule) and measure UPOS accuracy on out-of-dictionary tokens of EWT dev through en tag-eval. (the paper)
to try — Lexicon as an input: the paper's central claim is that a UD-compatible lexicon (CoNLL-UL) matters more than the parser for Hebrew. Measure whether a lattice built from the lex table beats a data-driven one on unpointed text, with two identical runs before any claim. (the paper)
to try — en/src/morph.rs: add suffix-rule splitting for contractions that the dictionary misses, applied only when a rule does not fire too often on the training data, and measure LAS on ATIS (currently 63.0, lowercase queries). (the paper)
to try — Number constructs by a small table: the noun heads a numeral of three and above, and a multi-word number such as "thousand and forty" is compound inside one headed noun. Use the numeral table of the Hebrew lexicon to write these rules and. (the paper)
to try — with : a dictionary of imported words with their UPOS (the paper's Table A4 gives a starting list), consulted before the suffix rules; measure UPOS accuracy on tokens that are outside the EWT vocabulary. (the paper)
to try — : a fallback lemmatizer for forms that are missing from the dictionary. It uses suffix edit rules counted and. Measure lemma accuracy on EWT test out-of-dictionary forms only, against the dictionary-only route. (the paper)
to try — : a seed list of AAE and Twitter spellings (da/tha, dis, ion, yo, finna, bouta, tryna, gonna) mapped to standard lemmas and UPOS; metric: lemma and UPOS accuracy on the AA subset, and the number of variant tokens that change tag. (the paper)
to try — Modal verbs as a closed list: the paper's Verb-Type=Mod marking depends on a lexical class. A short table in the Hebrew verb notes, checked against the lexicon. (the paper)
to try — A CoNLL-UL-style export of the tables (lex, pre) with one row per analysis (form, lemma, UPOS, features), so that other parsers can read the same lattice and the lexicon can be shared. (the paper)
to try — Make the lexicon the measured variable: a ladder of runs on a Hebrew dev set, each step with a single change (no lexicon, lexicon, lexicon plus forced gold path, gold morphology). The paper's numbers are the reference shape for the ladder. Run every step twice before reporting. (the paper)
to try — Edit-script fallback for lemmas: classify the remaining lemma errors of on EWT test into edit-script classes (suffix change, prefix change, stem change), and test a learned edit-script classifier only for words the dictionary and suffix rules miss. (the paper)
to try — Lexicon-first multi-word expansion: check how handles contractions and clitics (don't, can't, gonna) against EWT test tokens, and count the forms the lexicon misses; the paper uses a lexicon with a model as fallback. (the paper)
to try — Verbal discourse markers in : treat I mean, I guess and you know as discourse, not parataxis, as the paper proposes, and measure LAS on the discourse relation in the ESLSpok test set. (the paper)
to try — Train on all genres together but measure per genre, and keep the genre-specific embedding/lexicon choice as a switch: the period-specific model helps its own period and hurts the other, so a single "best" setting would hide the trade-off. (the paper)
to try — : add predicate-operator entries (=, <, >) with a verbal tag, then measure UPOS accuracy of en tag-eval on math-bearing sentences. (the paper)
to try — Lemmatisation transfer between related languages (the paper finds Classical Latin and Medieval Latin, and Old Church Slavonic and Old East Slavic, work across each other). A test that checks Hebrew and Aramaic lemma pairs through the shared lexicon, as a measure of coverage, before any transfer is attempted. (the paper)
to try — Build Mova's hedge lexicon the same way, as a table of term → hedge sense, non-hedge sense and one example of each, stored next. Each pair of examples doubles as a minimal-pair unit test for the disambiguation rules in (for example «there are about ten» vs «talk about Mark»). (the paper)
to try — Use the overlap test of the SMM as a merge check for the lexicon. Keep a concept merge only when the two words' sets of translations overlap above a threshold, in both directions. Test it on the known noisy merges (*be* with *have*, phrases with function words). Measure bits of reader+encoder residual and without the check, twice, to see whether it removes the noise. (the paper)
to try — Define a deterministic overload count: for each sense, the number of non-semantic tokens shown with it. Report the mean per entry on the school dictionary and track it across builds. (the paper)
to try — Put non-semantic material (etymology, derived forms) behind a separate field in the pupil dictionary, so that the sense lines carry only meaning and examples. Measure the SLM student's multiple-choice accuracy and without the extra fields; a drop would show that the fields are needed for the lesson. (the paper)
to try — and : when a headword has senses of different origin (the paper's "bore" as drill and as tedious), keep one entry per origin with an enum ID, and one entry per part of speech otherwise. Measure the number of split entries and the school answer accuracy before and after. (the paper)
to try — Do not merge concepts across senses in the lexicon. A merge of a literal "shadow" with a figurative "protection" would be the kind of noise open problem 2 is about. Test by keeping the two senses apart and checking the residual in bits. (the paper)
to try — Treat the be-type disambiguation as a sense class. The paper's example, two Spanish verbs for one English verb, is the analogue of merging two concepts that differ in use, such as *be* and *have* in the current lexicon (open problem 2). (the paper)
to try — Morphological heuristics for unknown words in : ordered rules on the form of the word (ending, prefix) that propose a reading, with a flag for the rule that made it. Measure the accuracy on words missing from the lexicon. (the paper)
to try — A demonstrator in one narrow domain (appointment scheduling) before broad text. For Mova: report the translator's residual on narrow, repetitive chapters (the genealogies of 1 Chronicles 1–9, a training book; never the test books) next to the narrative chapters. That separates lexicon gaps from construction gaps. (the paper)
to try — : the head's number of senses, from the lexicon, as a gate on the compositional reading. Heads with more than two senses get a "lexicalised" flag. Measure on English compounds with the standard ratings, if a licensed set can be used; control with a random flag. (the paper)
to try — : the same ambiguity feature on pupil compounds ("birthday cake", "sunflower"). Check that the compound's sense is chosen from the correct head sense in the school questions. (the paper)
to try — Use the one-edit stem rule as a near-match flag in the lexicon report, so that a plural or suffix variant is shown as a near merge and not as a failure. (the paper)
to try — Name handling: a proper name unseen in the lexicon should be passed through character by character. Test the translator's name copy on the name challenge set against a word-level copy. (the paper)
to try — Add the paper's presupposition types, with one search rule per type. (the paper)
to try — Lexicon candidates from divergence. Pairs that the reader treats as identical but whose graphs differ point to concept merges that drop meaning. List them as candidates for the residual-driven lexicon. (the paper)
to try — Transfer through English first. Compare the residual when source words are first replaced by English glosses (MACULA glosses for Hebrew) with the direct reading. The paper's finding predicts a smaller difference for the gloss route. (the paper)
to try — and : store the 12 gold scales as ordered BTreeMap entries with a membership flag, so that word-sense lookups can answer "which is larger, medium or huge" by scale position. Keep the scale file small and build it from the lexicon through. (the paper)
to try — : a region label per sense, stored as a sorted set (BTreeSet), so that a dialect word is available only with its region. Measure coverage against a held-out list of region-tagged words. (the paper)
to try — Negative control: a modern sentence with a word from the glossary ("bide", "bate") must keep its modern reading. The Early Modern reading should switch on only when the text is marked as biblical register. (the paper)
to try — Use the entries as a closed table of Early Modern senses for and the early-modern lexicon, each sense with its verse and its date range. A sense is used only where the entry says it applies, so the choice can be explained. (the paper)
to try — Bracketing control: report bracket precision on a held-out set of verse pairs, and repeat it with a shuffled lexicon. If the shuffled run scores as well, the gate is not red and the test is too weak. (the paper)
to try — Let a lexicon decide when a multiword unit (*give a kiss to*) is one meta token (open problem 2). Keep a merge only if the joint parse with it has fewer inverted productions and the residual bits of the reader and encoder go down. (the paper)
to try — Dynamic repair for the Hebrew reader: when MACULA or attaches a word wrongly, fix the rule or the lexicon entry, regenerate, and count the verses that change. Open problem 1 (reader attachment noise) is then answered with a count of changed verses per fix, not with hand edits. (the paper)
to try — Store frequent phrases with their correct analysis, looked up like dictionary entries. For Mova, this is the phrase level of the lexicon (e.g.), which should carry a fixed parse, not a merged concept with no structure. (the paper)
to try — Run the paper's own control on Mova's merges: the shuffled-lexicon control is not red for close authors, so apply a name-only shuffle and count the merges that still survive. (the paper)
to try — Use the dictionary-substitution baseline as the first control for any author delta. The paper's phrase table of about 68,000 word pairs is a count-based substitution. Before claiming that a delta helps, measure the substitution-only output on the same author pairs and report both. (the paper)
to try — Diagnostic on rule pairs. List the most frequent source-to-target pairs in the meta-lexicon, and mark each as a real paraphrase or a merge error. This gives a direct view of noisy merges such as "be" with "have", which is open problem 2. (the paper)
to try — Use the core weight as a priority in the lexicon work (open problem 2): a merge that breaks the agreement on a core concept counts as a larger residual than one that breaks a function word. (the paper)
to try — Keep each author's concept inventory separate and map it to shared identifiers only through the lexicon, so that an unaligned inventory does not leak into the meta-world. (the paper)
to try — An animacy feature on adjectives and nouns, as the paper suggests, keyed in : "oiled" with an animate subject is a different sense from "oiled" with an inanimate subject. Test on the school senses, with the feature as a deterministic check and not as a learned score. (the paper)
to try — Retrieval miss analysis: count questions whose if-clause content words do not occur in the gold passage (the "sea level" case). Use to map counterfactual wording to passage wording by a fixed lexicon. (the paper)
to try — Treat the 20% accept-rate as a threshold: when a rule's evidence (from the lexicon or a corpus count) is split near 20/80, return "uncertain" instead of a hard factive label, and log the reason. (the paper)
to try — Add these construction presuppositions as rules to Mova's absurdity checks (, the absurdity matrix triggers). Each of the eleven constructions maps to a tree pattern («except» + noun; «in particular» + noun; «a type of» + noun; «more than» / «rather than» with two nominal arguments) and a required relation between the two head nouns, looked up in Mova's noun hierarchy (WordNet hypernyms or the ontology categories). A flout gives a stated reason: «"in particular" needs pugs to be a kind of dogs; animals is a superset». (the paper)
to try — Hedging for unknown words: when a noun in the input is outside the dictionary, prefer a general word ("thing") plus attributes, and count how often a wrong category name is produced for such objects in the school texts. (the paper)
to try — Connective detection as a lexicon with a disambiguation rule: an inventory of explicit connectives (BTreeMap), each with the frame that makes it a connective ("in order to" before a verb, not "in order" as a noun). The paper says the task needs memorised vocabulary, so count coverage first. (the paper)
to try — Split Mova's pronoun precision table (, by form) by the paper's main perspective: cases where agreement alone leaves one candidate in the window vs cases where two or more agreeing candidates remain. Computed deterministically from Mova's own features in on GUM dev. If errors concentrate in the first group, the fault is in features (gender, animacy, number); if in the second, in ranking (Hobbs order, salience). It decides between the two «What next» directions: an animacy lexicon or Lappin–Leass weights. (the paper)
to try — A deterministic perturbation test for en parse-eval: for each school sentence, replace one word by a same-tag word from the lexicon, keeping the rest. The parse should stay the same. Report the share of sentences that change, and run it twice to check that it is identical. (the paper)
to try — Per-tag sensitivity table, like the paper's Table 3: count the attachment errors per replaced tag (adjective, noun, adverb, verb, preposition), and use the table to decide where lexicon entries need more careful tags. (the paper)
to try — The paper's pseudo-source is a reorder-then-gloss translation. Its English-to-Hebrew mirror, which writes an English verse in Hebrew order and glosses it word by word, gives the Hebrew reader a synthetic source with known alignments. Use it as a sanity check and : the aligner should recover the identity alignment on it. Measure the alignment error per class. (the paper)

This page in Markdown → https://md.dictionary.elamur.ai/free/en/ideas/page/7/