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 — Precompile the tokenization into each lexicon entry, as CALIMA Star does, rather than regenerating it after analysis. For, store the split of each prefix-stem-suffix triplet (the lex, pre tables) with its clitic segmentation, so the segmenter reads the split from the table instead of recomputing it. (the paper)
to try — Baseline rows: every historical test table should start with the identity baseline and a dictionary-only lookup, as the paper's zero-shot comparison does. Report both with the date of the run. (the paper)
to try — Out-of-vocabulary policy: for an unknown token, offer the most likely analyses seen in training (the paper uses ten), and measure a domain lexicon extension as an ablation, reporting the change without claiming a gain before two-run checks. (the paper)
to try — Build a vector or a feature for an unseen Hebrew word from its parts: the longest stem-like part plus the averaged rest, with a fixed weight, is a rule that needs only the segmenter. For, use it to give an unknown lemma the features of its prefix and suffix pieces when the stem is missing from the lexicon. (the paper)
to try — A review protocol: sample 200 neighbour pairs from a distributional graph, have a human judge them, and report precision by source, the way the paper does. Use it for any lexicon-neighbour suggestion before it is shown to a reader. (the paper)
to try — Jaro similarity with the paper's thresholds (0.67 Hebrew-Hebrew, 0.82 Aramaic-Aramaic, 0.78 Aramaic-Hebrew) as a deterministic filter for candidate gloss pairs. Test it in on the Onkelos gloss pairs, against a hand-checked sample. (the paper)
to try — The word-level measure: represent each word by its two rarest letters, order, after dropping prefixes and matres lectionis, and require an exact match. This links Aramaic spelling variants to one lemma. (the paper)
to try — A Topk gate for lexicon-level embeddings: for a set of Hebrew categories (biblical figures, plants, body parts) take k=3 nearest neighbours of each word and score how many fall in the same category; use the score to choose window and dimension for any embedding the SLM builds from a small corpus. (the paper)
to try — Build the category sets from the Wikidata data we already hold (outside the repo,) for the Biblical and Mishnaic lexicon, and check them against dictionary hebrew lemmas before use. (the paper)
to try — Keep the top-5 beam candidates and let a word-level lookup choose among them, which the authors suggest; a dictionary check is deterministic and can say which candidate it kept. (the paper)
to try — Measure out-of-vocabulary rate for morpheme-level vocabularies. The paper segments transcripts into morphemes and finds this helps, but OOV words remain. In, count OOV on held-out unpointed text for a word-level lexicon against the segmenter's morpheme lexicon. (the paper)
to try — Lexicon expansion with a polarity check: when a base word is expanded to all its inflections, drop any unpointed form that collides with a word of the opposite class (the paper's examples are a pair meaning "sun" and "anger" and a pair meaning "relaxed" and "calm down" that share a root). The check is a table lookup over unpointed forms. (the paper)
to try — Count-based baseline first: on about 2,600 posts, unigram counts plus a sentiment feature beat fine-tuned transformers. Before any Hebrew model is claimed to help, Mova needs this kind of baseline, and a sentiment score from a lexicon can be tested as one more feature. (the paper)
to try — Route by variety. A small identifier chooses which disambiguator runs. For Aramaic, already labels segments (arc, arc-x-targum, tmr, jpa, arc-x-zohar); a rule-based or small statistical identifier could choose the variety-specific lexicon and tests. (the paper)
to try — Use the paper's per-book coverage (Genesis 64.2%) as the first yardstick for the lexicon: report coverage per book, and treat any drop as a regression. (the paper)
to try — Keep the morphology-only model as a control: a feature set that drops the lexicon should lose accuracy; if it does not, the style signal is not what we think. (the paper)
to try — Keep only candidates that two independent sources agree on (a character-level guess and a word-level guess). The intersection is a rule Mova can implement with BTreeMap ordering, and it can be run with a dictionary such as dictionary hebrew in place of a neural word model. (the paper)
to try — Link broken plurals to their lemma. A broken plural is then reached from the singular lemma, not from a suffix rule. Test it: take the MACULA noun lemmas, count how many plurals are reached by pattern rather than suffix, and check the links against dictionary hebrew. (the paper)
to try — Seed-based lexicon growth: from a short list of offensive words, take the 100 nearest neighbours by Dice coefficient in an unannotated corpus, then accept candidates by hand and record the rejection rate. This is a reusable Hebrew lexicon build. (the paper)
to try — The paper finds that "hero" has no single Hebrew equivalent (gibbor or nabi). Record this as a lexicon entry in : one English gloss with two Hebrew candidates, and the rule that a query gloss may map to several lemmas. (the paper)
to try — Morpheme-level language tag: an Arabic clitic on a Hebrew stem (the paper's *al-* on *maskilim*) gets its own tag, and the stem gets the other, when the stem is rare in one lexicon and frequent in the other. This is a rule for, which now cuts text only at word level, and would give mixed Hebrew and Aramaic words a split. (the paper)
to try — Lexical-linking test: build a small gold set of token-to-entry links (a word in a verse linked to its lexicon entry, as in the *shemà* example) and measure the lemma output of against it. The paper has no measures, so this set is the part Mova can add. (the paper)
to try — Confidence tiers with a stated precision: the paper's thresholds give three tiers (precision 100%, 84%, 54%), each with a fixed hedge in the gloss ("likely", "perhaps"). Mova can output the same three tiers with the rule that fired, so the gloss says why. (the paper)
to try — Paired readings in the lexicon: for each ambiguous form, store both analyses (a bare verb, and a verb with a person-gender object clitic) with the clitic's features. The paper's examples ("they ordered" against "he ordered it") make good seed pairs. In UD this is a verb with a pronoun object merged in the surface form, so the segmenter has to split it. (the paper)
to try — Match predicted lemmas to gold through a fixed, documented normalisation (macron and circumflex, optional final mimation, optional initial waw), applied at evaluation time only. For Hebrew, the same kind of table can normalise spelling variants of lemmas in dictionary hebrew before scoring. (the paper)
to try — Report every score on the same four splits (all, function, non-function, in-vocabulary, OOV). The gap between in-vocabulary and OOV accuracy is the number to watch, and a rule system can be tested on it separately. The dictionary baseline's 0.00 OOV shows what a pure lookup cannot do. (the paper)
to try — A dubious-stem check: generate root × pattern combinations and count how many occur in the MACULA lexicon. The paper's method (count the combinations that are not words, then review by hand) is a measure of how far the SLM over-generates in morphology. (the paper)
to try — Character trigrams plus TF–IDF as the default feature for forms the lexicon does not know. The paper finds these competitive with BERT on fragmentary text, so the SLM can use them for unknown or damaged words in before it reaches for a neural model. (the paper)
to try — A positional inverted index over consonant skeletons, with a fixed edit-distance threshold, to find candidate parallels. It is a table and a test, and it can use the morphological lemma lexicon that the paper's pipeline also builds. (the paper)
to try — Keep one lemma table per convention and fall back to a merged table when a token is missing from the domain table (the paper's domain-aware dictionary: about +5 points over one table). This is a deterministic, testable step for any multi-edition lexicon, such as. (the paper)
to try — Keep the layers separate, keyed by token id: a morphology rule can be re-run without touching the text, the same split that dictionary hebrew already makes between lexicon and treebank. (the paper)
to try — A rule-based candidate generator: take all pointed forms of each consonantal word from a lexicon with frequencies (for example from dictionary hebrew), filter them by the shape of the word, and rank them by frequency. Measure coverage (how often the right form is in the set) against candidate-set size, as the paper does; its curve suggests three to four candidates. (the paper)
to try — Operationalise offensiveness in Mova as four checkable features from the tree: intent (a verb or predicate aimed at a person), target (a person or group referred to by a pronoun or noun), tone (a lexicon hit), impact (a second rule). Each rule records the reason it fired, so a decision can be explained. (the paper)
to try — A deterministic table of rabbinic abbreviations (letter pairs with a gershayim, such as the ones for "perek kama" and "amud aleph") as lexicon anchors for the Rabbinic layer, tested on a held-out set of citations. (the paper)
to try — A flag-then-decide step after the segmenter in : each prefix split is checked by a rule (or by a tiny classifier trained on a held-out slice of MACULA) that flags splits no lexicon entry supports; a flagged split falls back to the unsplit form and the log says why. Measure flag precision and recall against MACULA gold on verses kept out of training. (the paper)
to try — Acceptance gate with per-criterion numbers: for each generated sentence, report the pronoun form, the verb ending and the lexical attestedness (is each word in the Early Modern lexicon?) as separate pass or fail values, and admit a sentence only when all pass. Mirrors the paper's threshold of 7 but is explainable. (the paper)
to try — : keep a table of verb-particle pairs with their period and register, from the paper's lists. Use it as a scored multi-word lexicon for the parser, so that the particle is attached to the verb as compound:prt. Measure particle attachment on UD EWT dev, run twice. (the paper)
to try — A contrastive suite for the translator's accuracy errors: for each error the translator can make (thou against you, the name of the LORD, negation, number agreement, a dropped clause, an invented name), write a correct and a wrong output and count how often the scorer prefers the correct one. Keep it as a regression gate and add a red control (a shuffled lexicon). (the paper)
to try — Reject focus on scalar quantifiers that lie outside the negation's scope, and on neg-raising verbs such as "believe", using small lexicons. Count each rejection with its rule. (the paper)
to try — : an audit of the English irregular list for lemmas whose present and simple past share a stem while the participle differs (*beat, beat, beaten* is one). Expect very few; report the count and the list, and keep the test in the lexicon checks. (the paper)
to try — Newman's three-way comparison (source, prose literal version, verse rendering) is the translator's baseline table: MACULA glosses word by word, a literal Bible and the system. Keep all three in every Hebrew report, as already does for the glosses. (the paper)
to try — : make a cross-table between Mova's dictionary senses and WordNet synsets for the verbs in the school lessons. Count one-to-one, one-to-many and many-to-one matches as the paper does. The cells with many-to-one matches are the places where the lexicon needs a new sense. (the paper)
to try — Attach a frame name to each sense, so a school word can be asked for its frame. Measure how much of the school vocabulary gets a frame. (the paper)
to try — A quality protocol: two or more readers tag the same sample of about 800 words independently, and the full and pairwise agreement are reported before a lexicon is called gold. (the paper)
to try — Use their reasons for "no solution" as the error log of the lexicon: adjective-versus-noun confusion, typos, passive forms, dialectal forms, grammar problems, names, numerals. logs every unparsed word under one of these classes; the counts per class are the to-do list for the lexicon, and the dialectal class is a route to the Mishnaic segment. (the paper)
to try — Function-tag style second stage: after en parses a sentence, label each dependent with a grammatical group (syntactic, semantic, direction, temporal, as in the paper's Table 1) with a small classifier on the UD relation and the lexicon, and measure F1 per group on PTB section 23 (test only). (the paper)
to try — Compound lexicon feature in : add a feature from the AGID and UD lemma lists that records whether a noun pair is a known compound, which is the lexical information the paper says the parser lacked, and measure LAS on the NP-internal subset of EWT. (the paper)
to try — Partitive and contraction blacklist: check whether the cases the paper blacklists (de la, du, and similar) are covered or, and count the sentences that change when those rules are added (dev set). (the paper)
to try — Light-verb marking in the en lexicon: mark light verb constructions (take a shower, make a wish, have a nap) with compound:lvc in the mova dialect, check that accepts the relation for English, and measure LAS on EWT test restricted to the tokens of those constructions. (the paper)
to try — Principle D as a query: have en write the MWE type as a MISC feature and check that every idiom in is retrievable as one set; measure the share of EWT and GUM idioms that the rule finds (test only for GUM). (the paper)
to try — Adjective order in : order prenominal adjectives by a subjectivity score from Mova's own lexicon rating (not the paper's crowd data), and measure accuracy on the 155 EWT AAN triples (test only) against the current rule. (the paper)
to try — A count form of the translator's lexicon: the source chunk (Hebrew) plays the role of the foreign phrase, the English chunks aligned to it are the candidate paraphrases, and the probability is a product of two counts. No neural model is involved, and each merge carries its counts. (the paper)
to try — Keep the three-level plausibility marker: store for each lexicon word a level for each tag (common, possible, rare), in a BTreeMap, and use the level as the cost in the tag decision. Report the share of tokens decided by the marker alone. (the paper)
to try — Verb subcategories from evidence: a count-based frame table per verb, drawn from UD trees (object, prepositional, clausal complement), with the frame tested against a lexicographer's list. Use the paper's eleven-class scheme as the first pass. (the paper)
to try — and : find pairs of English words that are close in meaning but not in complementary distribution (as *fela* and *manig* were), and record which one a school text uses. This is a lexicon-level check, not a grammar rule. (the paper)
to try — "Shards of one vessel" as a lexicon test across more than two authors. A concept is well formed when its members across the four Bibles and the Hebrew complement each other. Each author's member set covers part of the concept's contexts, and together they cover all of them. Report for each concept the share of its contexts covered by one author alone versus by the union. Concepts where one author covers everything are spelling variants. Concepts where coverage is split are real "ways of meaning". (the paper)
to try — The interlinear as an ideal endpoint, not a baseline. Exact mode already gives a fully literal path: source words, MACULA glosses and target words, aligned by verse. Publish it as an interlinear view in the run's : Hebrew word, its meta tokens, the target chunk, and the residual class of each. That is the most transparent form of "why" the translator can give. (the paper)
to try — Compose alignments instead of building a new aligner. Chain the reader's Hebrew-to-meta alignment with the meta-to-English links of the lexicon, which gives a Hebrew-to-English alignment for error counts, in the spirit of the via-source option. (the paper)
to try — Mark concepts that are theologically marked in the meta-lexicon, so that the encoder does not silently normalise them; count how many of them change between standard and author. (the paper)
to try — Audit the lexicon the same way: sample a fixed number of frequent merges, inspect them, and report the share that are noise. The paper's error rate of about 12 percent is a useful baseline to beat. Run the audit after each merge round. (the paper)
to try — Use the Scots POS corpus as a measure for : run the English tagger on it, report accuracy by tag and by the words that are not, and keep the result as a note on what the dictionary misses. (the paper)
to try — Sense-level frames in : store the complement frame per word sense, not per entry, because the paper shows that frames in one entry are often split across senses. Check each sense with its own frame when the pupil dictionary gives the sense. (the paper)
to try — Provenance queries: add a command that answers which verse, which merge and which author produced a given meta token or output word, as ARIANE answers which lexical units come from which dictionary. It helps debugging the lexicon (open problem 2). (the paper)
to try — Require an English gloss for every concept before a merge is accepted, and compare the glosses of the two members. Pairs whose glosses are far apart are rejected, a cheap form of the duplicate check that the paper proposes. (the paper)
to try — Read light verb constructions as frames of the noun's verb: «the fox made an offer to the crow» → Event fox offer recipient=crow, «the bear took a nap» → Event bear nap, «gave the dog a bath» →. Use the noun→verb links of (WordNet) and the light-verb entries of the expression base (already tags LVCs as light-verb); the subject keeps the agent, a dative or *to* object becomes the noun's recipient or patient by the noun verb's VerbNet frame. Print How as «light verb make + noun offer». Measure frame bench verb/patient F1 and world knowledge diagnose on questions that use the plain verb («What did the fox offer?»). (the paper)
to try — Before a phrase merge, record how many function words it contains and how many bits the residual changes when the phrase is split. Report both for every merge. This is the gate that the lexicon needs. (the paper)
to try — Keep MWE types as a typed table in the translator's lexicon: fixed, light-verb, particle, idiom, proverb, each with its own merge rule. A fixed phrase may merge into one concept, while a light-verb pattern must keep its predicate's arguments. This addresses open problem 2: today phrases such as *unto_thou* merge function words into one concept. (the paper)
to try — Disjoint verb classes in : tag each verb as manner or result (the school dictionary already has word senses), using the disjoint split and not Vendler's overlapping one. Check that every verb gets exactly one class. (the paper)
to try — Regression gate for the morphology: when the regular verbs are split into classes, the generated forms must be byte-identical to the unsplit lexicon for every verb. Run it twice. (the paper)
to try — Keep the list of additive and focusing particles in the lexicon as a closed BTreeSet, with the presupposition each contributes, so that the trigger list is explicit and checkable. (the paper)
to try — Explicit lexical rules: hyponym pairs entail, sister terms contradict. Keep them as a small lexicon, and log each rule that fires. (the paper)
to try — Pairwise ranking test. Take five systems per verse set (copy, word-by-word table, meta-world, delta, the MACULA gloss baseline) and count how many pairwise comparisons each metric gets right against gold. This is cheap and deterministic. (the paper)
to try — Generate candidate multi-word chunks only from spans that occur in the input text, not from every combination of dictionary entries. For the encoder, this means building chunks from the anchors of the frames in the sentence, which avoids the combinatorial junk the paper warns about. (the paper)
to try — Give multi-word units a fixed alignment instead of letting the aligner split them. The paper leaves them unaligned and gives them a weight of 1.0. In the Hebrew reader, a dictionary multi-word unit could be fixed before IBM Model 1 runs. (the paper)
to try — The sense map: write a checked table from the world's concept names to the WordNet-style senses of the dictionary, and keep every mapping that has more than one candidate in a list for hand review, since the paper's mapping step resolves ties by frequency. (the paper)
to try — and : keep the sense-to-word choice as two stages, a fixed slot filter first and a count-based example score second, each stage writing its reasons into the output. Ties are broken by lemma order in a BTreeMap. (the paper)
to try — Duplicate words cause 17% of the errors, because the lexicon key is the word alone. Adding the tree relation or the role to the key would separate the two occurrences of an ordinal such as *first*. Test this as a change to the key and measure the residual. (the paper)
to try — Keep only the rules that match the dev verses, as the paper does with its filter. For the residual-driven lexicon (open problem 2), keep a merge only if it occurs in the verses it is meant to read; the rest stay out of the table. (the paper)
to try — Length limits as parameters. Make the rule length and gap limits explicit parameters of the lexicon, and report the rule count and the residual for each setting, so the gain of a larger limit is shown with its cost. (the paper)
to try — Compare exact and fuzzy extraction: the paper finds that exact tree-to-tree rules are worse than fuzzy ones. For the lexicon, count the concept merges that a strict rule (same concept, same tag, same head) would allow, and compare with a fuzzy rule (same concept, tags may differ), on the same dev split. (the paper)
to try — Symmetric closure as a lexicon check: wherever the pupil dictionary or the archaic table links A→B (an archaic form to its modern lemma, or a sense to a synonym), test whether B→A is present. Report the missing reverse links as a gap list instead of filling them silently. (the paper)
to try — Verse alignment as free supervision for the biblical register: the KJV, the JPS 1917 Torah and a modern open translation share verse ids. Mova can align their tokens with a deterministic score (identity, dictionary link, relative Levenshtein, relative position), the paper's measures without the neural parts. It can then report which Early Modern forms (hath, spake, unto) line up with which modern ones. The alignment table is a checkable source for entries, with a verse id for each. (the paper)
to try — Sibling-form voting for the unpointed mode of : when a token has several analyses, count how often each candidate lemma appears with unambiguous forms in the corpus, choose the analysis that its siblings support, and leave the token ambiguous otherwise. The lemma list can come from dictionary hebrew. (the paper)
to try — and onto: map the verb of each PP to its VerbNet class and the PP-internal noun to its WordNet lexicographer file, and feed both as features to the en perceptron. The paper finds classes more useful than single words; measure LAS and PP accuracy on EWT test. (the paper)
to try — A deterministic root guesser for dictionary hebrew, which has no roots yet: per-radical rule tables with the weak-paradigm rules (a first radical n or ʾ may vanish, a second w or y, a third y, a doubled second and third) generate candidates, a root list built from the lemmas filters them, and edit distance ranks them. This uses the lexicon instead of a classifier. (the paper)
to try — Keep their split of results by paradigm (regular, weak, mixed) and their 200-word hard set (a missing radical or metathesis) as a test of the root step in dictionary hebrew. (the paper)
to try — Lexicon coverage split: as the paper does, split EWT test sentences into those whose words all have the right tags and those that do not. Report LAS for each group with en lexicon output. (the paper)
to try — Add a second metric to en morph-eval beside accuracy, for example the share of lemma errors that are for words in the dictionary, since the paper argues accuracy alone does not separate the models. (the paper)
to try — Filter a concept merge by its evidence: keep a lexicon entry only if its pair probability from aligned counts is above a threshold and the part of speech of both sides agrees (the paper's two filters). This applies to the lexicon in (open problem 2). (the paper)
to try — : find intensifiers without a list. Take the base and comparative pairs in our own lexicon (hard → harder), and mark each adverb that links a base form to a stronger one, in a deterministic count table (BTreeMap). Check the list by hand against 50 known intensifiers and 50 de-intensifiers. (the paper)
to try — Calibrate the similarity against a null distribution. The paper sets its threshold at the mean plus three standard deviations of unrelated pairs. For Mova, which is count-based, the null comes from the shuffled-lexicon control: the score of a translation under a shuffled lexicon gives the floor a result must clear. (the paper)
to try — Replace unknown words by a typed placeholder in a comparator run. This separates errors caused by a missing lexicon entry from errors of structure. Mova's residual can report unknown-concept tokens as their own class. (the paper)
to try — Coverage as the headline measure: report the out-of-vocabulary rate of the Hebrew analyzer on a large unseen corpus, split by period, not the size of the lexicon. (the paper)
to try — Audit protocol: keep only the pairs that two independent methods accept, then check a sample by a linguist and report the error rate. Use the same protocol before any auto-built root or cognate list enters the Mova Semitic lexicon. (the paper)
to try — Deterministic cognate gate: the Turchin rule (two words are candidates when their first two consonant classes match) is a table lookup. Apply it to Hebrew, Aramaic and Arabic forms of the same gloss, and measure its precision and recall against a small expert-marked set in the style of the 74 sets. (the paper)
to try — Sound-class alignment for roots: map the consonants of Hebrew and Aramaic lemmas to sound classes and align them with SCA-style scoring, to find forms that share a root. This is a candidate route to root data for the dictionary hebrew lemmas, which have none yet. (the paper)
to try — A small glossary of archaic English spellings taken from the orthography section, to test for coverage. Only the list of forms and their English meanings may be used; the wording is not copied. (the paper)
to try — Build a lexicon of English fixed expressions (as well, as well as, in spite and similar) and a rule that attaches the later words as fixed to the first,. Measure the fixed arc accuracy on EWT dev before and after, run twice. (the paper)
to try — : the English strong verbs with a distinct past participle (ABC, for example "drive, drove, driven") can be listed with a feature "past and participle differ". Test whether the feature predicts that the verb keeps its irregular form in the lexicon, using an open frequency list; keep the list as a count table. (the paper)
to try — Read marker strings as one decision: *and then* → next, *and so* → cause, *but then* → contrast + next; the second marker decides the link. A short list in the connective table (see das-2018-constructing-lexicon-english-discourse-connectives I1). (the paper)

This page in Markdown → https://md.dictionary.elamur.ai/free/en/ideas/page/2/