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 — Dictionary first, model second: look each transliterated token up in a table of forms already transcribed in the corpus, and call the neural or statistical model only for forms that stay ambiguous. The authors propose this themselves (section 6). In Mova it is a table lookup before any learned step, and it can be measured as coverage and accuracy on a held-out split. (the paper)
to try — Compatibility as flags: suffix sets that are only allowed on some stems (the middle-weak ♂/♀ marks in the paper) are controlled by a flag on the stem, not by duplicating the lexicon. This is a good model for weak verbs whose endings depend on the final radical. (the paper)
to try — A two-reading rule in the particle lexicon: *be-sax ha-kol* + a gradable predicate with an upper-bound scale gives the approximative reading (covert *pos* as focus associate); with an overt focused phrase it gives the exclusive reading. Test it on the paper's examples as gold. (the paper)
to try — A feature bundle for the cline in the lexicon: each article form lists which environments it may occur, so that a rule can read off the weakest article that fits a context. (the paper)
to try — A root search as a test for the lexicon: the wildcard search for the root of "to become" in the lemma list finds 197 forms in the text, of which some are false hits from a different root. A rule with the root fixed by the lemma table gives a gold test of root extraction. (the paper)
to try — Stem buffers for weak roots: list one stem for a verb whose surface forms alternate (third radical w/y, middle or final weak, geminates) and put the alternation in a buffer class with a condition term, instead of listing each surface stem. This is the cleanest way to write the weak-verb paradigms in Hebrew and Aramaic without multiplying lexicon entries. (the paper)
to try — Choose features by measured information: before writing a rule for a noun class, compute how much each candidate feature (final cluster, vowel count, a stem-final weak radical) tells about the class, conditioned on gender, and keep the features that add something beyond the others. This replaces intuition with a number on the lexicon. (the paper)
to try — Keep vowel colour "unspecified" in the lexicon and fill it by a default rule only after segmentation, as Puech does. The unpointed mode of can then report which vowels were guessed by default. (the paper)
to try — An ambiguity count for the unpointed mode of : for each unpointed consonantal skeleton in the Biblical Hebrew lexicon (dictionary hebrew), count the vocalised forms it can stand for. The list shows which skeletons need context and which the segmenter may leave open. This is a measurement, not a rule. (the paper)
to try — A generation constraint in the verb paradigms: a nonce or new root never receives a transitive in *niCCaC* or *hitCaCeC*. The constraint is one table entry, and the per-root template pairs go into the lexicon as listed data. (the paper)
to try — Per-word fuzzy lexicon lookup: for each source word, return the top entries whose normalised Levenshtein similarity is at least 0.5. For unpointed Hebrew, test this against the lex table of on Genesis, as a deterministic table lookup in Rust. (the paper)
to try — Gloss in the prompt: their third prompt, which gives the base word's meaning, lowers both error and hallucination for most models. If a teacher is used to label pointed forms, include the gloss. (the paper)
to try — Headword-chain matching on meta trees: take the head-modifier chains of the reference's trees, read by the same reader, and check that each chain appears in the output's trees in the same order. This measures open problem 3, whether the encoder writes trees. (the paper)
to try — Treat ordinal and sequence word pairs as next signals across sentences: *first … then/next/finally*, *at first … later*, numbered steps in how-to lessons. A short closed list in the connective data (see das-2018-constructing-lexicon-english-discourse-connectives I1); measure next precision on lessons from the how-to-like courses. (the paper)
to try — Use the predicate alignment as a check on the lexicon (open problem 2). When a concept merge makes two predicates align that the frames treat differently, the merge is counted as an error. (the paper)
to try — Run a feature ablation as the paper does (remove one feature, retrain, compare). The paper finds that adverbials and the infinitive matter most, so a lexicon-level feature may carry most of the signal; test that with a deterministic perceptron, run twice. (the paper)
to try — An entailment rule in : when a causative event (agent causes theme to change) is in the world, add the change-of-state event for the theme, if the verb is marked as labile in the lexicon. Mark the added fact as derived, so the answer can say why. (the paper)
to try — Verbs with many equivalents show a lexicon problem: one source word with many targets inflates the data. Check the merges of open problem 2 on exactly such verbs. (the paper)
to try — Use the survey's sources as labels for the translator's lexicon: each concept merge is tagged with the method that proposed it (pivot, parallel, comparable, distributional), and the residual is reported per source. This shows which source of merges reduces the residual and which only adds noise. (the paper)
to try — Alignment agreement as the merge rule: a link is used only when two directional alignments agree. Use the same rule for the lexicon's concept merges (open problem 2) and count how many merges the rule removes, compared with the current count thresholds. (the paper)
to try — Audit the rules for direction: the paper's examples "X feeds Y entails Y feeds X" show symmetric errors. Mark which pairs are symmetric and check them against the dictionary. (the paper)
to try — Report paraphrase coverage by word class, as the paper's Table 1 does. Mova can see where its lexicon lacks nouns and adjectives, using the word-class chapter. (the paper)
to try — Write a rule detector for superfluous *do* in pupil or learner sentences: an auxiliary *do* in a positive clause, followed by a bare verb. Label the verb as agentive or stative from a lexicon, and count the classes per text with BTreeMap order. (the paper)
to try — Guard each author delta with the target features it needs. A delta that rewrites a chunk should apply only when the standard encoder's chunk agrees in number and person with the delta's target, the same way the restrictions block gender changes. Test this as a rule over the dictionary features, not over the tokens. (the paper)
to try — Hypernym source for : for nominal pairs with no string overlap, check whether the head of one mention is a hypernym of the other in the lexicon, and count how many shared recall errors that would fix before adding it to the sieve. (the paper)
to try — Align the trees of the two languages for each parallel verse, with priority-ordered criteria (same label first, then POS match, then head alignment). This is an alternative to the longest-common-subsequence anchors of meta.rs, and it can be measured by counting merges and checking a hand-labelled sample against the current lexicon. (the paper)
to try — Separate pair-specific alignments from alignments usable for a standard. The paper's example, an idiom that mentions a different object in each language, is correct for the pair but must not enter the lexicon. Mova's merge needs the same label, so an idiomatic merge stays with its author. (the paper)
to try — One lexicon, two views: keep, per concept, the source-side facts (tag, frame, feature) and the target-side facts in one entry, and let a compile step derive the reader's and the encoder's tables from it. The two sides then cannot drift apart, which is the residual problem the translator already has with separate tables. (the paper)
to try — A two-tier role lookup in the frame reader: (verb, sense) gives a verb-specific thematic role; when the verb is not in the lexicon, fall back to a coarse abstract label (A0 or A1 style). Implement the order, and measure accuracy at each step on school sentences. (the paper)
to try — Learning curve at 444 sentences, as in the paper's reduced setting. It shows where the lexicon and the rules stop helping and the statistics take over. (the paper)
to try — Use typed wh-phrases as answer constraints: «What town…», «Which animal…» require an option whose head is a hyponym of the type (WordNet, as for). An option of the wrong type is unsupported however well its words overlap. Measure on «right option outscored» cases. (the paper)
to try — A deterministic span pass for MWEs in en: greedy longest match over the lexicon, with the gaps the paper marks. Measure: link-based F1 on the paper's reviews annotation. That annotation (STREUSLE, over EWT reviews) is CC BY-SA like EWT; keep it as the test, not as a source for the span list. (the paper)
to try — Add a coarse semantic class to nouns and verbs and, filled from WordNet through onto. Measure: the share of correct senses in on a hand-checked sample of 200 EWT review tokens, and without the class. (the paper)
to try — Precompile the meta lexicon into a lookup table keyed by concept and tag, as the paper compiles its mapping rules. The output must stay byte-identical, and the speed gain is measured, not assumed. (the paper)
to try — Build the concept lexicon from counts, then let a linguist review each concept in its contexts. A concept view that lists the verses where each merge occurs, with the residual before and after, is the KWIC browser for the meta-world. (the paper)
to try — Measure the reader's accuracy separately on sentences that satisfy a controlled-grammar test (simple clauses, no coordinated verb phrases) and on the rest. This shows whether the reader's errors come from sentence complexity or from the lexicon. (the paper)
to try — Write the phrase-level rule as a test. A multi-word concept (such as abide) should contain no function-word-only member, the same check that would catch unto_thou in the current lexicon. The test is a count of phrase concepts that fail it. (the paper)
to try — Derivation check in : for each suffix, compare the class of its base with the class the suffix expects (for example -ion with Latinate bases). Flag mismatches as candidate lexicon errors. (the paper)
to try — Use the paper's representativeness score (a log ratio of cluster probabilities) to list the most diagnostic substrings for each class, as a table in en for the pupil dictionary. Keep only substrings with a frequency of at least ten. (the paper)
to try — : add a counted option that joins noun-noun pairs when the pair is frequent in a training corpus and the lexicon has no entry for the joined form. Keep the counts in BTreeMap order and run twice before any claim. (the paper)
to try — : the book treats passive as a lexical operation on argument structure. A passive rule over arg-st-like lists in the English lexicon would make the voice decision explicit; measure with the passive cases of the UD EWT dev set. (the paper)
to try — Character-level features for misspellings: a closed set of edit-distance-1 mappings (Becose to Because, thier to their) to the nearest lexicon form, stored as explicit rules with a reason string, so each recovered tag says which correction produced it. (the paper)
to try — Check the claim that feature-based transfer needs no interlingua by measuring the lexicon size and without the features on the same data. (the paper)
to try — Keep the set of meta path features closed, like the prepositions of a language, and list its size after each lexicon round. Open classes (manners) should stay out of the concept table. (the paper)
to try — Keep the concept dictionary and the instances in two tables. The meta-world concepts (in) should not change with each verse, while the instances for one verse are rebuilt. Count how many concept entries change between runs; that count should be near zero. (the paper)
to try — Check that every open-class word points to exactly one concept. The paper's lexicon links each unit to an ontology term. In, count the chunks mapped to two concepts and the concepts with no chunk; both counts should be reported. (the paper)
to try — Keep regression suites that are never tuned on: one covering the constructions the translator should handle, one covering domain lexicon. Run them on each change, and let a failing case block the change. (the paper)
to try — Generate all perfect and imperfect forms of a list of strong verbs from their roots (roots still have to come from a source; dictionary hebrew has no roots yet), strip the points, and check that recovers each lexeme. The list of failures is a gap list for the lex and pre tables. (the paper)
to try — Report verbs apart from nouns, with the ambiguity beside each score (the paper gives 4.8 for nouns and 10.4 for verbs). A verb with more than ten senses is too fine for a sense list (the paper's example, *keep*, has 22 WordNet senses). Test whether a frame choice from beats the first sense on those verbs; if it does not, the explanation says that the first sense was used. (the paper)
to try — Graded concept match: compare two meta-worlds by a score per concept, with credit for concepts that share a lexicon merge, rather than exact equality. This is the graded Smatch idea applied to the meta-world. (the paper)
to try — The n:m alignment matters for similarity. A meta token may cover two source spans, as in the paper's play-with-music example. Count how many gold n:m cases the current lexicon misses, before changing the merge rule. (the paper)
to try — Report the oracle best-of-two number (rule-based sense versus frame reader) on the school test, as the paper does, to see whether the dictionary carries information the frame reader misses. (the paper)
to try — A control/raising flag in the English lexicon: gets one enum per predicate (the eight types, not strings) with a flag for whether its valent is semantically selected. The enum is built into the dictionary at build time, so it is deterministic. (the paper)
to try — Keep the paper's refusal of null elements: do not add PRO or traces to the parser; the controller is a lexicon flag resolved in the frame, which keeps the output explainable. (the paper)
to try — : keep the unresolved-sense marker, as the paper's ? tag, for senses the rules cannot separate. Exclude those tokens from the accuracy count and report them as their own number. (the paper)
to try — Store one roleset per (lemma, sense) as a BTreeMap in the pupil dictionary, with the sense's roles and their descriptions. Measure how many verbs in the school lessons get a roleset, and how many of their sentences have every role filled. (the paper)
to try — A held-out gain. Hold one translation of a verse out, and compare its edit distance to the meta-world paths with its edit distance to the other translations as a plain set (the paper's ed gain, mean 0.91 over 899 groups). It is a measure of the lexicon that uses no test set. (the paper)
to try — A lexical feature resolutive in the dictionary : know, tell, discover, forget, guess accept a multiple-wh remnant as an object; ask, wonder, investigate, discuss do not. Keep the verb list in a BTreeMap and test the contrast in (2b) against (21b) of the paper, with a negative control on the ask/wonder class. (the paper)
to try — en/seeds/lexicon: count the distinct valid UD patterns per word, as the paper does for "cat" and "time", and list the words whose pattern count is largest, to find the seeds that are too broad. (the paper)
to try — en/src/parse.rs: build a per-word table of allowed head-dependent label pairs from (link-style), and use it to forbid impossible arcs during the arc-standard search; measure LAS on EWT test (must stay the same) and the candidate transitions per sentence. (the paper)
to try — Profile grammar rules, not only the parser: for each rule or lexicon class in en, log how many candidate edges it creates and how many survive into the final tree on a fixed test suite, and report the most wasteful ones. Two runs must give identical profiles. (to check: which profiling measures the paper uses) (the paper)
to try — Keep a fixed list of human-judged merges as a held-out check. A new lexicon rule must raise the correlation on that list, measured twice, before it is accepted. The check should be kept apart from the rules that were fitted to it. (the paper)
to try — Keep the sense lexicon as an explicit table with a count of covered senses per entry. The paper finds that a complete sense lexicon moves the score more than any single rule. In this suggests counting how many merges are between senses, not words, and reporting that count with each run. (the paper)
to try — Dynamic versus static verbs from : a lexicon flag, checked against the paper's split (dynamic conjuncts vs. "the food is super" type stative ones). (the paper)
to try — Re-ranking precision for choosing among the translator's variants (author deltas, lexicon versions): for each verse, a choice is correct if it is among the variants with the lowest residual. Report it per verse set. (the paper)
to try — : rank candidate modern forms for archaic English words that are not in the dictionary, by a weighted edit distance with the paper's limits (length change from +1 to -3, distance at most 4). Measure the exact-match rate on held-out tools/gutenberg lessons against the archaic rules alone. (the paper)
to try — : split unknown words into two dictionary parts, as the compound splitter does, and report how many unknown historical words are compounds. Use the count to decide whether the step is worth building. (the paper)
to try — Bound candidate pairs by length difference. The paper shows that most true normalizations differ in length by only a little. For the lexicon, a length-ratio bound on candidate merges would cut noise merges cheaply. (the paper)
to try — Split compounds before a merge, so phrase concepts such as *unto_thou* are broken into parts that keep their function words. This targets open problem 2 directly: merges that drop function words are the noise the lexicon needs to avoid. (the paper)
to try — Use the dev set as a regression gate for the lexicon. A merge or delta is accepted only if the dev score does not fall, and the log records which failing token each change was for. The translator's residual-driven lexicon (open problem 2) can be checked the same way. (the paper)
to try — Add priority tiers to the lexicon, so that early-modern entries shadow modern senses (the *putt* case). Measure lemma accuracy and without the tier, and run twice. (the paper)
to try — Report precision, recall and F in the same way for the early-modern transcription. Add a negative control: with the transcription lexicon shuffled, the score must fall well below the real run. (the paper)
to try — Track separation as a rule for claims: a result that uses a resource outside the closed set (WordNet, a gazetteer) is reported under an "open" label, so that no open-track gain is taken for a closed-track one. (the paper)
to try — A larger out-of-domain yardstick for the frame reader: PropBank's annotation of the English Web Treebank and of the Brown fiction subset (closest to tales), mapped through function tags (PAG → agent, PPT → patient, GOL → recipient or place, LOC → place, TMP → when:, CAU/PRP → cause). It runs outside the repo as a test, like the PropBank tables already converted in prag lexicon convert, and gives per-role recall on thousands of predicates where the frame bench has 240 lines. Mova's parser is trained on EWT, so use EWT's test split only. (the paper)
to try — Adjective and noun predicates with arguments, as PropBank's unified rolesets treat them: «the hare was afraid of the fox», «proud of his speed», «fond of honey» become frames of the related verb (fear, pride, like) with the subject as experiencer and the *of*-phrase as patient. Build the link from WordNet derivations (WordNet already feeds), not from PropBank. Measure with world knowledge diagnose on questions «What was the hare afraid of?» and course test counts. (the paper)
to try — Build verb-type lists for Mova from our own lexicon: which verbs take a sentential complement, which copula adjectives apply only to propositions ("true", "wrong", "right"). The paper says such lists are the weak part; use them as data in en, with a source note for each entry. (the paper)
to try — Keep an AltLex list in with counts per expression. The paper says about 15% of AltLex expressions meet the criteria for explicit connectives, so promote an expression once its count passes a threshold. (the paper)
to try — Scenario flag: mark each telic or atelic verb phrase in the lexicon with a closed feature "can be abandoned or interrupted" (for example "eat a strawberry", "walk up a mountain"). The entailment gate in prag then answers "unclear" rather than "no". Test on the 50 phrases listed in the paper. (the paper)
to try — A support-verb rule in the lexicon : a verb that takes an NP and an XP where the XP is an argument of the NP's head ("take a walk", "give a standing ovation") gets the noun as the predicate and the verb as support, with the paper's exception for idioms. Test on the paper's examples, with "make a mistake" as the case where the subject selection is inherited from the noun. (the paper)
to try — Test whether a merge is justified by its context alone: compare the replacement table and without the neighbour context, and drop merges that the context does not support. The paper's WordNet result suggests lexical choice is where the errors lie. (the paper)
to try — Keep the lexicon linked by concept across languages, one entry per concept carrying all its forms. The meta-world is already a concept table, so the test is whether the forms of one concept in a new target language can be added without touching the reader. (the paper)
to try — The formality classifier can filter candidate outputs. The count form is the register lexicon score, a sum of per-word values over the chunk. (the paper)
to try — Unit compatibility by the paper's three rules (string match, lexical relation from the dictionary, and a person-noun rule). Log which rule fired, so a wrong comparison names its cause. (the paper)
to try — and : use the same index to interpret a new compound. For "turtle boat" the rule "boat used to catch marine animals" gives a paraphrase; record the paraphrase and the rule id, so the reader can say why. (the paper)
to try — A per-author concordance report from : for each concept, the glosses per author with counts and lower bounds. This is the evidence the author delta needs (fix at least three more tokens, hold in half the context), in a form a human can read. (the paper)
to try — Use verse-level co-occurrence as a baseline for the Hebrew reader: for each Hebrew lemma, the count of verses in which a gloss occurs, compared with the IBM Model 1 alignment in on the same Genesis verses. (the paper)
to try — Test a monotonic constraint in the Hebrew word alignment, the counterpart of the monotonic attention used here. Measure the alignment error on the MACULA glosses before and after. (the paper)
to try — An animacy-based choice rule for the English genitive or the generator: a possessor with the animate feature takes the s-genitive, an inanimate one takes, and the choice is logged with the feature that decided it. Keep the animacy values in the lexicon as an enum, not as strings. (the paper)
to try — : write the role rules over UD trees, with the case marker (case relation) as the first cue and the lemma of the head noun's WordNet hypernym as the second, which overrides the first. Keep the rule table in BTreeMap order and log which rule chose each label, so each choice can be explained. (the paper)
to try — : a completion table in code, for example "PLU+ implies not inflectable", "ER-comparative implies not inflectable", written as a list of rules with a reason for each, and applied to the lexicon at build time. (the paper)
to try — Store irregular forms in the lexicon rather than in rules, as the paper recommends, and measure the size of each table. Report both the table size and the lemma accuracy on a held-out list, to show the cost. (the paper)
to try — The spelling rules as one transducer per rule: the epenthesis rule (insert e before s after sibilants, as in "boxes") can be a small deterministic table, tested over all lemmas in the dictionary, with a count of changed forms as the measure. (the paper)
to try — Check functionality of the mapping: in the learned lexicon, count the many-to-one and one-to-many cases per concept. The paper's problem with a mapping that is not a function appears there as a count, and a gate can reject merges that break it. (the paper)
to try — Keep a spelling-pattern table (vowel swaps and unaccented endings) as data, updated only from errors seen in a checked dictionary, so each pattern has a source. (the paper)
to try — Plot quality against data size for the translator's concept lexicon. Apertium's pilot pairs suggest that a small curated table already gives usable output on close pairs. In tools/translate, record the exact-match rate per author as merges are added, and find where the next merge stops paying. (the paper)
to try — Build a small compound-family index from the lexicon (constituent, count of members) and report how the family size of a head varies, so the school dictionary knows where a compound's meaning is predictable and where it must be listed. (the paper)
to try — Control for the known noise: phrases that carry function words (the open case of unto_thou) should be dropped by a gate. Run the lexicon learner and without this gate, twice each, and compare the residual. (the paper)
to try — The interpreter and translator split predicts where a plain lexicon is enough. Report the translator's lexical residual separately for "business" text, the school's fact and explain lines with concrete objects, and for the Bibles. If Schleiermacher is right, the meta-world lexicon should come much closer to zero on the first. (the paper)
to try — : a layered normaliser in the same spirit as VARD2. First a closed dictionary of Early Modern forms, then letter-replacement rules with fixed weights chosen by counts on a training split, then a dictionary check. Report precision and recall per century, run twice. (the paper)
to try — Report economy next to accuracy: the number of meta concepts per sentence and the lexicon size after each round. Keep a lexicon merge only when accuracy and economy both do not get worse. (the paper)

This page in Markdown → https://md.dictionary.elamur.ai/free/en/ideas/page/6/