[Elamur Dictionary](https://md.dictionary.elamur.ai/free/en/) · [this page in HTML](https://dictionary.elamur.ai/free/en/ideas/page/5/)

[dictionaries](https://md.dictionary.elamur.ai/free/en/) / ideas

# 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** — An evaluator for Mova's own output: compare a generated Hebrew sentence with a reference at lemma level, using the dictionary hebrew lemma column, and count matches per token. This is a deterministic metric for paraphrase tests. ([the paper](https://research.elamur.ai/free/papers/langstream-2014-el-shishtawy-lemma-based-evaluator-semitic-language/))
- **to try** — Their evaluation design, generate a lexicon and test it on a random sample against an independent analyser, with the sample size stated (44%), can be copied for verb paradigms against MACULA. ([the paper](https://research.elamur.ai/free/papers/langstream-2014-martinez-jabalin-comprehensive-computational-model-modern/))
- **to try** — A template list for Hebrew verbs and nouns and, checked by the rule of the paper: a candidate template is kept only when its pairs agree in a second, independent measure (for Mova, the lemma pairs of dictionary hebrew, not embeddings). ([the paper](https://research.elamur.ai/free/papers/langstream-2017-sakakini-fixing-infix-unsupervised-discovery-root/))
- **to try** — A coverage measure for the Hebrew lexicon: the share of running-text word forms that get at least one candidate, reported per book, so that gaps show where the lexicon is thin. ([the paper](https://research.elamur.ai/free/papers/langstream-2019-neme-restoring-arabic-vowels-through-omission/))
- **to try** — Pattern-first inflection: keep a lemma-level lexicon with its pattern class and generate forms from pattern tables, instead of storing every form. Test it as a coverage check against the forms in dictionary hebrew (lemmas only, no roots yet). ([the paper](https://research.elamur.ai/free/papers/langstream-2019-neme-restoring-arabic-vowels-through-omission/))
- **to try** — A gap-filler for damaged Hebrew and Aramaic text in : remove a word, rank candidates from the lexicon and the grammar, and report MRR and hit@k on held-out verses, using the same cloze protocol as the paper. ([the paper](https://research.elamur.ai/free/papers/langstream-2020-fetaya-restoration-fragmentary-babylonian-texts-using/))
- **to try** — Build a Hebrew emotion lexicon test from the paper's category list (anger, fear, joy, surprise and the others), using the labelled comments if their licence allows, and otherwise a small set Mova labels itself. ([the paper](https://research.elamur.ai/free/papers/langstream-2021-chriqui-hebert-hebemo-hebrew-bert-model/))
- **to try** — Measure transcript noise before any use: sample utterances, hand-check them, and report a word error rate. The same deterministic correction approach as (rules over a lexicon) could then be tested on the transcripts, with its gain measured. ([the paper](https://research.elamur.ai/free/papers/langstream-2023-marmor-ivrit-ai-comprehensive-dataset-hebrew/))
- **to try** — A dialect and register test set for : colloquial Modern Hebrew clitic forms that the Hebrew segmenter cannot split, kept as a fixed list of cases, as the paper's Egyptian example shows. ([the paper](https://research.elamur.ai/free/papers/langstream-2024-alrefaie-exploring-tokenization-strategies-vocabulary-sizes/))
- **to try** — A register cloze test for : hide one word per sentence, ask a candidate model which word fits, and report top-1, top-2 and top-5 agreement. This measures how well the register's lexicon is covered, and it is a check. ([the paper](https://research.elamur.ai/free/papers/langstream-2024-goldin-knesset-dictabert-hebrew-language-model/))
- **to try** — Record the bias caveat in the data card: any lexicon or frequency list taken from parliamentary text should carry a label saying it reflects that register and its politics. ([the paper](https://research.elamur.ai/free/papers/langstream-2024-goldin-knesset-dictabert-hebrew-language-model/))
- **to try** — Do not commit early to one reading. The method drops the diacritisation step that picks a single reading before synthesis. For in unpointed mode, carry the candidate splits of an unpointed word (from the prefix and lexicon tables) as a small lattice to the parser, instead of one split, and measure against a gold set whether this lowers errors. ([the paper](https://research.elamur.ai/free/papers/langstream-2024-zeldes-enhancing-tts-stability-hebrew-using/))
- **to try** — Measure the compression of the Hebrew segmenter on a corpus: tokens per word, before and after a change of the prefix or lexicon tables. The paper uses this as a cheap check before any downstream test. For, it is a fast regression test for the segmenter. ([the paper](https://research.elamur.ai/free/papers/langstream-2025-hashiloni-building-patient-journeys-hebrew-language/))
- **to try** — A deterministic finite-state transducer for the grapheme-to-sound step, with a dictionary for exceptions, that handles letters read out of linear order (their example: a vowel letter read after the consonant it follows). Write it as a rule table and test it on a word list with gold readings. This fits Mova's rules-first design. ([the paper](https://research.elamur.ai/free/papers/langstream-2025-kolani-phonikud-overcoming-phonetic-underspecification-hebrew/))
- **to try** — Mark the end of a cliticised prefix (the one-letter particles ו, ה, ב, ל, כ, מ, ש and their like) so that a dictionary lookup can match irregular words after the prefix is removed. For (the prefix tables) and, test how many irregular forms match and without the boundary mark. ([the paper](https://research.elamur.ai/free/papers/langstream-2025-kolani-phonikud-overcoming-phonetic-underspecification-hebrew/))
- **to try** — An order-sensitivity band test for any Mova matcher that compares two forms (for example pointed and unpointed lookup in the lexicon). Use five bands: a true variant, an adjacent transposition, a full permutation, a reversal, and an unrelated form. A good matcher keeps the variant and the transposition, and drops the permutations and the reversal. Score the bands, not single examples. ([the paper](https://research.elamur.ai/free/papers/langstream-2026-gadd-symphonym-universal-phonetic-embeddings-cross/))
- **to try** — A root check in the lexicon: for a Semitic pair, compare the consonantal roots and flag pairs that share the root but differ in meaning (the paper's example: the root l-H-m gives Hebrew \*lehem\*, "bread", and Arabic \*laHm\*, "meat"). Such a rule lists false-friend candidates for a human to review. ([the paper](https://research.elamur.ai/free/papers/langstream-2026-liang-similar-means-different-evaluating-llms/))
- **to try** — Use the parliamentary transcripts as a formal-register check for Modern Hebrew. They are non-verbatim, so they are good for register and lexicon statistics, but not for parsing gold. ([the paper](https://research.elamur.ai/free/papers/langstream-2026-marmor-voxknesset-large-scale-longitudinal-hebrew/))
- **to try** — Lemma convention: the lemma is the fully pointed citation form of the singular, and a derivational difference makes a new lexeme. Use this rule when dictionary hebrew chooses lemmas, so that derived nouns are not merged into their base. ([the paper](https://research.elamur.ai/free/papers/langstream-2026-neme-pattern-root-inflectional-morphology-arabic/))
- **to try** — One plural per lexeme: when a noun has two plurals, make two entries, not one entry with two plurals. Check it on the Hebrew noun lists before the lexicon grows. ([the paper](https://research.elamur.ai/free/papers/langstream-2026-neme-pattern-root-inflectional-morphology-arabic/))
- **to try** — Define an out-of-vocabulary set by the system under test and print the definition next to each number. The paper's OOV set is defined by its own expanded tokenizer, so it is not "unseen words". Any "unknown" set in tests should be defined the same way, from the lexicon used. ([the paper](https://research.elamur.ai/free/papers/langstream-2026-teklehaymanot-expanding-lexicon-ge-ez-based/))
- **to try** — Separate the new material from the extra training: the paper cannot tell whether gains come from adapted new embeddings or from more training, and says so. When Mova adds entries to a lexicon, include a control that receives the same extra text but not the new entries. ([the paper](https://research.elamur.ai/free/papers/langstream-2026-teklehaymanot-expanding-lexicon-ge-ez-based/))
- **to try** — Collocations as counted lexicon entries. Multi-word concepts in the meta-world, such as the translator's, are collocations in this sense. Keep each as an entry with its count and its standard form, and report how many entries are used per author. ([the paper](https://research.elamur.ai/free/papers/lareau-2018-gendr-generic-deep-realizer-complex/))
- **to try** — Error-driven filtering of merges: keep a concept merge in the lexicon only if it lowers the reader and encoder bits on the training verses, and drop it otherwise. This is the paper's filter applied to open problem 2, with the existing rule of three or more tokens fixed as a second check. ([the paper](https://research.elamur.ai/free/papers/lavoie-2002-learning-domain-specific-transfer-rules/))
- **to try** — Controller rule in the lexicon: store for each control verb a flag, subject-to-subject or object-to-subject (the valency types of the Osborne 2017 account), in as an enum. Resolve the reflexive in a non-finite clause to that controller. Measure accuracy on minimal pairs generated from the paper's design, and use a negative control with the flags swapped, which must drop to zero. ([the paper](https://research.elamur.ai/free/papers/lee-2022-can-language-models-capture-syntactic/))
- **to try** — Keep the lexical fact in the lexicon, not in a learned model: the paper's finding that promise and persuade differ in meaning, not in surface form, is an argument for explicit lexical flags over a statistical learner. ([the paper](https://research.elamur.ai/free/papers/lee-2022-can-language-models-capture-syntactic/))
- **to try** — Keep the lexical roles apart from the concept. The lexicon entry says which argument goes where, and the meta-world concept comes from the ontology. In, store the per-author role names in a table keyed by concept, learned from the parses. ([the paper](https://research.elamur.ai/free/papers/levin-1994-correct-place-lexical-semantics-interlingual/))
- **to try** — Plot the frequency curve of meta concepts. If a small head of concepts covers most tokens, the lexicon is compact; the long tail is where the lexicon merges are most likely to be noise, so check those merges first. ([the paper](https://research.elamur.ai/free/papers/levin-2002-balancing-expressiveness-simplicity-interlingua-task/))
- **to try** — A verb class with two frames: mark precipitation and emission verbs with an event class in the lexicon, and let the class choose the subject role. Rule: a source subject with an object is an activity; a substance subject with a directional phrase is a scalar change. ([the paper](https://research.elamur.ai/free/papers/levin-2019-talking-about-weather-two-construals/))
- **to try** — Build a semantic class lexicon for mentions with the paper's unified ten types, from licence-clean sources (WordNet's top categories, Mova's own gendered-noun and animacy lists), and use GUM's entity types on dev only to measure its accuracy. The coref README lists semantic classes as needed for «it» and; the paper's table gives the target inventory and the GUM mapping for scoring. ([the paper](https://research.elamur.ai/free/papers/levine-2024-unifying-scope-bridging-anaphora-types/))
- **to try** — onto: measure the coverage of the open English WordNet, the 120,564 synsets in onto, on EWT test content words. Report the share of lemmas with at least one synset, and the share with none, so the gap is visible next to the 1994 growth figures. ([the paper](https://research.elamur.ai/free/papers/lex-1994-miller-wordnet/))
- **to try** — or the lexicon build: compute logDice for candidate multiword pairs (adjacent and within a 5-token window) from an open corpus, and rank them. Measure precision at the top 100 against the expression base in prag, on a held-out split of it. ([the paper](https://research.elamur.ai/free/papers/lex-2008-rychly-logdice/))
- **to try** — : apply a feature list to the English relations that the lexicon already holds, and report the precision and coverage of each feature on a hand-checked sample, following the paper's per-feature table. Keep the features that are precise on the sample. ([the paper](https://research.elamur.ai/free/papers/lex-2012-meyer-sense-disambiguated-wiktionary/))
- **to try** — Lexicon collection probe: take a list of English nouns from raw text, rank them by how well they fit the induced paradigms, and report the precision at 100 and 1000 words, as the paper does for Swedish. The precision check goes into en lexicon. ([the paper](https://research.elamur.ai/free/papers/lex-2014-ahlberg-paradigm-extraction/))
- **to try** — : add an extra gloss source to the OEWN senses only where the source is licence-clean, and measure the WSD change for two Lesk variants on EWT sense-tagged tokens. The paper's result, a gain for one variant and a loss for the other, is the negative control. ([the paper](https://research.elamur.ai/free/papers/lex-2014-miller-wordnet-wikipedia-wiktionary/))
- **to try** — Keep forms unlemmatised in the lexicon checks: run en morph-eval on forms that the lemmatiser merges (computer and computers, alternative spellings), and report how many of them would be lost if they were merged. This is a test, not a change to it. ([the paper](https://research.elamur.ai/free/papers/lex-2016-grieve-lexical-emergence-american-english/))
- **to try** — Layout consistency check: group English lemma tables by layout and count how many layouts give more than one feature reading. Use this as an offline check on the lexicon build, not as a training source. ([the paper](https://research.elamur.ai/free/papers/lex-2016-kirov-wiktionary-paradigms/))
- **to try** — : write the English inflection paradigms as a small set of rules, count how many paradigms each rule set produces, and report the number that have only one example, following the paper's Table 2 layout. Use it as a diagnostic for the lexicon, not as a training source. ([the paper](https://research.elamur.ai/free/papers/lex-2017-liu-fst-paradigm-extraction-wiktionary/))
- **to try** — Out-of-lexicon test: hold out 10% of EWT test forms that are missing, then measure lemmatization of those forms with the rule-based path against the current dictionary path. The paper's split is 90% train and 10% held-out, and this is the English analogue. ([the paper](https://research.elamur.ai/free/papers/lex-2017-liu-fst-paradigm-extraction-wiktionary/))
- **to try** — Productivity as a rule feature: compute the paper's percentage-of-more-frequent-family value for each suffix from the en lexicon counts, and test whether it separates the suffix rules that en morph-eval gets wrong. Measure lemma accuracy on EWT and without the feature. ([the paper](https://research.elamur.ai/free/papers/lex-2018-sanchezgutierrez-morpholex/))
- **to try** — Gloss provenance: count the glosses in onto that come from Wikipedia-style text, as the paper flags them as a licence risk. Keep only glosses with a clear source, and measure the share removed. ([the paper](https://research.elamur.ai/free/papers/lex-2019-mccrae-english-wordnet/))
- **to try** — onto: build the paper's example check, an edit-distance test that flags synset examples too close to the gloss or to other examples of the same synset. Measure how many OEWN synsets it flags out of the 120,564, and hand-check 100 of them before any change. ([the paper](https://research.elamur.ai/free/papers/lex-2019-mccrae-english-wordnet/))
- **to try** — : handle a target that is not in WordNet (an adjective-noun phrase, for example) by taking the sense of its head noun and adding the modifier's synonym set as a restriction. Measure the accuracy on adjective-noun phrases in EWT that have no WordNet entry, against the head-only baseline. ([the paper](https://research.elamur.ai/free/papers/lex-2020-bevilacqua-generationary/))
- **to try** — Multi-inventory candidates: put the WordNet, FrameNet and Wikidata glosses of a word into one candidate pool in onto, and keep the source of each chosen gloss in the explanation. Measure F1 on the Raganato sets (test only), per inventory and for the pool, and check that two runs are identical. ([the paper](https://research.elamur.ai/free/papers/lex-2020-bevilacqua-generationary/))
- **to try** — Offline judge: annot produces alternative glosses for a sample of targets, and a Rust check keeps only those that contain the target lemma's WordNet synonym or hypernym. Use this as silver data only; the paper's 4.3-point zero-shot gain is a neural result and is not a target for Mova. ([the paper](https://research.elamur.ai/free/papers/lex-2020-bevilacqua-generationary/))
- **to try** — Gap report: run the en lemmatiser over the school reading texts and list the forms it cannot analyse. Measure the share that a later dictionary build would cover, as a check on the lexicon, before any neural step. ([the paper](https://research.elamur.ai/free/papers/lex-2020-ryskina-neologisms-distributional/))
- **to try** — Growth-rate feature for : for a set of words, compute the per-book frequency trend against publication date across the public-domain books, and check the sign against the known history of a few words (for example "hast" and "doth" should fall). Use this as a test of the trend function, not as a new lexicon rule. ([the paper](https://research.elamur.ai/free/papers/lex-2020-ryskina-neologisms-distributional/))
- **to try** — and : for every token in an unseen modern text that is not in the en lexicon (110,308 lemmas), count the number of known lemmas within a small co-occurrence distance, as a sparsity score. Measure whether the score separates the unknown words that are later accepted from those that are not, on a hand-labelled list. ([the paper](https://research.elamur.ai/free/papers/lex-2020-ryskina-neologisms-distributional/))
- **to try** — : a lexicon-driven clipping link (mathematics to math, telephone to phone) with a list of English clipping pairs written from our own lists. Measure precision and recall on a held-out list, with a negative control of random word pairs. ([the paper](https://research.elamur.ai/free/papers/lex-2020-wu-wiktionary-normalization/))
- **to try** — : flag lexicon entries with a missing earlier part where a numbered component is present, following the paper's typo check. Hand-check the flagged entries and report how many are real typos. ([the paper](https://research.elamur.ai/free/papers/lex-2020-wu-wiktionary-normalization/))
- **to try** — Compound-first measure: check the compound splits first, since the paper names compounds as the hardest category across languages. Measure the share of English compounds in the lexicon with a correct two-part split. ([the paper](https://research.elamur.ai/free/papers/lex-2022-batsuren-morpheme-segmentation/))
- **to try** — and : replace the masked-model substitutes with the English synonyms and hypernyms from onto (Open English WordNet). For each token, take the candidate set, and cluster the occurrences of one lemma by the overlap of their candidate sets with a fixed rule. Measure F1 on 100 hand-annotated sentences per word for 20 ambiguous words. ([the paper](https://research.elamur.ai/free/papers/lex-2022-eyal-substitution-wsi/))
- **to try** — : label each sense cluster of a word with the WordNet gloss that best matches its usages, using a bag of co-occurring content words, and keep the gloss choice as an explanation. Measure the share of acceptable labels on a hand-judged sample of clusters; the paper's 80% is the bar to beat. ([the paper](https://research.elamur.ai/free/papers/lex-2023-giulianelli-definition-generation-senses/))
- **to try** — Diachronic check in : take a word that changes sense between an early and a late book (for example "wit" or "prevent") and report the usage counts per period per WordNet sense. This makes the change readable without a neural model. Use only public-domain texts. ([the paper](https://research.elamur.ai/free/papers/lex-2023-giulianelli-definition-generation-senses/))
- **to try** — : the extended Lesk step compares a sense's gloss with the context. Add the WordNet examples to the gloss bag, and measure WSD accuracy on a held-out English WSD test set and without the examples. This tests whether WordNet examples carry signal for the deterministic system. ([the paper](https://research.elamur.ai/free/papers/lex-2024-almeman-wordnet-examples-llm/))
- **to try** — onto or annot: for synsets with unclear glosses or no examples, generate candidate examples offline with a large model, then keep only those that pass Rust GDEX-style penalties (common-word share, pronoun share, length window, ambiguity). Measure the share of kept examples on a 87-word style sample rated by hand. ([the paper](https://research.elamur.ai/free/papers/lex-2024-almeman-wordnet-examples-llm/))
- **to try** — Offline gate for silver glosses: have annot produce free-form explanations for a sample of target words, then let a Rust check accept an explanation only if it names the WordNet sense's main lexeme or synonym set. Measure the agreement with the expert's good/bad/borderline tags on a small hand-tagged sample. ([the paper](https://research.elamur.ai/free/papers/lex-2025-meconi-llms-word-senses/))
- **to try** — : a deterministic check for binomial names (a capitalised genus followed by a lowercase species word) as a single token group for the lexicon, and a count of how often it fires on GUM test text (test only). ([the paper](https://research.elamur.ai/free/papers/lex-2026-mccrae-open-english-namenet/))
- **to try** — Coarse check in onto: score the deterministic WSD at WordNet lexfile (supersense) level next to the fine sense level on the same test items. The paper reports coarsening lifts agreement and accuracy together; check whether the same holds for the deterministic system. ([the paper](https://research.elamur.ai/free/papers/lex-2026-philippov-english-wsd-labels/))
- **to try** — and : write the particle verbs as macros, one per type (V+P, V+NP+P, V+P+NP) with the constraint "no NP after the particle" as a rule, not as a list. The macro gives the lexicon one line per verb, and the rule is explainable. ([the paper](https://research.elamur.ai/free/papers/li-2003-expert-lexicon-approach-identifying-english/))
- **to try** — Duplicate check before coining: the Academy searches all earlier glossaries so as not to create needless synonyms. Build the same check for new lemmas against the lexicon and the field lists, and flag same-meaning or same-form entries as a rule-based test. ([the paper](https://research.elamur.ai/free/papers/lingbulk-1982-yannai-innovation-neologism-modern-hebrew/))
- **to try** — A selection field on each Hebrew verb entry: the kinds of direct or prepositional complement it takes (human, animate, the object marker את), so that a generator can refuse suffixed forms that break the restriction. Start with a short list of verbs and count the rejections. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2003-dichy-roots-patterns-vs-stems-plus/))
- **to try** — Count the attested forms before building a generator: for dictionary hebrew, generate full forms from the lemmas with the paradigm rules and measure the share that occurs in the Torah. The paper's ratio (72 million against 6.2 million) is the reason to measure. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2003-dichy-roots-patterns-vs-stems-plus/))
- **to try** — Roots as labels only: keep the root in dictionary hebrew for lookup and for grouping derived forms, but build word-forms from the stem, as the paper's footnote on roots suggests. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2003-dichy-roots-patterns-vs-stems-plus/))
- **to try** — Exceptions in the lexicon, not the rules: store only the idiosyncratic plural endings (the default is im for masculine and wt for feminine) and the secondary verb bases that are hard to generate. Check the Biblical noun and verb chapters against this split. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2003-wintner-resources-processing-israeli-hebrew/))
- **to try** — Derivation cross-check: generate deverbal nominals from every verb and match them with the nominals in the lexicon; unmatched pairs show gaps and possible errors. This is a test for. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2006-itai-computational-lexicon-contemporary-hebrew/))
- **to try** — : lemmatization as a class of edit rules (strip a suffix, add a suffix), learned and, used only for forms that the dictionary does not contain. Measure lemma accuracy on the unseen forms of EWT test against the current rule set. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2008-chrupala-learning-morphology-morfette/))
- **to try** — Report the lemma accuracy on unseen words separately, as the paper does, since the English dictionary hides this error on known words. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2008-chrupala-learning-morphology-morfette/))
- **to try** — A difficult-by-design test subset: select EWT sentences that contain a verb form not seen in en training, using en lexicon for the check, and measure en parse LAS on this subset against the full EWT test. The paper's selection rule is the model here. Test only, since the subset is used for measurement. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2008-foster-parser-evaluation-bnc-evaluating-4/))
- **to try** — Run an ablation on gparse: and without morphological features, and with rare lexical items replaced by an unknown token at thresholds 20 and 100. The paper finds that a second-order graph model with pruned lexicon can beat the full one. Check this on Mova's data before adopting it, and keep the full run as the control. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2009-goldberg-hebrew-dependency-parsing-initial-results/))
- **to try** — Classify every disagreement before fixing anything: label each one as gold error, rule error or lexicon error, as the paper does (it finds only about 18% of precision errors are real). Print that split for every macula compare run against UD PTNK. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2010-altantawy-morphological-analysis-generation-arabic-nouns/))
- **to try** — The lemma convention: the lemma is the singular citation form of one lexeme, and a derivational difference makes a new lexeme. Use it when dictionary hebrew picks lemmas, so that derived nouns are not merged into their base. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2010-altantawy-morphological-analysis-generation-arabic-nouns/))
- **to try** — Their verb rule is a finite-state pattern: a template with fixed vowels and an infix, applied to a consonantal root from the lexicon (their hitpael example, with the root g.b.r). Mova can check its Modern Hebrew verb paradigms against such a rule set, with the root lexicon as the only data. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2010-nir-morphologically-analyzed-childes-corpus-hebrew/))
- **to try** — Build the Hebrew affix lists from MACULA as the analyzer's first step: count prefix and suffix strings with their tags and keep the table as the pre and lex tables of the segmenter. Report the lexicon coverage on a held-out part of the text, as the paper does (about 85%). ([the paper](https://research.elamur.ai/free/papers/lingbulk-2010-sawalha-fine-grain-morphological-analyzer-part/))
- **to try** — Fill gaps in a paradigm by analogy: when a lemma lacks some inflected stems, copy them from a lemma of the same class that has them (the paper's planned step for hollow verbs). For Hebrew, apply it to the weak-verb classes (I-nun, III-aleph, III-he, geminate) in the lexicon, and measure on held-out forms that were removed on purpose. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2012-habash-morphological-analyzer-egyptian-arabic/))
- **to try** — A hybrid lemmatizer: a lookup table of stem-to-baseform pairs seen in the dictionary or in dictionary hebrew first, a rule-based or statistical guesser only for unknown stems. The paper shows the table is near-perfect on seen stems and the guesser is needed for the rest. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2014-black-evaluating-lemmatization-models-machine-assisted/))
- **to try** — Report lemmatizer accuracy in the same six error classes (known stable unambiguous, known latent ambiguous, known stable ambiguous, known dynamic ambiguous, unknown unambiguous, unknown ambiguous). For and dictionary hebrew, this shows whether a gain comes from known stems or from unknown ones, which a single overall figure hides. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2014-black-evaluating-lemmatization-models-machine-assisted/))
- **to try** — Take the 19% interlink coverage as a limit: a cross-lingual lexicon transfer through English links reaches only a small part of Hebrew. Modern Hebrew concept lists need Hebrew sources of their own. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2014-chrizman-how-construct-multi-lingual-domain/))
- **to try** — A small, readable classifier for gender and number of unseen nouns: features are the pattern, the stem length, the POS, the suffix (feminine -ah, -ot, dual and plural endings), and a dictionary tag. Test with a split by lemma, so no lemma is seen in both parts. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2014-darwish-stem-templates-improve-arabic-pos/))
- **to try** — A maximum-match segmenter over the lex and pre tables as a second deterministic baseline for unpointed text. The paper's finding that dictionary matching beats learners on a small corpus argues for the tables before any learned step. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2016-homburg-word-segmentation-akkadian-cuneiform/))
- **to try** — Build the paper's glossed examples into a small gold test set for the verb tables, with the glosses as given, and count how many forms the rules get right. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2016-kalin-morphosyntax-aspect-stacking-northeastern-neo/))
- **to try** — Keep root, transfix pattern and word-form as separate tables in the Torah dictionary. dictionary hebrew has lemmas and Strong's numbers but no roots yet; a root column would make the lemma-to-root lookup deterministic. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2016-klimek-creating-linked-data-morphological-language/))
- **to try** — The cleaning rules (drop rows with invalid syntax, several entries in one column, or no word class) are a validator for any lexicon import Mova makes, such as the Wiktionary-derived tables. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2016-klimek-creating-linked-data-morphological-language/))
- **to try** — Use the lexicon as a soft feature in the gparse input: for each token, the set of values per FEATS category that the lexicon allows (dictionary hebrew), added as a feature and never as a filter. A word with no entry gets an empty set, so the right analysis can still win. Where:, described. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2017-inoue-joint-prediction-morphosyntactic-categories-fine/))
- **to try** — Use the per-category table as an error template for : which features are easy (clitics, prefixes) and which are hard (case, mood), and the share of errors that come from a missing candidate in the lexicon, which the paper reports as a large part of its baseline errors. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2017-inoue-joint-prediction-morphosyntactic-categories-fine/))
- **to try** — A root feature in the lexicon for hitXaYeZ and niXYaZ verbs: Self-Oriented or Other-Oriented. The reflexive reading then becomes a lookup on the root, not a rule on the template. Roots not yet labelled are the list to check first. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2017-kastner-reflexive-verbs-hebrew-deep-unaccusativity/))
- **to try** — Store the plural exponents as an explicit table keyed by (gender, State, possessed or not), not as string rules, so that a lexicon or a segmenter reads the same table: masculine Free -im, masculine Construct -ej, feminine -ot in both States, feminine Construct -at. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2018-faust-exponence-allomorphy-haplology-number-state/))
- **to try** — A rule for the CSN head: a noun in its Construct form (for example -at, -ey, or the bare construct) followed by a complement with ha- is a CSN head, and a ha- before that head is a violation to flag. Lexicalised compounds (beged-yam, 'swimsuit') are listed in the lexicon so that they are not treated as CSNs; the paper notes that native speakers do this. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2018-friedmann-head-construct-construct-state-nominals/))
- **to try** — Use a second-order (distributional) similarity only as a test of the lexicon, never as part of the parser, since Mova is rule-based. The paper's lesson for that test is that the comparison is only meaningful with the same term representation on both sides. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2018-liebeskind-automatic-thesaurus-construction-modern-hebrew/))
- **to try** — A cost channel for segmentation in : each prefix or suffix cut carries a penalty, so an unsplit word wins unless the stem is in the lexicon. This stops over-segmentation of names, the same way the paper stops ʔuɡanɨda from being read as a possessive -a. Where: the ranking of the lex and pre splits. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2018-littell-parser-combinators-tigrinya-oromo-morphology/))
- **to try** — Templates as explicit patterns for Hebrew and Aramaic root-and-pattern plurals: a consonant-class pattern over the root triple, applied after the lexicon lookup, as the paper does for Tigrinya internal plurals. Where:,, implemented as a fallback. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2018-littell-parser-combinators-tigrinya-oromo-morphology/))
- **to try** — For unpointed Hebrew, in unpointed mode: when a consonantal form has several analyses, prefer the one whose pointed form is most frequent in a pointed corpus (MACULA gives the pointed forms). It is table-based and needs no context, as the paper does. Where:. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2018-mubarak-build-fast-accurate-lemmatization-arabic/))
- **to try** — Order lemmas by usage in the lexicon, not in the code: where two analyses share a form (the paper's car/walker case), the lemma with more corpus occurrences comes first, and the order is stored in dictionary hebrew. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2018-mubarak-build-fast-accurate-lemmatization-arabic/))
- **to try** — Mark idiosyncratic verbs with a lexical flag (the paper's l-syntax) and let the regular rule apply to everything else. Exceptions stay listed in the lexicon, so the rule can be tested without them. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2019-kaabi-rethinking-templates-syntactic-analysis-verbal/))
- **to try** — : lemmatize English MWEs (phrasal verbs, compound nouns) part by part, with each word's lemma from the in-code dictionary and the agreement rule of the MWE as a table. Measure the MWE-level accuracy on the EWT test MWEs, and the share of MWEs whose lemma is wrong in the current lexicon. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2019-schmitt-neural-lemmatization-multiword-expressions/))
- **to try** — Align the consonant skeletons of Hebrew and Aramaic cognates monotonically, with one-to-one character links, and read off the regular sound correspondences. Use the cognate sets and, and check the result against hand-annotated pairs. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2020-gasser-character-alignment-morphologically-complex-translation/))
- **to try** — Keep morphological glosses separate from the lexical gloss, as the extractors do, so that the Hebrew verb analysis can be checked feature by feature. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2020-jyra-helfi-hebrew-greek-finnish-parallel/))
- **to try** — The epsilon link (no counterpart in the translation) is cleaner than forcing a gloss. Mova's Hebrew-to-English bridge should allow an explicit null link. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2020-jyra-helfi-hebrew-greek-finnish-parallel/))
- **to try** — Build an analyser from the annotated part of a small corpus by paradigm completion, and filter it before use. Where the same lemma, gloss, POS and features appear with different spellings, keep the entry with the highest count. This suits the Aramaic varieties with little data, and, once a small annotated set exists. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2020-khalifa-morphological-analysis-disambiguation-gulf-arabic/))
- **to try** — Use the object marker et as a diagnostic in the lexicon: an internal argument with et means verbal structure; a noun whose argument is a PP with a noun-selected preposition is nominal. Add this as a feature on nouns in the Hebrew lexicon, not as a rule about word order. ([the paper](https://research.elamur.ai/free/papers/lingbulk-2020-preminger-functional-structure-noun-phrase-revisiting/))