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

[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.

- **kept** — Constrain by a wordlist: in unpointed mode, restrict each token's candidates to forms in Mova's lexicon, then score them. The paper's largest single gain (6 to 11 points) comes from this lookup step, not from a bigger model. It fits the unpointed mode, with the lex table as the list. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2021-rubinstein-nlp-dh-pipeline-transfer-learning/))
- **tried** — Illusory correspondences inside English. The interlingua lexicon anchors two Bibles on equal lemmas (the longest common subsequence). Equal lemmas across centuries are Postgate's illusory correspondences: KJV \*let\* (hinder), \*prevent\* (go before), \*conversation\* (conduct), \*quick\* (alive), \*meat\* (food), \*suffer\* (allow). Build a small test list of such pairs. Measure how often the reader anchors on them, and whether WEB's word at that place is a different concept. Anchors that fail it should be cut by context, not by lemma. ([the paper](https://research.elamur.ai/free/papers/postgate-1922-translation-translations-theory-practice/))
- **tried** — A lower-bound gate for lexicon merges (open problem 2): accept a merge, or a chunk choice in the encoder, only when the 95 per cent lower bound of the co-occurrence rate clears a threshold. This is a closed formula over counts that could decide cases like \*be\* with \*have\*, where today's merges are noise. Test it on the residual in bits. ([the paper](https://research.elamur.ai/free/papers/riding-2008-statistical-glossing-language-independent-analysis/))
- **tried** — A count test for lexicon merges. Merge two concepts (or two meta tokens) only when their proportions over the same contexts agree under a dissimilarity like the paper's, computed from counts. This targets open problem 2. Measure the merges kept, including the \*be\*/\*have\* case, and the reader plus encoder bits on dev. ([the paper](https://research.elamur.ai/free/papers/sanchez-2020-martinez-learning-synchronous-context-free-grammars/))
- **measured, no gain** — Tag-conditioned lemma rules in : when the dictionary has several candidates or a suffix rule is ambiguous, choose the rule with the UPOS and FEATS predicted; measure lemma accuracy on EWT dev, reporting ambiguous tokens (more than one lemma) separately from the rest. ([the paper](https://research.elamur.ai/free/papers/ud-2019-malaviya-simple-joint-model-improved-contextual/))
- **to try** — Give each concept a WordNet-style sense identifier, as the task does, and compare it with the meta-world's own concept identifiers on a sample of verses. ([the paper](https://research.elamur.ai/free/papers/abzianidze-2019-first-shared-task-discourse-representation/))
- **to try** — Cognate matching as a minimum-cost flow over an Aramaic-Hebrew word list: one side is Aramaic lexicon entries, the other Hebrew lemmas, each edge costed by a weighted consonant edit distance, each word matched at most once. The solver is deterministic, so this needs no neural model at all. ([the paper](https://research.elamur.ai/free/papers/acl-2019-luo-neural-decipherment-minimum-cost-flow/))
- **to try** — Unlabelled data for the lookup: the in-code dictionary in gives modern forms without labels. Run a simple EM over archaic-to-modern pairs from the lessons, and measure accuracy at 500, 1,000 and 5,000 tokens against the paper's curve. ([the paper](https://research.elamur.ai/free/papers/acl-2020-makarov-semi-supervised-contextual-historical-text/))
- **to try** — Orthographic normalization as a table: the decoder learns to map non-standard spellings to a conventional form. For Hebrew, the analogue is plene and defective spelling (matres lectionis) in unpointed text. A deterministic normalization table in could run before lemma lookup, measured against a gold list from dictionary hebrew. ([the paper](https://research.elamur.ai/free/papers/acl-2020-zalmout-joint-diacritization-lemmatization-normalization-fine/))
- **to try** — Two-lexicon intersection for Hebrew verbs: a root lexicon (three consonants, slots open) and a pattern lexicon (binyan vowels and affixes) whose intersection yields the stem. The root list is the missing piece, since dictionary hebrew has no roots yet. ([the paper](https://research.elamur.ai/free/papers/aclbulk-1988-kataja-finite-state-description-semitic-morphology/))
- **to try** — A count probe: list the candidate compound tokens in EWT test that the dictionary covers, and report how many of them the tokenizer currently splits. This gives the size of the problem before any change. ([the paper](https://research.elamur.ai/free/papers/aclbulk-1992-webster-tokenization-as-initial-phase-nlp/))
- **to try** — Keep UD tokens as they are, but mark compound units (phrasal verbs, fixed expressions such as \*in spite of\*) as spans from a list in the in-code dictionary, and pass the span to the parser as a feature; measure on EWT test the LAS of fixed, compound:prt and flat arcs with en parse-eval, and without the span feature. Joining them into single tokens would break UD tokenization, so the span is the safe form of the paper's idea. ([the paper](https://research.elamur.ai/free/papers/aclbulk-1992-webster-tokenization-as-initial-phase-nlp/))
- **to try** — Write the definition of a token used in Mova as a short rule list, and check it against the EWT tokens that the paper's criteria would split or join. ([the paper](https://research.elamur.ai/free/papers/aclbulk-1992-webster-tokenization-as-initial-phase-nlp/))
- **to try** — Count threshold check on the lexicon: test whether removing rare-word features in hurts ATIS and EWT differently, which is the paper's low-count claim transferred to tagging. ([the paper](https://research.elamur.ai/free/papers/aclbulk-1995-collins-prepositional-phrase-attachment-through-backed/))
- **to try** — Concept back-off through onto (WordNet hypernyms) for nouns not seen in the counts. Measure: accuracy on held-out compounds whose nouns are unseen, and without the back-off. ([the paper](https://research.elamur.ai/free/papers/aclbulk-1995-lauer-corpus-statistics-meet-noun-compound/))
- **to try** — Filter instead of rewriting: let the lexicon overgenerate and compose a filter that forbids illegal clitic combinations (an article with an indefinite ending, a preposition with a nominative or accusative). For, the prefix and suffix tables could be checked with such a filter, so that a vav, bet, lamed or mem prefix cannot combine with a suffix the grammar forbids. ([the paper](https://research.elamur.ai/free/papers/aclbulk-1998-beesley-arabic-morphology-only-finite-state/))
- **to try** — Lexical flags that forbid an explicit partitive construction for some quantifiers (the paper's col case). Keep them as a feature on the lexical entry, not as a rule. ([the paper](https://research.elamur.ai/free/papers/aclbulk-1998-netzer-generating-determiners-quantifiers-hebrew/))
- **to try** — Use the paper's round-trip check: generate every form from its lemma and tag, analyse it back, and count the mismatches. Run it on the in-code dictionary and report the count per run, twice, to see that it is stable. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2000-minnen-robust-applied-morphological-generation/))
- **to try** — Replace the thesaurus with a deterministic back-off: for a rare noun2, use its WordNet lexicographer file. Measure accuracy on the subset of PP cases whose noun2 has fewer than five counts. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2000-pantel-unsupervised-approach-prepositional-phrase-attachment/))
- **to try** — A WordNet analogue in onto: map each noun to its hypernym chain at a chosen depth, as the paper maps to MeSH, and classify relations by a counted table with back-off. Measure: accuracy on held-out EWT compounds whose nouns are absent from training, against the word-only baseline. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2001-rosario-classifying-semantic-relations-noun-compounds/))
- **to try** — : group the forms of that are not yet linked to a lemma into conflation sets from a raw English corpus, using the orthographic and distributional cues of the paper. Measure the F-score of the sets against the dictionary on a held-out part of the forms. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2001-schone-knowledge-free-induction-inflectional-morphologies/))
- **to try** — : rank candidate pairs of English words in that are not yet linked, by edit distance and by a co-occurrence score from a raw corpus. Measure precision at 500 and 1,000 pairs against the existing links in the dictionary. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2002-baroni-unsupervised-discovery-morphologically-related-words/))
- **to try** — For defective Hebrew roots, generate the variants of a short stem (doubled last letter, weak letter in the middle or at the ends) and keep only those that exist in the root list of the lexicon. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2002-darwish-building-shallow-arabic-morphological-analyser/))
- **to try** — An "unseen word" state that emits the bare word with a fixed mass, rather than a guessed vowel pattern. Gal names morphological analysis as the next step for unseen words; dictionary hebrew has no roots yet, so it is the place where that analysis would come. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2002-gal-hmm-approach-vowel-restoration-arabic/))
- **to try** — A lemma policy for the lexicon: lemma = least marked form, a different part of speech means a different lemma, and the root is an optional grouping key. Use it; the root grouping could also fill the gap that dictionary hebrew has no roots yet. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2002-kamir-comprehensive-nlp-system-modern-standard/))
- **to try** — Support-verb units as the first level: light-verb frames such as "take into account" or "make use of" with a PP, taken from the MWE list. Check the precision on EWT train before the rule is used. In the paper this level is very reliable (97 cases at 100% on German); the English count is unknown until measured. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2002-volk-combining-unsupervised-supervised-methods-pp/))
- **to try** — Ambiguity filter: count how many EWT noun tokens have more than one WordNet supersense (the paper's figure is 72% of tokens for its corpus) and train only on unambiguous nouns. Report the size of the training set before and after the filter. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2003-ciaramita-supersense-tagging-unknown-nouns-wordnet/))
- **to try** — Gloss instances for -style training: turn WordNet glosses into labelled noun instances and add them to the averaged-perceptron training of a noun-class model, and measure accuracy on a held-out split of the open treebanks, with two identical runs. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2003-ciaramita-supersense-tagging-unknown-nouns-wordnet/))
- **to try** — or the onto lookup: when a noun has no WordNet entry, give a supersense from a rule-plus-perceptron step over the head noun and its suffix, and print the rule in the explanation. Measure supersense accuracy on the unknown nouns of EWT against the "person" baseline. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2003-ciaramita-supersense-tagging-unknown-nouns-wordnet/))
- **to try** — A rule-only root candidate filter: list the roots from the root table (2,152 roots in the union of the two sources the paper uses) whose radicals occur order with only w or y allowed between them, and rank them with the paper's three-class score. It needs no learning, and it would give dictionary hebrew the root column it lacks. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2004-daya-learning-hebrew-roots-machine-learning/))
- **to try** — A bootstrapped lexicon: run a cheap, high-recall rule over a large corpus, hand-correct the output once, and keep the corrected list as a lookup table. The paper's 3,580-stem list shows the cost is small once the cheap rule exists. For Mova, the same step can build lemma lists from the MACULA-derived lexicon in dictionary hebrew (tera/dictionary/src/hebrew.rs). ([the paper](https://research.elamur.ai/free/papers/aclbulk-2004-goweder-identifying-broken-plurals-unvowelised-arabic/))
- **to try** — : add a domain-aware filter that removes WordNet senses with a lexfile or domain label that never occurs in the target genre. Measure the WSD accuracy on the EWT test set (which mixes genres) and on a single-genre subset, with the filter and off, and run twice to confirm. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2004-mccarthy-automatic-identification-infrequent-word-senses/))
- **to try** — Ranking check: compare the current predominant-sense ranking in onto with a ranking by the Lesk gloss overlap the paper favours for nouns, verbs, adjectives and adverbs. Report the agreement and the accuracy on EWT with each ranking. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2004-mccarthy-automatic-identification-infrequent-word-senses/))
- **to try** — Thesaurus from open text: build neighbours for nouns from en parses of an open corpus (not the BNC), and measure how often the removed sense is one that the test corpus never uses; the SemCor figure (38% random) is the baseline. Use SemCor only as a test set. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2004-mccarthy-automatic-identification-infrequent-word-senses/))
- **to try** — : when a nominal has no verb in a clause, fill the event from a default frame, with the verb chosen from VerbNet or WordNet coverage. Mark every filled role as default in the output. Measure the number of filled roles on EWT test and check 50 of them by hand. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2004-mcshane-ontosem-methods-processing-semantic-ellipsis/))
- **to try** — : a synonym vote for unknown nouns. Take the nouns that co-occur with the target in a large open corpus (parsed with en), let each WordNet neighbour vote with its lexicographer file, and keep the top 50 as the explanation. Measure accuracy on the unknown nouns in an open treebank, and check the first two runs are the same. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2005-curran-supersense-tagging-unknown-nouns-semantic/))
- **to try** — Store the lexeme as the deepest key and the root only as an index field. The paper argues that meaning cannot be predicted from root and pattern (its maktab and maktaba example), so a lexicon keyed on the root would mislead. The rule belongs, and for Aramaic. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2005-habash-morphological-analysis-generation-arabic-dialects/))
- **to try** — Nearest-compound classifier in with WordNet path similarity between modifier and head synsets, copying the relation of the nearest training compound. Measure: accuracy on EWT noun compounds against the majority-class baseline, two identical runs. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2005-kim-automatic-interpretation-noun-compounds-wordnet/))
- **to try** — Use the letter model selectively, only on words unknown to the word model; the authors found that better than running it on everything. A two-tier rule (lexicon first, letter model second, with the tier recorded) keeps each pointing explainable. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2005-nelken-arabic-diacritization-weighted-finite-state/))
- **to try** — Irregular entries as overrides: keep a general lexicon and separate add, replace and remove lists, combined as (general minus remove), plus add, with replace taking priority. For irregular nouns and verbs in this gives one visible place for exceptions that the general rules should not see. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2005-yona-finite-state-morphological-grammar-hebrew/))
- **to try** — A fallback for unknown tokens: enumerate all segmentations with affix tags, give the unanalysed base form a single tag UK, and set its tag from the corpus distribution of similar forms. This is a rule to try in for tokens the lexicon misses. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2006-adler-unsupervised-morpheme-based-hmm-hebrew/))
- **to try** — Initial probabilities from a context-free approximation: similar-word statistics (the paper's Levinger-style prior) lift the word model from 82.01 to 84.08 with a first-order HMM. Measure this prior for the Hebrew lexicon before any model is trained on it. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2006-adler-unsupervised-morpheme-based-hmm-hebrew/))
- **to try** — Start the lexicon small: the 321-entry list of closed-class words and frequent items gave most of the gain, and the 1,560-entry extension added ambiguity that lowered F1 with gold tags. For a Biblical-to-Mishnaic (or Aramaic) pair table, measure each addition on held-out verses before keeping it. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2006-chiang-parsing-arabic-dialects/))
- **to try** — A root list as a precision knob: the paper orders restrictions from "all roots" to a full lexicon and reports type precision and recall for each. For Hebrew, the roots and lemma list from the existing data can restrict the root guesses of an unpointed token, reported as a table of the same shape. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2006-habash-magead-morphological-analyzer-generator-arabic/))
- **to try** — Store the lexicon in a notation between orthography and phonology, and render orthographic forms from it on demand. The paper's reason is that the model is then written once and applies to unpointed and pointed text alike. This is the design that needs for its pointed and unpointed modes, and it fits the abstract notation idea in the Mova plan for. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2006-smrz-tips-tricks-prague-arabic-dependency/))
- **to try** — Keep lexical features only for the closed classes (prepositions, the conjunction "and", punctuation) in any rule-based Hebrew NP or chunk rule, and test by removing the open-class lexicon on noisy tagging. The paper reports that such a small lexical set is enough on gold tags and collapses on noisy tags, so the test must use noisy tags. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2007-goldberg-svm-model-tampering-anchored-learning/))
- **to try** — Gate any fallback analyses on the sentence probability: use an extra candidate source (for example a character model for unpointed forms) only when the best path under the lexicon model has low probability. Calibrate the threshold on a dev part and report accuracy by part, as the paper's ATB1, ATB2 and ATB3 split does. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2007-manour-smoothing-lexicon-based-pos-tagger/))
- **to try** — A root-first lexicon layout: entries sit under their root, grouped by pattern, with the gloss and the lexical class, as in the paper's figure. dictionary hebrew has no roots yet; a root-first file with the same layout would be the first step for, and the structure can be read and by the Rust structs. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2007-smrz-elixirfm-implementation-functional-arabic-morphology/))
- **to try** — Run the transducer twice, and without a root lexicon. Without the lexicon, the grammar proposes candidate roots for an unknown form; with it, the output is limited to known roots. For, a no-lexicon mode that lists candidate roots for unknown words would be a useful and explainable fallback. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2009-barthelemy-une-description-morphologique-structuree-en/))
- **to try** — Measure how much corpus Mova's lexicon needs and without prefix splitting, on the same corpus size, by counting distinct lexemes. This is a cheap check of the segmenter's value for world. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2009-dinur-unsupervised-concept-discovery-hebrew-simple/))
- **to try** — A root filter: once a list of roots exists, compose a root acceptor onto the analyser so that analyses with non-existent roots are rejected. Hebrew has no root list yet (dictionary hebrew has none), so the paper's gap is ours too; a roots table or the Wiktionary-derived tools would close it. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2009-gasser-semitic-morphological-analysis-generation-finite/))
- **to try** — A baseline: accept any trilateral root first, record the automaton size and the number of spurious roots, then add a root list and measure again. The difference is the cost of the missing lexicon, a number Mova can report honestly. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2009-hulden-revisiting-multi-tape-automata-semitic/))
- **to try** — : learn lemma rules that can change a prefix or infix and, and use them only for forms that the dictionary does not cover. Measure lemma accuracy on the uncovered forms of EWT test, against the suffix-only rules. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2009-jongejan-automatic-training-lemmatization-rules-that/))
- **to try** — A cheap check in before any sense induction: for each EWT compound, measure the overlap between the context words of the whole and those of its head. Label it compositional when the overlap is high. Measure: accuracy on a hand-checked sample of 200 WordNet MWEs, two identical runs. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2009-korkontzelos-detecting-compositionality-multi-word-expressions/))
- **to try** — Build a small compositionality set from WordNet multi-word lexemes the way the paper does, as a test set only. The WordNet data is open, but the judgments are the authors' and must be checked by hand. Measure: agreement with a second annotator on 50 items. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2009-korkontzelos-detecting-compositionality-multi-word-expressions/))
- **to try** — Keep the reason with the result: every grammatical value the rule file gives is traceable to a rule and a lexicon entry, and the paper's dump files show the trace. Mova's frames and world should keep the same trace so that each deduction can say why. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2009-peursen-how-establish-verbal-paradigm-basis/))
- **to try** — Measure the type reduction by splitting, and the growth curve by corpus size, for Hebrew and Aramaic text in the same way. The count is deterministic and needs only open text, and it says how big a segmenter lexicon must be before it stops growing. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2010-alotaiby-clitics-arabic-language-statistical-study/))
- **to try** — All-caps normalisation before tagging in : when a whole sentence is uppercase, lowercase it and restore names from the lexicon. Measure tag accuracy and LAS on an uppercased copy of EWT test that we make ourselves. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2010-foster-cba-check-spelling-investigating-parser/))
- **to try** — Negative control for spelling correction in : correct only out-of-lexicon tokens with a one-edit closed-class candidate, and log each change. Measure LAS on EWT test with misspelled function words added by script, and check that clean text does not change at all. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2010-foster-cba-check-spelling-investigating-parser/))
- **to try** — Homograph lexicon for unvocalized forms: where one written form has several readings (a particle reading and a verb reading, for example), store all candidates with a context rule that chooses among them, instead of one fixed tag. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2010-green-better-arabic-parsing-baselines-evaluations/))
- **to try** — Lemma classes for unknown lemmas: encode a lemma as the list of insert and delete operations (with positions) that turn the stem into it, and count which operations recur. In this gives a rule-based guess for an unseen Aramaic lemma, checkable against the dictionary. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2010-mcclanahan-probabilistic-morphological-analyzer-syriac/))
- **to try** — Report decision-level accuracy (the number of decisions a human annotator would correct) next to word accuracy, and keep the naive most-frequent baselines as the floor in the same table. The baseline for an unknown root is the first three consonants of the baseform; that is a cheap rule to test. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2010-mcclanahan-probabilistic-morphological-analyzer-syriac/))
- **to try** — Use the candidate list as a coverage probe for en tok-eval: how many EWT test multi-word units are absent from the lexicon. Measure: the share covered, before any lexicon change. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2010-ramisch-multiword-expressions-wild-mwetoolkit-comes/))
- **to try** — Keep two layers for each noun: the form-based features (what the word shows) and the function-based features (what it does in agreement). Derive the form layer from the MACULA morphology and add the function layer as a lexicon field. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2011-alkuhlani-corpus-modeling-morpho-syntactic-agreement/))
- **to try** — Corpus-frequency filter for a seed lexicon: keep a lemma only if it has at least one web hit and at least ten corpus occurrences. Run this per language layer (biblical, mishnaic) with counts from the corpora Mova already holds, so that layers are not mixed. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2011-attia-open-source-finite-state-morphological/))
- **to try** — Restrict each segment to the tags its lexicon entry licenses, as the paper does. In this is a deterministic check on the lex and pre tables: a segment is kept only with a tag its table entry allows. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2011-goldberg-joint-hebrew-segmentation-parsing-pcfgla/))
- **to try** — Synonym choice without a thesaurus: a synset is a set of source words that one translation renders with the same English word, with the concordance giving the senses. For Mova, the lemmas and Strong's numbers in dictionary hebrew could give Hebrew synsets, and the test is whether synonym choice separates the layers of the Pentateuch. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2011-koppel-unsupervised-decomposition-document-into-authorial/))
- **to try** — Measure the undiacritized lemma as a parser feature: the paper's strongest lexical feature is the lemma without vowels, which is exactly what unpointed Hebrew input offers. Compare gparse on unpointed output and without a consonantal-lemma feature, using dictionary hebrew as the lemma source. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2011-marton-improving-arabic-dependency-parsing-form/))
- **to try** — Report coverage in two numbers: tokens without an analysis and types without an analysis. Types show the lexicon gaps more clearly (5.3 percent) than tokens (1.75 percent). A coverage gate on a held-out slice would show where the lexicon of he is thin. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2012-albert-morphologically-annotated-hebrew-childes-corpus/))
- **to try** — Report seen and unseen words separately for every morphological table. Here the seen-word score is 97% and the unseen score is 21% for the same lookup. A table that reports only the total would hide this, so should state both numbers for any Hebrew table. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2012-alkuhlani-identifying-broken-plurals-irregular-gender/))
- **to try** — : encode each suffix-based lemma change of the dictionary as one label, and let the averaged perceptron of learn the labels with the tag and the neighbouring words as features. Measure lemma accuracy on EWT test, and the accuracy on forms that the dictionary does not contain. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2012-gesmundo-lemmatisation-as-tagging-task/))
- **to try** — Use the label set as a check on the dictionary: count the labels that occur only once, and list them for review, since a rare label is often a dictionary error. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2012-gesmundo-lemmatisation-as-tagging-task/))
- **to try** — Compile the segmenter's tables into one finite-state lexicon, with the compatibility rules composed in one at a time rather than intersected at once. For this means the prefix, stem and suffix tables learned from MACULA become a transducer, and the composition order is recorded and measured by state count. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2012-hulden-conversion-procedural-morphologies-finite-state/))
- **to try** — Keep the lexicon fully pointed and delete the vowels optionally on the surface, so that one table serves both the pointed and the unpointed mode. This removes a second table that has to be kept in step with the first. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2012-hulden-conversion-procedural-morphologies-finite-state/))
- **to try** — onto lexfile layer: tag nouns of English running text with their WordNet supersense from the first-sense lexfile, with no disambiguation, and measure what share of EWT noun tokens is covered. The paper's Arabic figure is about half of all tokens; the English figure is the one to measure. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2012-schneider-coarse-lexical-semantic-annotation-supersenses/))
- **to try** — A seed list of affect terms for Modern Hebrew: the questionnaire-derived emotion, conflict and trust words, with their English equivalents, as a labelled list the pragmatic layer can use to flag a word as affect-laden or conflict-laden instead of guessing from the dictionary entry. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2012-strapparava-corpus-based-explorations-affective-load/))
- **to try** — Keep fusion as data: each rule maps two atomic prefixes to one orthographic form and one gloss, with a wildcard for the stem. Any Hebrew fusion in the prefix chain (checked against the treebank first) goes into a small rule table with its examples. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2012-zaraket-arabic-morphological-analyzer-agglutinative-affix/))
- **to try** — A transliteration lexicon built from an English term list: generate the Hebrew spellings an English term may take, match them to unknown forms in a corpus, and keep the pairs that pass a heuristic filter. Check a random sample by hand to estimate precision, as the paper does. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2013-cohen-effect-out-vocabulary-terms-inferring/))
- **to try** — Classify unknown tokens before segmenting them: transliteration, misspelling or missing lexicon word (the paper's split is 71.5%, 16% and 13.5%). Transliterations go to a loan-word path; the other two go to the ordinary prefix-suffix search. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2013-cohen-effect-out-vocabulary-terms-inferring/))
- **to try** — Translate from scratch when building a gold set: the paper found that editing machine output took longer than translating by hand with dictionaries, and that the first text of each cluster took longest. Any Hebrew or Aramaic gold set Mova builds should follow the same rule, and budget the first items of each topic higher. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2013-elhadad-multi-document-multilingual-summarization-corpus/))
- **to try** — and : group multiword names and idioms as one unit, parse the unit internally, and reattach the arcs to its head, as the paper does; measure head attachment on EWT test sentences that contain such units. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2014-kong-dependency-parser-tweets/))
- **to try** — A squeezing generator for loan verbs: given a loan noun, delete its vowels to get a consonant root, place the root in the array, and fill the seven verb patterns. Output the candidates with a flag "unchecked" for a lexicon entry. This is a deterministic table, not a model. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2014-ornan-innovation-verbs-hebrew/))
- **to try** — Squeezing from phrases: because the procedure also works on two-word idioms, a verb-noun expression can yield a new verb root. Use this as a candidate rule when a Hebrew verb-noun expression has no single verb in the lexicon. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2014-ornan-innovation-verbs-hebrew/))
- **to try** — Multiword prepositions (\*because of\*, \*instead of\*, \*according to\*) stay separate tokens, as UD wants, but get one supersense through a span list in the in-code dictionary, joined in the tree by fixed. Measure how many such units EWT test has and how often gives them the fixed arcs, before any label is assigned. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2015-schneider-hierarchy-preposition-supersenses/))
- **to try** — Re-rank, do not generate: keep the segmenter's or lexicon's ranked analyses and choose the first one that is consistent with the parse (same number of clitics, same case). Their retagging step does exactly this and shows that it changes few tokens but many nominals. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2015-shahrour-improving-arabic-diacritization-through-syntactic/))
- **to try** — and : a closed-class lexicon feature, so that a wrong preposition or determiner is not allowed to cascade into the head choice; measure robustness F1 on closed-class errors before and after. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2016-hashemi-evaluation-parser-robustness-ungrammatical-sentences/))
- **to try** — Constituent order as a rule: the authors report that 82% of their VN-MWEs do not allow the order of the two words to change, and 87% have non-compositional syntax. In a UD tree, a verb with an object whose order cannot change, and whose pair is in the lexicon, is an MWE candidate. Implement the rule and measure it on a gold set. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2016-liebeskind-semantically-motivated-hebrew-verb-noun/))
- **to try** — : use a letter-to-sound key as a second check on candidate modern forms, built from the dictionary's pronunciations where available. A candidate must match the historical form on the sound-class key. Measure the candidate ranking accuracy on the test split of tools/gutenberg lessons. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2017-bollmann-learning-attention-historical-text-normalization/))
- **to try** — Report accuracy separately for templatic and concatenative (strong and weak) verbs. In Hebrew, the weak verb classes (pe-nun, lamed-he, ayin-waw) are where stem variation hurts, so a per-class table shows where a rule or a lexicon is needed. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2017-borg-morphological-analysis-maltese-language-challenges/))
- **to try** — Use the cluster removal rate as a measure of a generated lexicon: a stem that is too short (their example ittra, "letter", which swept in unrelated words) must be blocked by a minimum stem length rule. Measure how many false groupings the rule removes. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2017-borg-morphological-analysis-maltese-language-challenges/))
- **to try** — Dictionary matching against a predicted span: the paper finds span prediction beats dictionary matching by 0.8 points; compare matching with a perceptron span predictor on EWT test before adding the joint model. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2017-kato-english-multiword-expression-aware-dependency/))
- **to try** — and : merge named-entity and fixed MWE spans into one unit before parsing, with flat and fixed relations as in the converted corpus; metric: LAS on EWT test and ATIS (current LAS 63.0), with the MWE-internal attachments scored separately. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2017-kato-english-multiword-expression-aware-dependency/))
- **to try** — Keep the affix information as a rule. The paper's fixed three-character prefixes and suffixes give most of the gain without any lexicon. The Hebrew analogue is a table of proclitics (the conjunction ו, the prepositions ב, כ, ל, מ) read as features, not as a lookup. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2017-zalmout-don-t-throw-those-morphological/))
- **to try** — Rank the candidate analyses of a word by a weighted feature-match score: each candidate gets one point per feature that agrees with a per-feature tagger, with the weights tuned on held-out data. For Mova, the candidate splits and the readings of dictionary hebrew can be ranked this way. The weights are a small table, so the choice stays deterministic and can be explained. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2017-zalmout-don-t-throw-those-morphological/))
- **to try** — A sampled-precision check: draw a fixed random sample of any new links, check each by hand, and report the rate with its date, before any added relation enters world or the dictionary. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2018-batita-derivational-relations-arabic-wordnet/))
- **to try** — Root bags as a first step: group the entries of dictionary hebrew by root. The dictionary has lemmas and Strong's numbers but no roots yet, so a root source comes first. Measure how many bags contain more than one POS. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2018-batita-derivational-relations-arabic-wordnet/))
- **to try** — Keep two parallel layers in the Mishnaic lexicon: a context-free category and an in-context function for each form. The parser then decides the function, and the lexicon still shows that a form such as a verb-or-adjective reading exists. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2018-giovannetti-constructing-annotated-resource-part-speech/))
- **to try** — Use the 10-fold cross-validation with the 87.9% figure as the bar for a Mishnaic test, and first compare a deterministic lexicon-plus-rules tagger with it, before any learned model is added. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2018-giovannetti-constructing-annotated-resource-part-speech/))
- **to try** — A coverage report for every test text: the paper gives the share of types and tokens unseen in training (9% of types, 3% of tokens for its test). Mova should report the same two shares for each hbo or arc test text, so that a lemma back-off effect is measured where it matters,. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2018-shapiro-morphological-word-embeddings-arabic-neural/))
- **to try** — Lemma back-off for unseen forms: when a surface form is not in the world, look up its lemma's entry instead. The paper's out-of-vocabulary handling rests on the lemma part of the vector. For Mova this is a deterministic table: form to lemma, from dictionary hebrew or the segmenter, and a fallback to the lemma's entry in world. ([the paper](https://research.elamur.ai/free/papers/aclbulk-2018-shapiro-morphological-word-embeddings-arabic-neural/))