11. Modern Hebrew in Universal Dependencies

How Modern Hebrew is cut into words and annotated in Universal Dependencies (UD): the segmentation of prefixes and suffixes into syntactic words, multiword tokens, the tags and features, the relations specific to Hebrew (compound:smixut, case:acc, case:gen, nsubj:cop, mark:q, compound:affix), and the four UD Hebrew treebanks — HTB, IAHLTwiki, IAHLTknesset and PostRab — with what each one does differently. The counts in this chapter were made on the UD 2.18 release files on 2026-10-17 (see § 11.9).

← README · ← 10 Historical layers · 12 Mova and Hebrew →

Related chapters: the orthographic facts that make segmentation hard are in 02 Script and spelling; the construct state in 04 Nouns; the inflected prepositions in 06 Prepositions; the copula and existentials in 07 Syntax.

Contents

• 11.1 The problem: one space-delimited token, several words
• 11.2 Tokens, words and multiword tokens
• 11.3 Two tokenization schemes: old HTB and IAHLT
• 11.4 Parts of speech
• 11.5 Features
• 11.6 Relations
• 11.7 Constructions, one by one
• 11.8 The treebanks
• 11.9 How the counts were made
• 11.10 Parsing Hebrew: the state of the art
• 11.11 Pitfalls for a converter
• References

11.1 The problem: one space-delimited token, several words

Hebrew writes a closed set of function words as prefixes on the next word, and pronouns as suffixes on prepositions, nouns and (in formal style) verbs. The prefixes are the seven letters remembered by the mnemonic מש״ה וכל״ב Moshe ve-Khalev "Moses and Caleb": מ mi- "from", ש she- "that", ה ha- "the", ו ve- "and", כ ke- "as", ל le- "to", ב be- "in" (UD Hebrew guidelines, Tokenization; the same mnemonic is used in Hebrew school grammar). A single orthographic word can therefore hold a conjunction, a subordinator, a preposition, the article, a content word and a pronoun:

וכשבביתו
ve-khshe-ba-béyt-o
and-when-in.DEF?-house.CS-3MS.POSS
'and when in his house'

The real difficulty is not the length of the chain but its ambiguity. Unpointed text does not show vowels, so the same letters can be cut in different places. The textbook case from Adler and Elhadad (2006), repeated by Zeldes et al. (2022: 4331), is the four letters בצלם:

Two further facts make Hebrew segmentation non-concatenative in the sense of the older treebank:

1. The covert article. After ב, ל, כ the article ה is not written; its vowel fuses into the preposition: בבית is be-váyit "in a house" or ba-báyit "in the house". The letters do not tell which (UD Hebrew guidelines; see 02 and 06).
2. Fused pronouns. בית + הוא is written ביתו beyt-o "his house"; אותו ot-o "him" is the accusative marker את + הוא; לו l-o "to him" is ל + הוא. The surface letters of the pronoun (ו, ה, ם, ן, נו, כם…) are not the letters of the free pronoun.

Tsarfaty (2019: "What's wrong with Hebrew NLP") argues that this is why pipelines built for English fail on Hebrew: a tagger cannot tag what has not been segmented, and segmentation cannot be decided without syntax. The Hebrew NLP tradition therefore built joint morphological-syntactic models (Tsarfaty 2006; Cohen and Smith 2007; Goldberg and Tsarfaty 2008; Goldberg and Elhadad 2011; More et al. 2019) and lattice-based analysis where every possible segmentation is kept until the parser decides.

11.2 Tokens, words and multiword tokens

UD separates the token (what stands between spaces and punctuation) from the syntactic word (what gets a POS tag and a node in the tree). A token made of several words is a multiword token (MWT); in CoNLL-U it is a range line 3-4 followed by the word lines 3 and 4.

Source: the treebank cards in treebanks/he/ (Mova repo, generated from the UD 2.18 stats.xml) and our own count. About one token in three in contemporary Hebrew is a multiword token. Spoken Knesset Hebrew has fewer than Wikipedia (more pronouns as free words, fewer bound possessives), and the older, more literary PostRab texts have the most.

The shapes of the MWTs in IAHLTwiki train, by the UPOS of their parts (count 2026-10-17):

Note on the last row: שבו she-bo (lit. "that in-it") is the relative marker with a resumptive pronoun on a preposition. The possessive של "of" is one word in IAHLT, so שלו "his" is ADP + PRON (07 Syntax, relative clauses).

Note that ADP DET NOUN (231) is rare beside ADP NOUN (7 451). This is the covert article: in IAHLT only מ "from" keeps a written ה (מהבית), while ב/ל/כ swallow it and the definiteness moves to a feature (§ 11.3).

11.3 Two tokenization schemes: old HTB and IAHLT

UD Hebrew has two schemes in its releases, and a converter has to know which one a file follows.

The old HTB scheme (UD_Hebrew-HTB as released in UD 2.18)

The HTB was built by MILA from 1990–91 issues of the daily Ha'aretz as a constituency treebank (Sima'an et al. 2001), turned into dependencies by Goldberg and Elhadad (2009), brought into early UD by Tsarfaty (2013) and into UD v2 by Sade et al. (2018) (Zeldes et al. 2022: § 2). Its tokenization restores elements that are not written:

• the covert article becomes a word of its own, with an underscore on the side where it fused: ב + ה_ + בית;
• a fused possessive becomes noun + של + pronoun (house of him);
• a fused object pronoun becomes verb + את_ + _הוא;
• the pronoun keeps the lemma of the free pronoun and the form gets an underscore: _הוא for the ו of ביתו.

In the UD 2.18 HTB train file we count 2 547 explicit ה_ article words. The guideline page for Hebrew still describes this scheme ("Tokenization and Word Segmentation"; UD Hebrew docs, _he/index.md).

An example of the shape (HTB is CC BY-NC-SA, so we describe and cite only a fragment of one tree, he_htb-ud-train.conllu):

14-16  בידיים
14     ב       ADP   case
15     ה_      DET   det      PronType=Art
16     ידיים   NOUN  obl      Gender=Fem|Number=Plur

ba-yadáyim "with the hands": three words for a token of two written morphemes.

The IAHLT scheme (IAHLTwiki, IAHLTknesset, and the revised HTB on GitHub)

Zeldes, Howell, Ordan and Ben Moshe (2022) changed the scheme so that an MWT is always the concatenation of its parts: nothing is inserted. Their three examples (2022: 4333):

The covert article is the only change that loses information, so it is kept as a feature on the preposition: Definite=Def|PronType=Art on ב/ל/כ. We count 2 241 such prepositions in IAHLTwiki train. The pronoun suffix keeps the lemma of the free pronoun (ו → הוא) and carries Poss=Yes|Case=Gen for a possessive.

The IAHLTknesset example from its README (CC BY-SA 4.0; # sent_id = 13_ptm_532410-8) shows both points: בעולם is split into ב (ADP, Definite=Def|PronType=Art) + עולם, and לו into ל + ו (lemma הוא).

The authors give three reasons for the change: the new scheme is simpler, it matches UD Arabic and the other Semitic treebanks, and it allows character-level segmenters that only decide "does a new word start here?" (Zeldes 2018; Zeldes et al. 2022: § 4). Converting between the two is deterministic in both directions (2022: 4334) — a fact a Mova converter can use (§ 11.11).

The IAHLT project also released a revised HTB in the new scheme (github.com/IAHLT/UD Hebrew), valid for UD 2.10. The official UD_Hebrew-HTB in UD 2.18 has not taken it: our count of its files still finds ה_, case:acc (1 946) and case:gen (4 812). Whoever trains on "HTB" must say which HTB.

11.4 Parts of speech

UD Hebrew uses all universal tags except PART (guidelines). The points where Hebrew practice is not obvious:

UPOS frequencies (all splits, 2026-10-17):

The genre shows plainly: spoken Knesset Hebrew has more than twice the share of PRON, VERB and ADV of Wikipedia, and nine times as many INTJ; Wikipedia has the most PROPN. PostRab has no punctuation in the text.

The PROPN column is not fully comparable: Zeldes et al. (2022: § 5) found that HTB tags transparent place names such as תל אביב tel aviv or כפר סבא kfar saba as NOUN, while IAHLT tags them PROPN; likewise דומה dome "similar" and שונה shone "different" are participles (VERB) in HTB but ADJ in IAHLT.

11.5 Features

The universal features used, and the Hebrew-specific ones. Frequencies are for the train split (2026-10-17).

HebBinyan counts over all splits (2026-10-17):

The passive binyanim huf'al and pu'al are about seven times rarer in parliamentary speech than in Wikipedia — a register fact (05 Verbs, 08 Register). IAHLT adds NITPAEL for the literary nitpa'el (e.g. נתקבל nitkabel "was received"), a Mishnaic form (10).

The UD guideline page for HebBinyan (_he/feat/HebBinyan.md) lists the seven classic binyanim with an example of each and their typical meanings; the IAHLT data shows that the inventory had to be extended.

11.6 Relations

Hebrew subtypes

From the guidelines (_he/index.md, _he/dep/*.md) and the data:

The IAHLT scheme drops the subtypes that are predictable from the words they join — case:acc (always את), case:gen (always של), mark:q (always האם) — and writes plain compound for the construct (Zeldes et al. 2022: 4334). The information is still in the features: Case=Acc on את, Case=Gen on של, Definite=Cons on the construct head. A converter between the schemes is therefore a rewrite by lemma and feature (§ 11.11).

Relations not used in Hebrew at all, per the guidelines: clf and expl (we still count 1 and 4 expl in IAHLT, 4 iobj in HTB and 7 in IAHLTknesset — leftovers or edge cases).

Most frequent relations

Two things stand out. First, the construct state is the fifth most frequent relation in written Hebrew: about one word in every 14–16. Second, det is third everywhere because the article ה is a separate word in UD. The spoken corpus also has 228 vocative (forms of address to the chair: adoni ha-yoshev rosh "Mr. Chairman") against 1 in Wikipedia.

11.7 Constructions, one by one

11.7.1 The construct state (smixut)

The first noun of a construct is in a bound form and carries Definite=Cons; the second noun depends on it with compound:smixut (HTB, PostRab) or compound (IAHLT). Definiteness is marked once, on the last member, and spreads to the whole chain (04).

The UD Hebrew page compound:smixut (_he/dep/compound-smixut.md) defines it as the relation for "a noun modifying another noun in a construct state". The choice of compound (and not nmod) is a convention of UD Hebrew: the construct behaves like a single noun phrase head, cannot be interrupted by an adjective, and often lexicalises (בית ספר beyt séfer "school", lit. "house of book"). A construct whose second member is a pronoun suffix (ביתו) is instead nmod:poss.

Annotated example (IAHLTwiki, CC BY-SA 4.0, iahltwiki_penicillin-88):

# text = בנזילפניצילין הוא טיפול הבחירה לעגבת בכל שלביה.
1      בנזילפניצילין  NOUN  3  nsubj
2      הוא           PRON  3  cop        Polarity=Pos|PronType=Prs
3      טיפול         NOUN  0  root
4-5    הבחירה
4      ה             DET   5  det
5      בחירה         NOUN  3  compound
6-7    לעגבת
6      ל             ADP   7  case
7      עגבת          NOUN  3  nmod
8-9    בכל
8      ב             ADP   10 case
9      כל            DET   10 det        Definite=Cons|PronType=Tot
10-11  שלביה
10     שלבי          NOUN  3  nmod
11     ה             PRON  10 nmod:poss  Case=Gen|Poss=Yes|Person=3|Gender=Fem
12     .             PUNCT 3  punct

benzilpenitsilin hu tipul ha-bkhira le-agévet be-khol shlav-éha "Benzylpenicillin is the treatment of choice for syphilis in all its stages." Note three things: טיפול הבחירה "the treatment of choice" is a construct with the article on the second member (compound); כל "all" is itself in the construct (Definite=Cons) before its noun; and שלביה "its stages" is the construct plural שלבי + the feminine pronoun suffix (antecedent עגבת, feminine).

In the HTB and IAHLTknesset examples on the treebank card, the same pattern appears with ראש עיר rosh ir "mayor" (lit. "head of city": ראש Definite=Cons, עיר compound) and מדינת ישראל medinat yisra'el "the State of Israel" (מדינת is the construct of מדינה).

Counted: Definite=Cons is on 10 801 words of HTB, 11 014 of IAHLTwiki and 3 789 of IAHLTknesset — 6.7%, 7.8% and 5.7% of all words. Spoken Hebrew uses the construct less and the analytic של more (04).

11.7.2 Copular clauses

Present-tense Hebrew needs no verb "to be": הבית גדול ha-báyit gadol "the house [is] big" is a full clause with the adjective as root and הבית as nsubj. When a copula appears, UD follows its general rule: the predicate is the head and the copula depends on it with cop.

Hebrew has several copulas (guidelines, Non-verbal clauses):

Both schemes tag the pronominal copula PRON (HTB: 380 הוא/PRON as cop, beside 588 היה, 289 אינו and 5 הינו as AUX). They differ in the subject: HTB labels it nsubj:cop (360), IAHLT plain nsubj; IAHLT also adds Polarity=Pos to the pronoun. The guideline page remarks that זה as a copula ("chocolate זה tasty") is probably modelled on Slavic to/èto — a claim we report, not verify.

Example (IAHLTwiki, iahltwiki_mammography-58):

# text = השיטה הדיגיטלית היא המקובלת כיום.
2   שיטה      NOUN  7  nsubj
4   דיגיטלית  ADJ   2  amod
5   היא       PRON  7  cop    Gender=Fem|Number=Sing|Person=3|Polarity=Pos
7   מקובלת    ADJ   0  root
8   כיום      ADV   7  advmod

ha-shitá ha-digitálit hi ha-mekubélet kayom "The digital method is the accepted one today." The copula היא agrees with the subject (feminine singular). It is near-obligatory here because both sides are definite: without it, השיטה הדיגיטלית המקובלת would read as one noun phrase "the accepted digital method" (07 Syntax).

11.7.3 Existentials and possession: יש / אין

יש yesh "there is" and אין eyn "there is not" are tagged VERB (not AUX) and head their clause; the existing thing is nsubj in IAHLT. In the past and future the existential is a form of היה (היה, היתה/הייתה, יהיה), also VERB. HTB marks all of them HebExistential=Yes (462 in train); IAHLT marks Polarity and adds a construction label in MISC (§ 11.7.7).

Possession is the existential with ל: יש לי ספר yesh li séfer "I have a book" (lit. "there is to-me a book"). The possessor is obl (ל + pronoun). The guideline page notes that the possessed is grammatically a subject in formal Hebrew but behaves like an object in the colloquial יש לי את הספר "I have the book" (07 Syntax).

The existential sentence often has the order V-obl-S ("there-is to-him the-skill…"; guidelines, Word order).

Example (IAHLTwiki, iahltwiki_glasgow-coma-scale-13):

# text = ישנן ארבע רמות:
1  ישנן   יש    VERB  0  root    Gender=Fem|Number=Plur|Person=3|Tense=Pres   Cxn=Existential-ExistPred-VblPart
2  ארבע   ארבע  NUM   3  nummod
3  רמות   רמה   NOUN  1  nsubj

yeshnan arba ramot "There are four levels:". ישנן is the formal inflected יש (lemma יש, feminine plural to agree with רמות) — a form that exists for יש but not in colloquial speech (07).

The IAHLTknesset card sentence 13_ptm_532410-8 shows אין as root with נבחר ציבור "elected official" as nsubj, and a relative clause שיש לו סמכויות "who has powers" with יש as acl:relcl.

11.7.4 Objects and את

The object marker את precedes definite direct objects (07, differential object marking). It is ADP with Case=Acc and depends on the object noun: case:acc in HTB (1 946 in all splits, plus 41 plain case in train), plain case in IAHLT. With a pronoun it fuses: אותי oti "me", אותו oto "him" — split as את + pronoun (HTB את_ + _הוא).

A peculiarity of IAHLTwiki train: 568 standalone את carry Case=Acc, but 622 carry no feature at all. The feature is therefore not a reliable cue in that treebank; the lemma is.

Inflected verbs with an object suffix (ראיתיה "I saw her", הרגו "he killed him") are literary. IAHLT splits them into verb + pronoun with no inserted את.

11.7.5 Relative clauses with ש

The relative marker ש is SCONJ, a prefix on the first word of the clause, and depends on the clause's predicate with mark; the clause attaches to its noun with acl:relcl (2 969 HTB, 3 065 IAHLTwiki, 1 403 IAHLTknesset). The formal אשר is a separate word with the same analysis. A resumptive pronoun inside the clause (שבו "in which", lit. "that in-it") gets its normal relation (obl with its case). Free relatives (מי ש "whoever", מה ש "what") are headed by מי/מה.

The guideline example for a participial subject — המגדל נחש בביתו "whoever raises a snake at home" (literally "the raising a snake in his house") — shows the article ה used as a relative marker on a participle, a construction from earlier Hebrew still used in formal writing (10).

11.7.6 Questions and negation

• Yes/no questions in speech are marked only by intonation; formal writing adds האם. HTB labels האם mark:q (41); IAHLT gives it plain mark. Embedded questions use the same האם ("whether").
• Wh-words are PRON (מי, מה), ADV (איפה, מתי, למה, איך) or DET (איזה, כמה) with PronType=Int.
• Negation: לא is ADV with advmod; אין as an existential is VERB; אינו as a negative copula is AUX (cop) and as a negator of a participle (אינו יודע "does not know") is AUX (aux, 141 in IAHLT). The prohibitive אל + future (אל תדאג "don't worry") is ADV advmod.

11.7.7 Construction annotation (Cxn) in IAHLT

IAHLT adds labels from the UD construction layer (UCxn) in MISC: Cxn= on the head and CxnElt= on the parts. Counts over both IAHLT treebanks (2026-10-17):

VblPart means the existential is the particle יש/אין; FullVerb means a form of היה or a verb like קיים "exist"; CopPred means the existential is expressed by a copula. NPN covers patterns like יום יום "day by day" or פנים אל פנים "face to face". These labels are exactly the kind of enrichment the Mova dialect keeps in MISC (The Mova dialect: a UD metadialect, treebanks, and converters).

11.7.8 Named entities

Both IAHLT treebanks carry nested named entities in MISC (Entity=(PER…PER)) with twelve types: ANG (language), DUC (product), EVE, FAC, GPE, LOC, ORG, PER, TIMEX, TTL (title), WOA (work of art), MISC (treebank READMEs). The IAHLTwiki README example (iahltwiki_piskat-hitgabrut-27) nests an ORG בית המשפט העליון "the Supreme Court" inside a TTL נשיא בית המשפט העליון "President of the Supreme Court".

11.8 The treebanks

Cards with the full metadata are in the Mova repo at treebanks/he/.

HTB. The historical benchmark. Its weaknesses are age (pre-internet: Zeldes et al. 2022 count about a dozen mentions of "computer" and none of mobile phones), a single newspaper's house style (היתה, never the common spelling הייתה "was"), and conversion errors flagged in MISC as HebSource=ConvUncertainHead|ConvUncertainLabel (guidelines). For Mova it is NC: tests and measurements only, never in what is built (Data and model licences).

IAHLTwiki. The first Hebrew treebank annotated natively in UD. Inter-annotator agreement on ~440 double-annotated sentences: word segmentation 99.1%, Cohen's κ lemma 98.5, UPOS 94.8, FEATS 90.9, head 95.8, deprel 95.7 (Zeldes et al. 2022: Table 3). Annotated in Arborator-Grew and continuously checked by grewv, a validator of Grew patterns written as Hebrew-specific rules.

IAHLTknesset. Spoken, parliamentary Hebrew: speaker turns with # speaker, # turnnumber, # is_chairman. Drawn from the Knesset Corpus (Goldin et al. 2024), which itself holds tens of millions of sentences of proceedings since 1992 with automatic morphology. It is the only UD Hebrew corpus with conversational features: vocatives, interjections, code-switching (the card's example has English balance as NOUN in a Hebrew sentence).

PostRab. Not Modern Hebrew: legal, exegetical and other texts from the medieval and early-modern periods, with adaptations "to account for differences in orthography, morphology, and syntax relative to modern Hebrew" (treebank README). The paper reports over 11 800 words (Tal et al. 2025); UD 2.18 holds 8 029 words as a test set. It is relevant here for two reasons: it shows where Modern Hebrew parsers break on older Hebrew (10), and it uses compound:svc and compound:redup which the modern treebanks do not.

Not in UD. The Ancient Hebrew treebank PTNK (hbo, Biblical Hebrew) is a separate language in UD; see the Ancient Hebrew manual (../hbo/, if present) and tools/macula in the Mova repo. Earlier Hebrew dependency resources (Goldberg and Elhadad 2009's scheme; the SPMRL 2013/2014 shared-task Hebrew data) are not UD.

11.9 How the counts were made

All counts above were made on 2026-10-17 from the UD 2.18 CoNLL-U files of the four treebanks (UD_Hebrew-HTB, -IAHLTwiki, -IAHLTknesset, -PostRab), by streaming the files and counting word lines (integer ids), range lines (MWTs), and the UPOS, FEATS and DEPREL columns. "All splits" means train + dev + test; "train" means the train file only. Token and sentence totals come from the treebank cards, which check their counts against each release's stats.xml. The counts are simple, deterministic tallies; a recount on the same files gives the same numbers.

11.10 Parsing Hebrew: the state of the art

End-to-end scores from plain text (gold sentence splits), from Zeldes et al. (2022: Table 4); MWT F1 is segmentation, the others are aligned accuracy:

Out of domain (HebPipe), LAS drops by roughly 3 to 4.5 points; a joint model trained on both corpora scores 88.29 LAS on HTB and 90.36 on Wiki (Table 6). The most confused labels across corpora are NOUN/PROPN (the name convention above) and obl/nmod (prepositional attachment) (Table 7).

The segmentation leap comes from combining a character-level classifier (RFTokenizer, Zeldes 2018: each character either begins a new word or not) with a word-level transformer that predicts whether a word has prefixes and suffixes. A further cue is the position of a word inside its MWT (B/I/E/S): the letter ו is "and" at the start of an MWT and "his" at its end (Zeldes et al. 2022: § 4).

Pretrained Hebrew encoders: AlephBERT (Seker et al. 2022), DictaBERT (Shmidman et al. 2023), and the generative DictaLM (Shmidman et al. 2024). Dicta also built Nakdan, the diacritizer (Shmidman et al. 2020). The rule-based and lexicon-based tradition before them: the MILA morphological analyser and lexicon (Itai and Wintner 2008; Itai et al. 2006), Hspell's analyser (Har'El and Kenigsberg, Hspell), the HMM disambiguator of Adler and Elhadad (2006), and the YAP joint parser (More et al. 2019).

11.11 Pitfalls for a converter

A list for whoever writes a Hebrew reader or converter for the Mova dialect:

1. Know the scheme. Old-HTB files have inserted ה_, _של_, את_ and underscored pronoun forms; IAHLT files never insert. Detect by the presence of underscores in FORM or case:acc / case:gen.
2. Old → new is deterministic (Zeldes et al. 2022): drop an inserted ה_ and put Definite=Def|PronType=Art on the preposition; drop an inserted _של_ or את_ and keep the pronoun as the dependent (nmod:poss / obj); strip the underscores.
3. New → old must re-insert the article from the feature, and the של/את from the relation of the suffix pronoun.
4. Subtypes are predictable: case:acc ⇔ lemma את; case:gen ⇔ lemma של; mark:q ⇔ lemma האם; compound:smixut ⇔ compound whose head has Definite=Cons. flat:name ⇔ flat between PROPNs.
5. The copular subject is nsubj:cop in HTB (only some: 360 of over 1 400 copulas) and nsubj in IAHLT; the pronominal copula is PRON in both.
6. Present tense is unmarked in HTB, Tense=Pres in IAHLT.
7. Names: HTB tags transparent place names NOUN; IAHLT PROPN. A converter cannot fix that by rule: it needs a gazetteer.
8. Spelling: HTB follows Ha'aretz house spelling (closer to ktiv haser, e.g. היתה), IAHLT has free web spelling (הייתה). A lexicon keyed on the spelling must hold both (02).
9. Right-to-left text in a left-to-right file: CoNLL-U stores Hebrew in logical order; viewers show it reversed or mixed with digits and Latin. Never reverse strings when reading; compare by code points. Punctuation like the gershayim ״ is often typed as ASCII " (צה"ל), which collides with quotation marks in TSV and JSON.
10. Licences: HTB is NC (tests only); IAHLT and PostRab are CC BY-SA (quote, test and learn from; do not carry into a release as modified data).

References

• Adler, Meni and Michael Elhadad. 2006. An Unsupervised Morpheme-Based HMM for Hebrew Morphological Disambiguation. ACL 2006. aclanthology.org/P06-1084 our card — local: aclbulk-2006-adler-unsupervised-morpheme-based-hmm-hebrew.
• Cohen, Shay B. and Noah A. Smith. 2007. Joint Morphological and Syntactic Disambiguation. EMNLP-CoNLL 2007. aclanthology.org/D07-1022 — local: aclbulk-2007-cohen-joint-morphological-syntactic-disambiguation.
• Goldberg, Yoav and Michael Elhadad. 2009. Hebrew Dependency Parsing: Initial Results. IWPT 2009. aclanthology.org/W09-3819 our card — local: lingbulk-2009-goldberg-hebrew-dependency-parsing-initial-results.
• Goldberg, Yoav and Michael Elhadad. 2011. Joint Hebrew Segmentation and Parsing using a PCFGLA Lattice Parser. ACL 2011. aclanthology.org/P11-2124 our card — local: aclbulk-2011-goldberg-joint-hebrew-segmentation-parsing-pcfgla.
• Goldberg, Yoav and Reut Tsarfaty. 2008. A Single Generative Model for Joint Morphological Segmentation and Syntactic Parsing. ACL 2008. aclanthology.org/P08-1043 — local: aclbulk-2008-goldberg-single-generative-model-joint-morphological.
• Goldin, Gili, Nick Howell, Noam Ordan, Ella Rabinovich and Shuly Wintner. 2024. The Knesset Corpus: An Annotated Corpus of Hebrew Parliamentary Proceedings. arXiv:2405.18115. arxiv.org/abs/2405.18115 our card — local: arxivrecent-2024-goldin-knesset-corpus-annotated-corpus-hebrew.
• Itai, Alon and Shuly Wintner. 2008. Language Resources for Hebrew. Language Resources and Evaluation 42: 75–98. (Not held; the MILA resources are described in Wintner 2003/2004, local: lingbulk-2003-wintner-resources-processing-israeli-hebrew.)
• Itai, Alon, Shuly Wintner and Shlomo Yona. 2006. A Computational Lexicon of Contemporary Hebrew. LREC 2006. Local: lingbulk-2006-itai-computational-lexicon-contemporary-hebrew.
• More, Amir, Amit Seker, Victoria Basmova and Reut Tsarfaty. 2019. Joint Transition-Based Models for Morpho-Syntactic Parsing: Parsing Strategies for MRLs and a Case Study from Modern Hebrew. TACL 7: 33–48. aclanthology.org/Q19-1003 our card (not held locally).
• Sade, Shoval, Amit Seker and Reut Tsarfaty. 2018. The Hebrew Universal Dependency Treebank: Past, Present and Future. UDW 2018. aclanthology.org/W18-6016 our card (not held locally).
• Seker, Amit, Elron Bandel, Dan Bareket, Idan Brusilovsky, Refael Shaked Greenfeld and Reut Tsarfaty. 2022. AlephBERT: Language Model Pre-training and Evaluation from Sub-Word to Sentence Level. ACL 2022. aclanthology.org/2022.acl-long.4 our card — local: acl-2022-seker-alephbert-language-model-pre-training.
• Shmidman, Avi, Shaltiel Shmidman, Moshe Koppel and Yoav Goldberg. 2020. Nakdan: Professional Hebrew Diacritizer. ACL 2020 demos. aclanthology.org/2020.acl-demos.23 our card — local: aclbulk-2020-shmidman-nakdan-professional-hebrew-diacritizer.
• Shmidman, Shaltiel, Avi Shmidman and Moshe Koppel. 2023. DictaBERT: A State-of-the-Art BERT Suite for Modern Hebrew. arXiv:2308.16687 — local: arxivrecent-2023-shmidman-dictabert-state-art-bert-suite.
• Shmidman, Shaltiel et al. 2024. Adapting LLMs to Hebrew: Unveiling DictaLM 2.0. arXiv:2407.07080 — local: arxivrecent-2024-shmidman-adapting-llms-hebrew-unveiling-dictalm.
• Sima'an, Khalil, Alon Itai, Yoad Winter, Alon Altman and Noa Nativ. 2001. Building a Tree-Bank of Modern Hebrew Text. Traitement Automatique des Langues 42(2). (Not held.)
• Tal, Rachel, Shlomit Fuchs, Orly Albeck, Elisheva Brauner, Avi Shmidman, Yitzchak Lindenbaum and Ephraim Meiri. 2025. A New Hebrew Universal Dependency Treebank: The First Treebank of Post-Rabbinic Historical Hebrew. TLT 2025. aclanthology.org/2025.tlt-1.11 our card
• Tsarfaty, Reut. 2006. Integrated Morphological and Syntactic Disambiguation for Modern Hebrew. COLING-ACL 2006 Student Research Workshop. aclanthology.org/P06-3009 our card — local: aclbulk-2006-tsarfaty-integrated-morphological-syntactic-disambiguation-modern.
• Tsarfaty, Reut. 2013. A Unified Morpho-Syntactic Scheme of Stanford Dependencies. ACL 2013. aclanthology.org/P13-2103
• Tsarfaty, Reut, Shoval Sadde, Stav Klein and Amit Seker. 2019. What's Wrong with Hebrew NLP? And How to Make it Right. EMNLP-IJCNLP 2019 demos. aclanthology.org/D19-3044 our card — local: aclbulk-2019-tsarfaty-what-s-wrong-hebrew-nlp.
• Tsarfaty, Reut, Djamé Seddah, Sandra Kübler and Joakim Nivre. 2013. Parsing Morphologically Rich Languages: Introduction to the Special Issue. Computational Linguistics 39(1). aclanthology.org/J13-1003 — local: cl-2013-tsarfaty-parsing-morphologically-rich-languages-introduction.
• UD Hebrew documentation: language page and the pages for compound:smixut, case:acc, case:gen, nsubj:cop, mark:q, compound:affix, HebBinyan, HebExistential, Prefix. universaldependencies.org/he — local snapshot (commit 3ae7161, 2026-10-17): _he.
• UD 2.18 treebanks UD_Hebrew-HTB, -IAHLTwiki, -IAHLTknesset, -PostRab: github.com/UniversalDependencies — local: ud-2.18 (HTB under nc/); Mova cards: treebanks/he/.
• Zeldes, Amir. 2018. A Characterwise Windowed Approach to Hebrew Morphological Segmentation. SIGMORPHON 2018. aclanthology.org/W18-5811 our card — local: aclbulk-2018-zeldes-characterwise-windowed-approach-hebrew-morphological.
• Zeldes, Amir, Nick Howell, Noam Ordan and Yifat Ben Moshe. 2022. A Second Wave of UD Hebrew Treebanking and Cross-Domain Parsing. EMNLP 2022, 4331–4344. aclanthology.org/2022.emnlp-main.292 our card

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/he/11-ud-analysis/