Amit Seker

8 paper cards in Mova's library · 2018–2022

Amit Seker wrote 8 of the papers Mova reads, in 2018–2022, most of them on Dependency parsing, Treebanks, Named entity recognition; often with Reut Tsarfaty, Dan Bareket, Amir More.

OpenAlex · h-index 7 · 185 citations · 13 works — openalex, 2026-10-18

The profile is the library's match of the name; two people with one name are not always told apart.

Elsewhere: Google Scholar · Semantic Scholar · dblp · ACL Anthology · ORCID

Topics: Dependency parsing 2 · Treebanks 2 · Named entity recognition 1

Ideas Mova took (6)

From the ideas ledger: an idea tried, kept or queued, or measured with no gain, and what it gave.

kept — Turn the morpheme head into a deterministic check: a segmenter's output for a token must concatenate back to the surface string. The paper's character-with-boundary view makes this a free invariant to test on every token. (AlephBERT: Language Model Pre-training and Evaluation from… · consonant check on every candidate (Model::spells): 0 failures on dev, test and the two-fold cv of train; kept as a gate over cut(); commit 5c24bc409)
kept — Split every wrong prediction into two kinds: "gold analysis missing from the candidate list" and "gold analysis present but not chosen". (A Pointer Network Architecture for Joint Morphological… · hbo seg eval splits errors into gold missing from the candidates vs present but not chosen, plus oracle@1/2/3/5: unpointed dev 36 missing / 79 not chosen…)
kept — A small choice model over the segmenter's candidates: the tools/hbo tables already give the candidate splits. (A Pointer Network Architecture for Joint Morphological… · choice over the lexicon's candidates by the neighbouring words (ctx:0.5, the lexicon share smoothed toward the counts with the word before, then the word…)
kept — Store the candidates per token as an ordered list, so that a choice is a single index. Log the index with the features that decided it. This matches Mova's "say why" rule and fits the world store. (A Pointer Network Architecture for Joint Morphological… · Model::candidates gives the ordered candidate list; the choice is logged in MISC (Seg= with the neighbour that decided, Lattice=<label>:<gain>); commits…)
measured, no gain — MDFirst as a re-ranker: keep the top-B segmentations from the Hebrew segmenter and let the parser rescore each one with its tree score, then take the best joint result. The first target is the segmenter in tools/hbo. (Joint Transition-Based Models for Morpho-Syntactic Parsing… · top-k segmentations rescored by gparse's tree score (hbo lattice/choose): no gain beyond noise on Chronicles (unpointed 79.73 -> 79.58..79.99, pointed 85.24…)
measured, no gain — Keep the lattice: let tools/hbo output every segmentation candidate of a token, with its tag and lemma per arc, and pass the whole lattice to gparse, so that the parser can choose the segmentation that fits the tree. (What's Wrong with Hebrew NLP? And How to Make it Right · lattice in tools/hbo (hbo lattice / gparse GP_SCORE / hbo choose): the tree score does not pick the gold split (gold candidate +0.13 mean tree gain vs -0.41…)

Papers

AlephBERT: Language Model Pre-training and Evaluation from Sub-Word to Sentence Level (2022) 1 idea kept
Seker et al. pre-train AlephBERT, a BERT-style model for Modern Hebrew, on about 1.9 billion words: the deduplicated Hebrew part of OSCAR, all of Hebrew Wikipedia, and Hebrew tweets, with a 52K wordpiece vocabulary. The base model trained for 10 epochs and took 8 days on eight V100 GPUs. The paper's main point is…
Large Pre-Trained Models with Extra-Large Vocabularies: A Contrastive Analysis of Hebrew BERT Models and a New One to Outperform Them All (2022)
The authors train AlephBERTGimmel, a Modern Hebrew BERT with a 128K word-piece vocabulary (earlier models had 2.4K–52K), on the same ~2 billion words as AlephBERT.
AlephBERT: A Hebrew Large Pre-Trained Language Model to Start-off your Hebrew NLP Application With (2021)
The authors build AlephBERT, a BERT-base style language model for Modern Hebrew, pre-trained on about 1.9 billion words: deduplicated OSCAR, a large set of Hebrew tweets (72% of sentences), and Hebrew Wikipedia.
A Pointer Network Architecture for Joint Morphological Segmentation and Tagging (2020) 3 ideas kept
Morphological disambiguation (MD) picks one decomposition of each token into morphemes. The authors feed a symbolic lattice of all analyses per token (HebLex for Hebrew, TRMorph2 for Turkish) into a neural pointer network. The lattice is flattened into a sequence of analyses, each analysis is embedded by averaging…
Joint Transition-Based Models for Morpho-Syntactic Parsing: Parsing Strategies for MRLs and a Case Study from Modern Hebrew (2019)
In Hebrew, morphological analysis and disambiguation (MA&D) normally comes before parsing, yet the syntax would help choose the morphology, and the reverse.
What's Wrong with Hebrew NLP? And How to Make it Right (2019)
The paper presents ONLP (YAP), an open system for Modern Hebrew that does segmentation, POS tagging, lemmas, features and labelled dependencies in one package.
The Hebrew Universal Dependency Treebank: Past, Present and Future (2018)
The Hebrew Treebank (HTB) grew from 500 newspaper sentences (2001) to 6221 sentences, through a chain of automatic and semi-automatic conversions: phrase structure, then Goldberg's functional-head dependencies, then Unified-Stanford, then UD v1 and UD v2.
Universal Morpho-Syntactic Parsing and the Contribution of Lexica: Analyzing the ONLP Lab Submission to the CoNLL 2018 Shared Task (2018)
The ONLP Lab submission used UDPipe for segmentation and morphological disambiguation, then yap's arc-eager dependency parser (a Zhang–Nivre variant with morphological-signature features).

Co-authors (17)

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/authors/amit-seker/