scienceclaw-benchmark-for47

Universal Dependencies parsing with gold tokenization

Task. Input: annotated sentences {"words": [{"form", "upos", "feats", "head", "deprel", ...}]} (head 1-based, 0 = root) and target sentences with id and form only. Deliverable: one dict per target sentence, {"head": [int] * n_words, "deprel": [str] * n_words}, in word order. Each tree needs exactly one root, heads in range, no cycle, and relations from the 37 UD v2 relations (udparse.UD_RELATIONS).

Quality. LAS (CoNLL 2018 evaluation, gold tokenization), higher is better: share of words with correct head and universal relation (deprel before :), punctuation included, micro over all words; UAS is auxiliary. Weak baseline: right-branching chain with each word's most frequent relation.

Tools.

• from scilib import udparse: UDParser(decode="eisner"|"cle", epochs, n_models, jackknife_folds, use_morph, seed).fit(train) with .tag, .parse, .arc_scores (operator ud_parser_train_parse); fit_predict(train, [targets_a, targets_b]) fits once, parses every list; cross_validate; las_uas(gold, parses) (ud_attachment_scores); check_parse(parse, n_words) (empty = valid); eisner, chu_liu_edmonds (dependency_tree_decode). Needs only target word forms; ~20-30 CPU s to fit 1000 sentences.
• Pretrained French: fit_predict(train, targets, pretrained=True) or udparse_pretrained.parse_gold_tokens(sentences) (french_dependency_parse_stanza; Stanza French-GSD on a CamemBERT-large encoder; train unused). Weights camembert_large and stanza_fr; needs torch, transformers, stanza or a GPU worker. Check udparse_pretrained.available() and scienceclaw_tools(operation=weights); nothing downloads implicitly.

Routes. Trained parser on the visible treebank (any language), selected on held-out annotated sentences or cross_validate. For French the pretrained parser is an alternative; compare on held-out sentences outside its training data.

Rules.

• Disclose overlap: the pretrained models were trained on the UD French-GSD train file and selected on its dev file (other treebanks, e.g. French-PUD, unseen); do not judge them on those sentences. Their scheme follows a newer UD release than 2.2 in places.
• Use the given words: no re-tokenisation, no multiword splitting.
• Pretrained labels may carry subtypes (nsubj:pass); strip them if only universal relations are allowed.
• Run check_parse on every output; never train on target trees.