#acl AlinaWroblewska:read,write,revert All:read = COMBO's models for Polish = [[https://gitlab.clarin-pl.eu/syntactic-tools/combo/-/tree/master|COMBO's]] models for Polish trained on the current version of [[http://zil.ipipan.waw.pl/PDB|Polish Dependency Bank]] using the [[https://huggingface.co/allegro/herbert-base-cased|HerBERT]] language model. === PDB-trained models === * [[http://mozart.ipipan.waw.pl/~alina/Polish_dependency_parsing_models/COMBO_pytorch/combo_PDB_parseonly_220906.tar.gz|model]] for dependency parsing only * [[http://mozart.ipipan.waw.pl/~alina/Polish_dependency_parsing_models/COMBO_pytorch/combo_PDB_full_220906.tar.gz|model]] for part-of-speech tagging, morphological analysis, lemmatisation, and dependency parsing (dependency relation types '''without''' semantic extensions, e.g. adjunct instead of adjunct_temp) * [[http://mozart.ipipan.waw.pl/~alina/Polish_dependency_parsing_models/COMBO_pytorch/combo_PDB_full_SEMLAB_220906.tar.gz|model]] for part-of-speech tagging, morphological analysis, lemmatisation, and dependency parsing (dependency relation types '''with''' semantic extensions, e.g. adjunct_temp) === PDB-UD-trained model === * [[http://mozart.ipipan.waw.pl/~alina/Polish_dependency_parsing_models/COMBO_pytorch/combo_PDBUD_full_220906.tar.gz|model]] for part-of-speech tagging, morphological analysis, lemmatisation, and dependency parsing {{{#!wiki comment [[https://github.com/360er0/COMBO|COMBO]], [[https://code.google.com/archive/p/mate-tools/|MateParser]] and [[http://maltparser.org|MaltParser]]. /* ''MaltParser'' is a transition-based dependency parser that uses a deterministic parsing algorithm. The deterministic parsing algorithm builds a dependency structure of an input sentence based on transitions (shift-reduce actions) predicted by a classifier. The classifier learns to predict the next transition given training data and the parse history. `MateParser`, in turn, is a graph-based parser that defines a space of well-formed candidate dependency trees for an input sentence, scores them given an induced parsing model, and selects the highest scoring dependency tree as a correct analysis of the input sentence. */ * [[http://mozart.ipipan.waw.pl/~alina/Polish_dependency_parsing_models/COMBO_pytorch/combo_PDB_parseonly_220906.tar.gz|COMBO-pytorch model]] for dependency parsing only (with [[https://huggingface.co/allegro/herbert-base-cased|HerBERT-base]] embeddings), * [[http://mozart.ipipan.waw.pl/~alina/Polish_dependency_parsing_models/COMBO/20200930_COMBO_PDB_nosem_parseonly.pkl|COMBO model]] for dependency parsing only * [[http://mozart.ipipan.waw.pl/~alina/Polish_dependency_parsing_models/COMBO/20200930_COMBO_PDB_nosem.pkl|COMBO model]] for part-of-speech tagging, lemmatisation, and dependency parsing * [[http://mozart.ipipan.waw.pl/~alina/Polish_dependency_parsing_models/COMBO/20200930_COMBO_PDB_sem.pkl|COMBO model]] for part-of-speech tagging, lemmatisation, dependency parsing, and semantic role labelling * [[http://mozart.ipipan.waw.pl/~alina/Polish_dependency_parsing_models/COMBO/191107_COMBO_PDB_semlab_parseonly.pkl|COMBO model]] for (semantic) dependency parsing only * [[http://mozart.ipipan.waw.pl/~alina/Polish_dependency_parsing_models/MATE/20190612_MATE_PDB.pkl|MATE model]] for dependency parsing * [[http://mozart.ipipan.waw.pl/~alina/Polish_dependency_parsing_models/MALT/190125_MALT_PDB.mco|MaltParser model]] for dependency parsing == PDB-UD-trained dependency parsing models for Polish == The PDB-UD-based models are trained on the current version of [[http://git.nlp.ipipan.waw.pl/alina/PDBUD|Polish Dependency Bank in Universal Dependencies format]] with the publicly available parsing systems – [[https://gitlab.clarin-pl.eu/syntactic-tools/combo/-/tree/master|COMBO-pytorch]], [[https://github.com/360er0/COMBO|COMBO]], [[http://ufal.mff.cuni.cz/udpipe|UDPipe]]. * [[http://mozart.ipipan.waw.pl/~mklimaszewski/models/polish-herbert-base.tar.gz|COMBO-pytorch model]] for for part-of-speech tagging, lemmatisation, and dependency parsing (with [[https://huggingface.co/allegro/herbert-base-cased|HerBERT-base]] embeddings), * [[http://mozart.ipipan.waw.pl/~mklimaszewski/models/polish-herbert-large.tar.gz|COMBO-pytorch model]] for for part-of-speech tagging, lemmatisation, and dependency parsing (with [[https://huggingface.co/allegro/herbert-large-cased|HerBERT-large]] embeddings), * [[http://mozart.ipipan.waw.pl/~mklimaszewski/models/polish-ud27.tar.gz|COMBO-pytorch model]] for for part-of-speech tagging, lemmatisation, and dependency parsing (with fastText embeddings), * [[http://mozart.ipipan.waw.pl/~alina/Polish_dependency_parsing_models/COMBO/20200930_COMBO_PDBUD_nosem.pkl|COMBO model]] for part-of-speech tagging, lemmatisation, and dependency parsing * [[http://mozart.ipipan.waw.pl/~alina/Polish_dependency_parsing_models/COMBO/20200930_COMBO_PDBUD_sem.pkl|COMBO model]] for part-of-speech tagging, lemmatisation, dependency parsing, and semantic role labelling * [[http://mozart.ipipan.waw.pl/~alina/Polish_dependency_parsing_models/UDPIPE/20200930_PDBUD_ttp_embedd.udpipe|UDPipe model]] for tokenisation, part-of-speech tagging, lemmatisation, and dependency parsing * [[http://mozart.ipipan.waw.pl/~alina/Polish_dependency_parsing_models/UDPIPE/20200930_PDBUD_tokeniser.udpipe|UDPipe model]] for tokenisation}}} {{{#!wiki comment * [[http://mozart.ipipan.waw.pl/~prybak/model_poleval2018/model_A_semi.pkl|COMBO]] model for Polish (the model estimated for the [[http://poleval.pl/tasks#task1|PolEval 2018]] competition) * [[attachment:180606_PDBUDPipe.udpipe|UDPipe]] model for Polish}}} === COMBO === * COMBO's [[https://gitlab.clarin-pl.eu/syntactic-tools/combo/-/tree/master|source code]] * Beginner's [[https://colab.research.google.com/drive/1D1P4AiE40Cc_4SF3HY-Mz06JY0XMiEFs?hl=en|tutorial]] (collab notebook) * COMBO's [[https://gitlab.clarin-pl.eu/syntactic-tools/combo/-/blob/master/docs/performance.md|performance]] on test sets for multiple languages from [[https://universaldependencies.org|Universal Dependencies]] * Web demos * [[http://combo-demo.nlp.ipipan.waw.pl/combo-eng|English]] * [[http://combo-demo.nlp.ipipan.waw.pl/combo-pl|Polish]] {{{#!wiki comment === Parsing performance === See [[http://clip.ipipan.waw.pl/benchmarks#Dependency_parsing|Dependency parsing]] section. * [[attachment:190115_COMBO_PDB_nosem.pkl]] – PDB-based COMBO model for part-of-speech tagging, lemmatisation, and dependency parsing * [[attachment: 190115_COMBO_PDB_sem.pkl]] – PDB-based COMBO model for part-of-speech tagging, lemmatisation, dependency parsing and semantic role labelling * '''NEW!''' PDB-based COMBO model compatible with the tagset of Morfeusz 2: [[attachment:180912_PDBCOMBO.pkl]] * MateParser * '''NEW!''' PDB-based Mate model compatible with the tagset of Morfeusz 2: [[attachment:180322_PDBMate.mdl]] * PDB-based Mate model compatible with the tagset of Morfeusz: [[attachment:170608_PDBMate.mdl]] * MateParser * [[attachment:190125_MATE_PDB.model]] – PDB-based MateParser model for dependency parsing * MaltParser * [[attachment:190125_MALT_PDB.mco]] – PDB-based MaltParser model for dependency parsing * '''NEW!''' PDB-based MaltParser model compatible with the tagset of Morfeusz 2: [[attachment:180322_PDBMalt.mco]] * PDB-basd MaltParser model compatible with the tagset of Morfeusz: [[attachment:170608_PDBMalt.mco]] === 10-fold cross-validation (avg.) === || '''Model''' || '''LAS''' || '''UAS''' || || `PDBMate` || 0.85 || 0.89 || || `PDBMalt` || 0.82 || 0.86 || === Precision, recall and f-score of individual dependency relations (avg.) === The description of Polish dependency relations types is available on [[http://zil.ipipan.waw.pl/PDB/DepRelTypes|Polish dependency relation types]]. || '''Dependency relation type''' |||| '''Precision''' |||| '''Recall''' |||| '''F-Measure''' || || Mate || Malt || Mate || Malt || Mate || Malt || ||abbrev_punct ||0.99 ||0.99 ||0.98 ||0.97 || 0.98 ||0.98 || ||adjunct || 0.89 || 0.73 || 0.92 || 0.77 || 0.82 || 0.75 || ||adjunct_qt || 0.74 || 0.51 || 0.76 || 0.58 || 0.75 || 0.55 || ||aglt || 1.00 || 0.98 || 1.00 || 0.98 || 0.98 || 0.98 || ||app || 0.75 || 0.58 || 0.69 || 0.52 || 0.72 || 0.55 || ||aux || 0.95 || 0.90 || 0.97 || 0.92 || 0.96 || 0.91 || ||comp || 0.90 || 0.85 || 0.87 || 0.82 || 0.88 || 0.84 || ||comp_ag || 0.95 || 0.90 || 0.96 || 0.91 || 0.94 || 0.90 || ||comp_fin || 0.87 || 0.75 || 0.86 || 0.79 || 0.87 || 0.77 || ||comp_inf || 0.95 || 0.91 || 0.96 || 0.90 || 0.93 || 0.90 || || cond || 1.00 || 0.97 || 1.00 || 0.96 || 1.00 || 0.96 || ||conjunct || 0.85 || 0.71 || 0.82 || 0.65 || 0.82 || 0.68 || || imp ||0.98 || 0.97 || 0.91 || 0.87 ||0.94 || 0.92 || || item || 0.87 || 0.4 || 0.73 || 0.37 || 0.61 || 0.39 || || mwe || 0.90 || 0.83 || 0.83 || 0.75 ||0.87 || 0.79 || ||ne || 0.87 || 0.78 || 0.73 || 0.64 || 0.76 || 0.70 || || neg ||0.99 || 0.97 || 1.00 || 0.98 ||0.99 || 0.98 || ||obj || 0.89 || 0.81 || 0.91 || 0.86 || 0.89 || 0.83 || || obj_th ||0.83 || 0.76 || 0.76 || 0.65 ||0.80 || 0.70 || ||pd || 0.86 || 0.77 || 0.80 || 0.72 || 0.87 || 0.74 || || pre_coord ||0.86 || 0.76 ||0.78 || 0.55 ||0.82 || 0.64 || ||punct || 0.97 || 0.75 || 0.98 || 0.76 || 0.88 || 0.76 || ||refl || 0.99 || 0.96 || 0.99 || 0.96 || 0.99 || 0.96 || ||root || 0.91 || 0.80 || 0.91 || 0.81 || 0.94 ||0.80 || || subj || 0.94|| 0.84 ||0.94 || 0.83 ||0.94 || 0.84 || === COMBO demos === * [[http://combo-demo.nlp.ipipan.waw.pl/combo-eng|English]] * [[http://combo-demo.nlp.ipipan.waw.pl/combo-pl|Polish]] * [[http://scwad-demo.nlp.ipipan.waw.pl:8000/dependency-parsing|COMBO demo]] * [[http://multiservice.nlp.ipipan.waw.pl|MaltParser demo in Multiservice NLP]] * To parse a Polish text in Multiservice "Select predefined chain of actions": 5: Concraft, !DependencyParser, input your text, and press the button "Run". * To download the parser's output in CoNLL format, "Select output format:".}}} === Publications === <> <> <> {{{#!wiki comment <> <>}}} === Licensing === Polish NLP models are released under the [[https://creativecommons.org/licenses/by-nc-sa/4.0/|CC BY-NC-SA 4.0]] licence and by downloading them you accept the conditions of that licence. === Acknowledgment === The research was founded by SONATA 8 grant no 2014/15/D/HS2/03486 from the National Science Centre Poland and by the Polish Ministry of Science, Higher Education as part of the investment in the CLARIN-PL research infrastructure and by Digital Research Infrastructure for the Arts and Humanities DARIAH-PL. The computing was performed at Poznań Supercomputing and Networking Center. === Contact === Any questions, comments? Please send them to <>.