README.md

DaCy small transformer

DaCy is a Danish language processing framework with state-of-the-art pipelines as well as functionality for analysing Danish pipelines. DaCy's largest pipeline has achieved State-of-the-Art performance on Named entity recognition, part-of-speech tagging and dependency parsing for Danish on the DaNE dataset. Check out the DaCy repository for material on how to use DaCy and reproduce the results. DaCy also contains guides on usage of the package as well as behavioural test for biases and robustness of Danish NLP pipelines.

Feature	Description
Name	da_dacy_small_trf
Version	0.1.0
spaCy	>=3.1.1,<3.2.0
Default Pipeline	transformer, morphologizer, parser, attribute_ruler, lemmatizer, ner
Components	transformer, morphologizer, parser, attribute_ruler, lemmatizer, ner
Vectors	0 keys, 0 unique vectors (0 dimensions)
Sources	UD Danish DDT v2.5 (Johannsen, Anders; Martínez Alonso, Héctor; Plank, Barbara) DaNE (Rasmus Hvingelby, Amalie B. Pauli, Maria Barrett, Christina Rosted, Lasse M. Lidegaard, Anders Søgaard) Maltehb/-l-ctra-danish-electra-small-cased (Malte Højmark-Bertelsen)
License	Apache-2.0 License
Author	Centre for Humanities Computing Aarhus

Label Scheme

View label scheme (192 labels for 3 components)

Component	Labels
morphologizer	AdpType=Prep|POS=ADP, Definite=Ind|Gender=Com|Number=Sing|POS=NOUN, Mood=Ind|POS=AUX|Tense=Pres|VerbForm=Fin|Voice=Act, POS=PROPN, Definite=Ind|Number=Sing|POS=VERB|Tense=Past|VerbForm=Part, Definite=Def|Gender=Neut|Number=Sing|POS=NOUN, POS=SCONJ, Definite=Def|Gender=Com|Number=Sing|POS=NOUN, Mood=Ind|POS=VERB|Tense=Pres|VerbForm=Fin|Voice=Act, POS=ADV, Number=Plur|POS=DET|PronType=Dem, Degree=Pos|Number=Plur|POS=ADJ, Definite=Ind|Gender=Com|Number=Plur|POS=NOUN, POS=PUNCT, POS=CCONJ, Definite=Ind|Degree=Cmp|Number=Sing|POS=ADJ, Degree=Cmp|POS=ADJ, POS=PRON|PartType=Inf, Gender=Com|Number=Sing|POS=DET|PronType=Ind, Definite=Ind|Degree=Pos|Number=Sing|POS=ADJ, Case=Acc|Gender=Neut|Number=Sing|POS=PRON|Person=3|PronType=Prs, Definite=Ind|Gender=Neut|Number=Plur|POS=NOUN, Definite=Def|Degree=Pos|Number=Sing|POS=ADJ, Gender=Neut|Number=Sing|POS=DET|PronType=Dem, Degree=Pos|POS=ADV, Definite=Def|Number=Sing|POS=VERB|Tense=Past|VerbForm=Part, Definite=Ind|Gender=Neut|Number=Sing|POS=NOUN, POS=PRON|PronType=Dem, NumType=Card|POS=NUM, Definite=Ind|Degree=Pos|Gender=Neut|Number=Sing|POS=ADJ, Case=Acc|Gender=Com|Number=Sing|POS=PRON|Person=3|PronType=Prs, Degree=Pos|Gender=Com|Number=Sing|POS=ADJ, Case=Nom|Gender=Com|Number=Sing|POS=PRON|Person=3|PronType=Prs, NumType=Ord|POS=ADJ, Gender=Com|Number=Sing|Number[psor]=Sing|POS=DET|Person=3|Poss=Yes|PronType=Prs|Reflex=Yes, Mood=Ind|POS=AUX|Tense=Past|VerbForm=Fin|Voice=Act, POS=VERB|VerbForm=Inf|Voice=Act, Mood=Ind|POS=VERB|Tense=Past|VerbForm=Fin|Voice=Act, POS=NOUN, Mood=Ind|POS=VERB|Tense=Pres|VerbForm=Fin|Voice=Pass, POS=ADP|PartType=Inf, Degree=Pos|POS=ADJ, Definite=Def|Gender=Com|Number=Plur|POS=NOUN, Number[psor]=Sing|POS=DET|Person=3|Poss=Yes|PronType=Prs, Case=Gen|Definite=Def|Gender=Com|Number=Sing|POS=NOUN, POS=AUX|VerbForm=Inf|Voice=Act, Definite=Ind|Degree=Pos|Gender=Com|Number=Sing|POS=ADJ, Gender=Com|Number=Sing|POS=DET|PronType=Dem, Number=Plur|POS=DET|PronType=Ind, Gender=Com|Number=Sing|POS=PRON|PronType=Ind, Case=Acc|POS=PRON|Person=3|PronType=Prs|Reflex=Yes, POS=PART|PartType=Inf, Gender=Neut|Number=Sing|POS=DET|PronType=Ind, Case=Acc|Number=Plur|POS=PRON|Person=3|PronType=Prs, Case=Gen|Definite=Def|Gender=Neut|Number=Sing|POS=NOUN, Case=Nom|Number=Plur|POS=PRON|Person=3|PronType=Prs, Case=Nom|Gender=Com|Number=Sing|POS=PRON|Person=1|PronType=Prs, Case=Nom|Gender=Com|POS=PRON|PronType=Ind, Gender=Neut|Number=Sing|POS=PRON|PronType=Ind, Mood=Imp|POS=VERB, Gender=Com|Number=Sing|Number[psor]=Sing|POS=DET|Person=1|Poss=Yes|PronType=Prs, Definite=Ind|Number=Sing|POS=AUX|Tense=Past|VerbForm=Part, POS=X, Case=Nom|Gender=Com|Number=Plur|POS=PRON|Person=1|PronType=Prs, Case=Gen|Definite=Def|Gender=Com|Number=Plur|POS=NOUN, POS=VERB|Tense=Pres|VerbForm=Part, Number=Plur|POS=PRON|PronType=Int,Rel, POS=VERB|VerbForm=Inf|Voice=Pass, Case=Gen|Definite=Ind|Gender=Com|Number=Sing|POS=NOUN, Degree=Cmp|POS=ADV, POS=ADV|PartType=Inf, Degree=Sup|POS=ADV, Number=Plur|POS=PRON|PronType=Dem, Number=Plur|POS=PRON|PronType=Ind, Definite=Def|Gender=Neut|Number=Plur|POS=NOUN, Case=Acc|Gender=Com|Number=Sing|POS=PRON|Person=1|PronType=Prs, Case=Gen|POS=PROPN, POS=ADP, Degree=Cmp|Number=Plur|POS=ADJ, Definite=Def|Degree=Sup|POS=ADJ, Gender=Neut|Number=Sing|Number[psor]=Sing|POS=DET|Person=1|Poss=Yes|PronType=Prs, Degree=Pos|Number=Sing|POS=ADJ, Number=Plur|Number[psor]=Sing|POS=DET|Person=3|Poss=Yes|PronType=Prs|Reflex=Yes, Gender=Com|Number=Sing|Number[psor]=Plur|POS=DET|Person=1|Poss=Yes|PronType=Prs|Style=Form, Number=Plur|POS=PRON|PronType=Rcp, Case=Gen|Degree=Cmp|POS=ADJ, Case=Gen|Definite=Def|Gender=Neut|Number=Plur|POS=NOUN, Number[psor]=Plur|POS=DET|Person=3|Poss=Yes|PronType=Prs, POS=INTJ, Number=Plur|Number[psor]=Sing|POS=DET|Person=1|Poss=Yes|PronType=Prs, Degree=Pos|Gender=Neut|Number=Sing|POS=ADJ, Gender=Neut|Number=Sing|Number[psor]=Plur|POS=DET|Person=1|Poss=Yes|PronType=Prs|Style=Form, Case=Acc|Gender=Com|Number=Sing|POS=PRON|Person=2|PronType=Prs, Gender=Com|Number=Sing|Number[psor]=Sing|POS=DET|Person=2|Poss=Yes|PronType=Prs, Case=Gen|Definite=Ind|Gender=Neut|Number=Plur|POS=NOUN, Number=Sing|POS=PRON|PronType=Int,Rel, Number=Plur|Number[psor]=Plur|POS=DET|Person=1|Poss=Yes|PronType=Prs|Style=Form, Gender=Neut|Number=Sing|POS=PRON|PronType=Int,Rel, Definite=Def|Degree=Sup|Number=Plur|POS=ADJ, Case=Nom|Gender=Com|Number=Sing|POS=PRON|Person=2|PronType=Prs, Gender=Neut|Number=Sing|Number[psor]=Sing|POS=DET|Person=3|Poss=Yes|PronType=Prs|Reflex=Yes, Definite=Ind|Number=Sing|POS=NOUN, Number=Plur|POS=VERB|Tense=Past|VerbForm=Part, Number=Plur|Number[psor]=Sing|POS=PRON|Person=3|Poss=Yes|PronType=Prs|Reflex=Yes, POS=SYM, Case=Nom|Gender=Com|POS=PRON|Person=2|Polite=Form|PronType=Prs, Degree=Sup|POS=ADJ, Number=Plur|POS=DET|PronType=Ind|Style=Arch, Case=Gen|Gender=Com|Number=Sing|POS=DET|PronType=Dem, Foreign=Yes|POS=X, POS=DET|Person=2|Polite=Form|Poss=Yes|PronType=Prs, Gender=Neut|Number=Sing|POS=PRON|PronType=Dem, Case=Acc|Gender=Com|Number=Plur|POS=PRON|Person=1|PronType=Prs, Case=Gen|Definite=Ind|Gender=Neut|Number=Sing|POS=NOUN, Case=Gen|POS=PRON|PronType=Int,Rel, Gender=Com|Number=Sing|POS=PRON|PronType=Dem, Abbr=Yes|POS=X, Case=Gen|Definite=Ind|Gender=Com|Number=Plur|POS=NOUN, Definite=Def|Degree=Abs|POS=ADJ, Definite=Ind|Degree=Sup|Number=Sing|POS=ADJ, Definite=Ind|POS=NOUN, Gender=Com|Number=Plur|POS=NOUN, Number[psor]=Plur|POS=DET|Person=1|Poss=Yes|PronType=Prs, Gender=Com|POS=PRON|PronType=Int,Rel, Case=Nom|Gender=Com|Number=Plur|POS=PRON|Person=2|PronType=Prs, Degree=Abs|POS=ADV, POS=VERB|VerbForm=Ger, POS=VERB|Tense=Past|VerbForm=Part, Definite=Def|Degree=Sup|Number=Sing|POS=ADJ, Number=Plur|Number[psor]=Plur|POS=PRON|Person=1|Poss=Yes|PronType=Prs|Style=Form, Case=Gen|Definite=Def|Degree=Pos|Number=Sing|POS=ADJ, Case=Gen|Degree=Pos|Number=Plur|POS=ADJ, Case=Acc|Gender=Com|POS=PRON|Person=2|Polite=Form|PronType=Prs, Gender=Com|Number=Sing|POS=PRON|PronType=Int,Rel, POS=VERB|Tense=Pres, Case=Gen|Number=Plur|POS=DET|PronType=Ind, Number[psor]=Plur|POS=DET|Person=2|Poss=Yes|PronType=Prs, POS=PRON|Person=2|Polite=Form|Poss=Yes|PronType=Prs, Gender=Neut|Number=Sing|Number[psor]=Sing|POS=DET|Person=2|Poss=Yes|PronType=Prs, POS=AUX|Tense=Pres|VerbForm=Part, Mood=Ind|POS=VERB|Tense=Past|VerbForm=Fin|Voice=Pass, Gender=Com|Number=Sing|Number[psor]=Sing|POS=PRON|Person=3|Poss=Yes|PronType=Prs|Reflex=Yes, Degree=Sup|Number=Plur|POS=ADJ, Case=Acc|Gender=Com|Number=Plur|POS=PRON|Person=2|PronType=Prs, Gender=Neut|Number=Sing|Number[psor]=Sing|POS=PRON|Person=3|Poss=Yes|PronType=Prs|Reflex=Yes, Definite=Ind|Number=Plur|POS=NOUN, Case=Gen|Number=Plur|POS=VERB|Tense=Past|VerbForm=Part, Mood=Imp|POS=AUX, Gender=Com|Number=Sing|Number[psor]=Sing|POS=PRON|Person=1|Poss=Yes|PronType=Prs, Number[psor]=Sing|POS=PRON|Person=3|Poss=Yes|PronType=Prs, Definite=Def|Gender=Com|Number=Sing|POS=VERB|Tense=Past|VerbForm=Part, Number=Plur|Number[psor]=Sing|POS=DET|Person=2|Poss=Yes|PronType=Prs, Case=Gen|Gender=Com|Number=Sing|POS=DET|PronType=Ind, Case=Gen|POS=NOUN, Number[psor]=Plur|POS=PRON|Person=3|Poss=Yes|PronType=Prs, POS=DET|PronType=Dem, Definite=Def|Number=Plur|POS=NOUN
parser	ROOT, acl:relcl, advcl, advmod, amod, appos, aux, case, cc, ccomp, compound:prt, conj, cop, dep, det, expl, fixed, flat, iobj, list, mark, nmod, nmod:poss, nsubj, nummod, obj, obl, obl:loc, obl:tmod, punct, xcomp
ner	LOC, MISC, ORG, PER

Accuracy

Type	Score
POS_ACC	95.83
MORPH_ACC	95.70
DEP_UAS	84.92
DEP_LAS	81.76
SENTS_P	86.04
SENTS_R	87.41
SENTS_F	86.72
LEMMA_ACC	84.91
ENTS_F	82.32
ENTS_P	81.72
ENTS_R	82.92
TRANSFORMER_LOSS	41746686.63
MORPHOLOGIZER_LOSS	3458966.49
PARSER_LOSS	15104898.38
NER_LOSS	546098.45

Bias and Robustness

Besides the validation done by SpaCy on the DaNE testset, DaCy also provides a series of augmentations to the DaNE test set to see how well the models deal with these types of augmentations. The can be seen as behavioural probes akinn to the NLP checklist.

Deterministic Augmentations

Deterministic augmentations are augmentation which always yield the same result.

Augmentation	Part-of-speech tagging (Accuracy)	Morphological tagging (Accuracy)	Dependency Parsing (UAS)	Dependency Parsing (LAS)	Sentence segmentation (F1)	Lemmatization (Accuracy)	Named entity recognition (F1)
No augmentation	0.98	0.974	0.868	0.836	0.936	0.844	0.765
Æøå Augmentation	0.955	0.948	0.823	0.783	0.922	0.754	0.718
Lowercase	0.974	0.97	0.862	0.828	0.905	0.848	0.681
No Spacing	0.229	0.229	0.004	0.003	0.824	0.225	0.048
Abbreviated first names	0.979	0.973	0.864	0.832	0.94	0.845	0.699
Input size augmentation 5 sentences	0.956	0.956	0.851	0.818	0.883	0.844	0.743
Input size augmentation 10 sentences	0.959	0.958	0.853	0.821	0.897	0.844	0.755

Stochastic Augmentations

Stochastic augmentations are augmentation which are repeated mulitple times to estimate the effect of the augmentation.

Augmentation	Part-of-speech tagging (Accuracy)	Morphological tagging (Accuracy)	Dependency Parsing (UAS)	Dependency Parsing (LAS)	Sentence segmentation (F1)	Lemmatization (Accuracy)	Named entity recognition (F1)
Keystroke errors 2%	0.931 (0.003)	0.929 (0.003)	0.797 (0.003)	0.753 (0.003)	0.884 (0.003)	0.772 (0.003)	0.657 (0.003)
Keystroke errors 5%	0.859 (0.003)	0.863 (0.003)	0.699 (0.003)	0.641 (0.003)	0.824 (0.003)	0.681 (0.003)	0.53 (0.003)
Keystroke errors 15%	0.633 (0.006)	0.662 (0.006)	0.439 (0.006)	0.358 (0.006)	0.688 (0.006)	0.459 (0.006)	0.293 (0.006)
Danish names	0.979 (0.0)	0.974 (0.0)	0.867 (0.0)	0.835 (0.0)	0.943 (0.0)	0.847 (0.0)	0.748 (0.0)
Muslim names	0.979 (0.0)	0.974 (0.0)	0.865 (0.0)	0.833 (0.0)	0.94 (0.0)	0.847 (0.0)	0.732 (0.0)
Female names	0.979 (0.0)	0.974 (0.0)	0.867 (0.0)	0.835 (0.0)	0.946 (0.0)	0.847 (0.0)	0.754 (0.0)
Male names	0.979 (0.0)	0.974 (0.0)	0.867 (0.0)	0.835 (0.0)	0.943 (0.0)	0.847 (0.0)	0.748 (0.0)
Spacing Augmention 5%	0.941 (0.002)	0.936 (0.002)	0.755 (0.002)	0.725 (0.002)	0.907 (0.002)	0.811 (0.002)	0.699 (0.002)

Description of Augmenters

No augmentation: Applies no augmentation to the DaNE test set.

Æøå Augmentation: This augmentation replace the æ,ø, and å with their spelling variations ae, oe and aa respectively.

Lowercase: This augmentation lowercases all text.

No Spacing: This augmentation removed all spacing from the text.

Abbreviated first names: This agmentation abbreviates the first names of entities. For instance 'Kenneth Enevoldsen' would turn to 'K. Enevoldsen'.

Keystroke errors 2%: This agmentation simulate keystroke errors by replacing 2% of keys with a neighbouring key on a Danish QWERTY keyboard. As this agmentation is stochastic it is repeated 20 times to obtain a consistent estimate and the mean is provided with its standard deviation in parenthesis.

Keystroke errors 5%: This agmentation simulate keystroke errors by replacing 5% of keys with a neighbouring key on a Danish QWERTY keyboard. As this agmentation is stochastic it is repeated 20 times to obtain a consistent estimate and the mean is provided with its standard deviation in parenthesis.

Keystroke errors 15%: This agmentation simulate keystroke errors by replacing 15% of keys with a neighbouring key on a Danish QWERTY keyboard. As this agmentation is stochastic it is repeated 20 times to obtain a consistent estimate and the mean is provided with its standard deviation in parenthesis.

Danish names: This agmentation replace all names with Danish names derived from Danmarks Statistik (2021). As this agmentation is stochastic it is repeated 20 times to obtain a consistent estimate and the mean is provided with its standard deviation in parenthesis.

Muslim names: This agmentation replace all names with Muslim names derived from Meldgaard (2005). As this agmentation is stochastic it is repeated 20 times to obtain a consistent estimate and the mean is provided with its standard deviation in parenthesis.

Female names: This agmentation replace all names with Danish female names derived from Danmarks Statistik (2021). As this agmentation is stochastic it is repeated 20 times to obtain a consistent estimate and the mean is provided with its standard deviation in parenthesis.

Male names: This agmentation replace all names with Danish male names derived from Danmarks Statistik (2021). As this agmentation is stochastic it is repeated 20 times to obtain a consistent estimate and the mean is provided with its standard deviation in parenthesis.

Spacing Augmention 5%: This agmentation replace all names with Danish male names derived from Danmarks Statistik (2021). As this agmentation is stochastic it is repeated 20 times to obtain a consistent estimate and the mean is provided with its standard deviation in parenthesis.

Hardware

This was run an trained on a Quadro RTX 8000 GPU.