|
--- |
|
tags: |
|
- spacy |
|
- token-classification |
|
language: |
|
- da |
|
license: apache-2.0 |
|
model-index: |
|
- name: da_dacy_large_DANSK_ner |
|
results: |
|
- task: |
|
name: NER |
|
type: token-classification |
|
metrics: |
|
- name: NER Precision |
|
type: precision |
|
value: 0.8100263852 |
|
- name: NER Recall |
|
type: recall |
|
value: 0.8203072812 |
|
- name: NER F Score |
|
type: f_score |
|
value: 0.8151344175 |
|
--- |
|
|
|
<a href="https://github.com/centre-for-humanities-computing/Dacy"><img src="https://centre-for-humanities-computing.github.io/DaCy/_static/icon.png" width="175" height="175" align="right" /></a> |
|
|
|
# DaCy_large_DANSK_ner |
|
|
|
DaCy is a Danish language processing framework with state-of-the-art pipelines as well as functionality for analyzing Danish pipelines. |
|
At the time of publishing this model, also included in DaCy encorporates the only models for fine-grained NER using DANSK dataset - a dataset containing 18 annotation types in the same format as Ontonotes. |
|
Moreover, 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](https://github.com/centre-for-humanities-computing/DaCy) 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_large_DANSK_ner` | |
|
| **Version** | `0.1.0` | |
|
| **spaCy** | `>=3.5.0,<3.6.0` | |
|
| **Default Pipeline** | `transformer`, `ner` | |
|
| **Components** | `transformer`, `ner` | |
|
| **Vectors** | 0 keys, 0 unique vectors (0 dimensions) | |
|
| **Sources** | DANSK - Danish Annotations for NLP Specific TasKs<br />KennethEnevoldsen/dfm-bert-large-v1-2048bsz-1Msteps (Kenneth Enevoldsen) | |
|
| **License** | `apache-2.0` | |
|
| **Author** | [Centre for Humanities Computing Aarhus](https://chcaa.io/#/) | |
|
|
|
### Label Scheme |
|
|
|
<details> |
|
|
|
<summary>View label scheme (18 labels for 1 components)</summary> |
|
|
|
| Component | Labels | |
|
| --- | --- | |
|
| **`ner`** | `CARDINAL`, `DATE`, `EVENT`, `FACILITY`, `GPE`, `LANGUAGE`, `LAW`, `LOCATION`, `MONEY`, `NORP`, `ORDINAL`, `ORGANIZATION`, `PERCENT`, `PERSON`, `PRODUCT`, `QUANTITY`, `TIME`, `WORK OF ART` | |
|
|
|
</details> |
|
|
|
### Accuracy |
|
|
|
| Type | Score | |
|
| --- | --- | |
|
| `ENTS_F` | 81.51 | |
|
| `ENTS_P` | 81.00 | |
|
| `ENTS_R` | 82.03 | |
|
| `TRANSFORMER_LOSS` | 63375.61 | |
|
| `NER_LOSS` | 158164.20 | |
|
|
|
|
|
### Performance tables |
|
|
|
The table below shows the F1, recall and precision of the three DaCy fine-grained models. |
|
| Score | DaCy large | DaCy medium | DaCy small | |
|
|:---:|:---:|:---:|:---:| |
|
| F1 | 0.823 | 0.806 | 0.776 | |
|
| Recall | 0.834 | 0.818 | 0.77 | |
|
| Precision | 0.813 | 0.794 | 0.781 | |
|
|
|
The table below shows the F1 of the three DaCy fine-grained models within each named entity type. |
|
| Named-entity type | DaCy large | DaCy medium | DaCy small | |
|
|:---:|:---:|:---:|:---:| |
|
| CARDINAL | 0.874 | 0.781 | 0.887 | |
|
| DATE | 0.846 | 0.859 | 0.867 | |
|
| EVENT | 0.611 | 0.571 | 0.4 | |
|
| FACILITY | 0.545 | 0.533 | 0.468 | |
|
| GPE | 0.893 | 0.838 | 0.794 | |
|
| LANGUAGE | 0.902 | 0.486 | 0.194 | |
|
| LAW | 0.686 | 0.625 | 0.606 | |
|
| LOCATION | 0.633 | 0.737 | 0.581 | |
|
| MONEY | 0.993 | 1 | 0.947 | |
|
| NORP | 0.78 | 0.887 | 0.785 | |
|
| ORDINAL | 0.696 | 0.7 | 0.727 | |
|
| ORGANIZATION | 0.863 | 0.851 | 0.781 | |
|
| PERCENT | 0.923 | 0.96 | 0.96 | |
|
| PERSON | 0.871 | 0.872 | 0.833 | |
|
| PRODUCT | 0.671 | 0.635 | 0.526 | |
|
| QUANTITY | 0.386 | 0.654 | 0.708 | |
|
| TIME | 0.643 | 0.571 | 0.71 | |
|
| WORK OF ART | 0.494 | 0.639 | 0.488 | |
|
|
|
The table below shows the F1 of the three DaCy fine-grained models within each domain of texts in DANSK. |
|
| Domain | DaCy large | DaCy medium | DaCy small | |
|
|:---:|:---:|:---:|:---:| |
|
| All domains combined | 0.823 | 0.806 | 0.776 | |
|
| Conversation | 0.796 | 0.718 | 0.82 | |
|
| Dannet | 0.75 | 0.667 | 1 | |
|
| Legal | 0.852 | 0.854 | 0.866 | |
|
| News | 0.841 | 0.759 | 0.86 | |
|
| Social Media | 0.793 | 0.847 | 0.8 | |
|
| Web | 0.826 | 0.802 | 0.756 | |
|
| Wiki and Books | 0.778 | 0.838 | 0.709 | |
|
|
