Datasets:

Languages:
Norwegian
Multilinguality:
monolingual
Size Categories:
1K<n<10K
Language Creators:
found
Annotations Creators:
expert-generated
Source Datasets:
original
ArXiv:
Tags:
License:
pere commited on
Commit
d3c13cd
1 Parent(s): 4e81ff1
Files changed (1) hide show
  1. README.md +23 -12
README.md CHANGED
@@ -59,26 +59,37 @@ This is a classification dataset created from a subset of the [Talk of Norway](h
59
 
60
  This dataset is meant for classification. Results can for instance be viewed in [this article](https://arxiv.org/abs/2104.09617).
61
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
62
  ### Languages
63
 
64
  The text in the dataset is in Norwegian.
65
 
66
  ## Dataset Structure
67
 
68
- ### Data Instances
69
-
70
- Example of one instance in the dataset.
71
-
72
- ```{'label': 0, 'text': 'Verre er det med slagsmålene .', 'date': '2016-01-15'}```
73
-
74
  ### Data Fields
75
 
76
- - `id`: index of the example
77
  - `text`: Text of a speech
78
- - `date`: Date (`YYYY-MM-DD`) the speech was produced
79
  - `label`: Political party the speaker was associated with at the time
80
- - 0 = Fremskrittspartiet
81
- - 1 = Sosialistisk Venstreparti
 
 
 
82
 
83
  ### Data Splits
84
 
@@ -95,7 +106,7 @@ The dataset is balanced on political party.
95
  This dataset is based on the publicly available information by Norwegian Parliament (Storting) and created by the National Library of Norway AI-Lab to benchmark their language models.
96
 
97
  ## Additional Information
98
- The [Talk of Norway dataset] is also available [here](https://www.nb.no/sprakbanken/ressurskatalog/oai-repo-clarino-uib-no-11509-123/).
99
 
100
  ### Licensing Information
101
 
@@ -123,4 +134,4 @@ in several token and sequence classification tasks for both Norwegian Bokm{aa}l
123
  languages present in the corpus such as English, Swedish, and Danish. For languages not included in the corpus, the weights degrade moderately while keeping strong multilingual properties. Therefore,
124
  we show that building high-quality models within a memory institution using somewhat noisy optical character recognition (OCR) content is feasible, and we hope to pave the way for other memory institutions to follow.",
125
  }
126
- ```
 
59
 
60
  This dataset is meant for classification. Results can for instance be viewed in [this article](https://arxiv.org/abs/2104.09617).
61
 
62
+ The dataset was also evaluated as part of the [North-T5](https://huggingface.co/north/t5_large_NCC) training. It then gave these results:
63
+
64
+ |**Model:** | **F1** |
65
+ |:-----------|:------------|
66
+ |mT5-base|73.2 |
67
+ |mBERT-base|78.4 |
68
+ |NorBERT-base|78.2 |
69
+ |North-T5-small|80.5 |
70
+ |nb-bert-base|81.8 |
71
+ |North-T5-base|85.3 |
72
+ |North-T5-large|86.7 |
73
+ |North-T5-xl|88.7 |
74
+ |North-T5-xxl|91.8|
75
+
76
+
77
  ### Languages
78
 
79
  The text in the dataset is in Norwegian.
80
 
81
  ## Dataset Structure
82
 
 
 
 
 
 
 
83
  ### Data Fields
84
 
 
85
  - `text`: Text of a speech
86
+ - `date`: Date (`YYYY-MM-DD`) the speech was held
87
  - `label`: Political party the speaker was associated with at the time
88
+ - 0
89
+ - 1
90
+ - `full_label`: Political party the speaker was associated with at the time
91
+ - Fremskrittspartiet
92
+ - Sosialistisk Venstreparti
93
 
94
  ### Data Splits
95
 
 
106
  This dataset is based on the publicly available information by Norwegian Parliament (Storting) and created by the National Library of Norway AI-Lab to benchmark their language models.
107
 
108
  ## Additional Information
109
+ The [Talk of Norway dataset](https://www.nb.no/sprakbanken/ressurskatalog/oai-repo-clarino-uib-no-11509-123/) is also available at the [LTG Talk-of-Norway Github](https://github.com/ltgoslo/talk-of-norway).
110
 
111
  ### Licensing Information
112
 
 
134
  languages present in the corpus such as English, Swedish, and Danish. For languages not included in the corpus, the weights degrade moderately while keeping strong multilingual properties. Therefore,
135
  we show that building high-quality models within a memory institution using somewhat noisy optical character recognition (OCR) content is feasible, and we hope to pave the way for other memory institutions to follow.",
136
  }
137
+ ```