add readme

README.md

@@ -0,0 +1,86 @@
+---
+language:
+- sv
+license: cc-by-4.0
+size_categories:
+- 1M<n<10M
+pretty_name: Open Riksdag-103
+tags:
+- diachronic
+- semantic change
+dataset_info:
+- config_name: sentences
+  features:
+  - name: sentence
+    dtype: string
+  - name: doc_type
+    dtype: string
+  - name: doc_id
+    dtype: string
+  - name: date
+    dtype: timestamp[s]
+  splits:
+  - name: train
+    num_bytes: 8436701719
+    num_examples: 56846721
+  download_size: 1516962051
+  dataset_size: 8436701719
+- config_name: target-103
+  features:
+  - name: sentence
+    dtype: string
+  - name: doc_type
+    dtype: string
+  - name: doc_id
+    dtype: string
+  - name: date
+    dtype: timestamp[s]
+  - name: lemma
+    dtype: string
+  - name: start
+    dtype: int32
+  - name: end
+    dtype: int32
+  - name: pos
+    dtype: string
+  splits:
+  - name: train
+    num_bytes: 8138875561
+    num_examples: 33393155
+  download_size: 1434826241
+  dataset_size: 8138875561
+---
+
+This is a dataset of text from the Riksdag, Sweden's national legislative body. 
+
+The original data is availble without a license under the Re-use of Public Administration Documents Act (2010:566) at https://data.riksdagen.se/data/dokument
+
+This dataset is derivative of a version compiled by Språkbanken Text (SBX) at the University of Gothenburg (Sweden). That version consists of XML files split by source document type (motions, questions, protocol, etc.) and includes additional linguistic annotations. It is available under a CC BY 4.0 license at https://spraakbanken.gu.se/resurser/rd 
+
+The focus of this huggingface dataset is to organise the data for fine-grained diachronic modeling. In a nutshell, this version offers:
+
+- all sentences including one or more of 103 target words, which were chosen by TF-IDF (described below)
+- per-month subsets (with all document types combined)
+- one line per sentence (sentences shorter than 4 words were discarded)
+- data includes: date, document_type, document_id, target_word, and text.
+
+The dataset builder requires a `years` argument, which must be an interable of years between 1979 and 2019 (inclusive). This can be supplied to the `load_dataset` function as a keyword argument.
+For example, to load all of the data:
+
+```python
+from datasets import load_dataset
+data = load_dataset('ChangeIsKey/openRD-103', years=range(1979,2020))
+```
+
+The data can take some time to load/extract. Using [dataset streaming](https://huggingface.co/docs/datasets/stream) may be an option. Size of data by decade can be found below:
+
+|       | bytes | sentences | tokens |
+|-------|-------|-----------|--------|
+| 1979  | 118Mb | 0.409M    | 10M    |
+| 1980s | 1.4Gb | 4.7M      | 118M   |
+| 1990s | 2.2Gb | 5.3M      | 202M   |
+| 2000s | 4.0Gb | 11.8M     | 338M   |
+| 2010s | 4.4Gb | 14.1M     | 361M   |
+| total | 13Gb  | 36.9M     | 279M   |
+
+License is CC BY 4.0 with attribution.