Edit model card

OALZ/1788/Q1/NER

A named entity recognition system (NER) was trained on text extracted from Oberdeutsche Allgemeine Litteraturueitung (OALZ) of the first quarter (January, Febuary, March) of 1788. The scans from which text was extracted can be found at Bayerische Staatsbibliothek using the extraction strategy of the KEDiff project, which can be found at cborgelt/KEDiff.

Annotations

Each text passage was annotated in doccano by two or three annotators and their annotations were cleaned and merged into one dataset. For details on how this was done, see LelViLamp/kediff-doccano-postprocessing. In total, the text consists of about 1.7m characters. The resulting annotation datasets were published on the Hugging Face Hub. There are two versions of the dataset

  • union-dataset contains the texts split into chunks. This is how they were presented in the annotation application doccano and results from preprocessing step 5a.
  • merged-union-dataset does not retain this split. The text was merged into one long text and annotation indices were adapted in preprocessing step 5b.

Note that both these directories contain three equivalent datasets each:

  • a Huggingface/Arrow dataset, *
  • a CSV, * and
  • a JSONL file.

* The former two should be used together with the provided text.csv to catch the context of the annotation. The latter JSONL file contains the full text.

The following categories were included in the annotation process:

Tag Label Count Total Length Median Annotation Length Mean Annotation Length SD
EVENT Event 294 6,090 18 20.71 13.24
LOC Location 2,449 24,417 9 9.97 6.21
MISC Miscellaneous 2,585 50,654 14 19.60 19.63
ORG Organisation 2,479 34,693 11 13.99 9.33
PER Person 7,055 64,710 7 9.17 9.35
TIME Dates & Time 1,076 13,154 8 12.22 10.98

NER models

Based on the annotations above, six separate NER classifiers were trained, one for each label type. This was done in order to allow overlapping annotations. For example, in the passage "Dieses Projekt wurde an der Universität Salzburg durchgeführt", you would want to categorise "Universität Salzburg" as an organisation while also extracting "Salzburg" as a location. This would result in an annotation like this:

{
  "id": "example-42",
  "text": "Dieses Projekt wurde an der Universität Salzburg durchgeführt",
  "label": [[28, 49, "ORG"], [40, 49, "LOC"]]
}

Example entry in CSV and Huggingface dataset

annotation_id line_id start end label label_text merged
$n$ example-42 28 49 ORG Universität Salzburg ???
$n+1$ example-42 40 49 LOC Salzburg ???

The columns mean:

  • annotation_id was assigned internally by enumerating all annotations. This is not present in the JSONL format
  • line_id is the fragment of the subdivided text, as shown in doccano. Called id in the JSONL dataset.
  • start index of the first character that is annotated. Included, starts with 0.
  • end index of the last character that is annotated. Excluded, maximum value is len(respectiveText).
  • label indicates what the passage indicated by $[start, end)$ was annotated as.
  • label_text contains the text that is annotated by $[start, end)$. This is not present in the JSONL dataset as it can be inferred there.
  • merged indicates whether this annotation is the result of overlapping annotations of the same label. In that case, annotation_id contains the IDs of the individual annotations it was constructed of. This is not present in the JSONL dataset.

To achieve this overlap, each text passage must be run through all the classifiers individually and each classifier's results need to be combined. For details on how the training was done, see LelViLamp/kediff-ner-training.

The dbmdz/bert-base-historic-multilingual-cased tokeniser was used to create historical embeddings. Therefore, it is necessary to use that in order to use these NER models.

The models' performance measures are as follows:

Model Selected Epoch Checkpoint Validation Loss Precision Recall F1 Accuracy
EVENT 1 1393 .021957 .665233 .343066 .351528 .995700
LOC 1 1393 .033602 .829535 .803648 .814146 .990999
MISC 2 2786 .123994 .739221 .503677 .571298 968697
ORG 1 1393 .062769 .744259 .709738 .726212 .980288
PER 2 2786 .059186 .914037 .849048 .879070 .983253
TIME 1 1393 .016120 .866866 .724958 .783099 .994631

Acknowledgements

The data set and models were created in the project Kooperative Erschließung diffusen Wissens (KEDiff), funded by the State of Salzburg, Austria 🇦🇹, and carried out at Paris Lodron Universität Salzburg.

Downloads last month
15
Safetensors
Model size
110M params
Tensor type
F32
·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.