initial commit

README.md

@@ -0,0 +1,66 @@
+---
+language: nl
+license: apache-2.0
+tags:
+- dighum
+inference: false
+---
+
+# Early-modern Dutch NER (General Letters)
+
+## Description
+This is a fine-tuned NER model for early-modern Dutch United East India Company (VOC) letters based on XLM-R_base [(Conneau et al., 2020)](https://aclanthology.org/2020.acl-main.747/). The model identifies *locations*, *persons*, *organisations*, but also *ships* as well as derived forms of locations and religions.
+
+## Intended uses and limitations
+
+This model was fine-tuned (trained, validated and tested) on a single source of data, the General Letters (Generale Missiven). These letters span a large variety of Dutch, as they cover the largest part of the 17th and 18th centuries, and have been extended with editorial notes between 1960 and 2017. As the model was only fine-tuned on this data however, it may perform less well on other texts from the same period.
+
+
+## Training data and tagset
+
+The model was fine-tuned on the General Letters [GM-NER](https://github.com/cltl/voc-missives/tree/master/data/ner/datasplit_all_standard) dataset, with the following tagset:
+
+| tag | description | notes |
+| --- | ----------- | ----- |
+| LOC | locations   | |
+| LOCderiv | derived forms of locations | by derivation, e.g. *Bandanezen*, or composition, e.g. *Javakoffie* |
+| ORG | organisations | includes forms derived by composition, e.g. *Compagnieszaken*
+| PER | persons |
+| RELderiv | forms related to religion | merges religion names (*Christendom*), derived forms (*christenen*) and composed forms (*Christen-orangkay*) |
+| SHP | ships |
+
+The base text for this dataset is OCR text that has been partially corrected. The text is clean overall but errors remain.
+
+## Training procedure
+The model was fine-tuned with [xlm-roberta-base](https://huggingface.co/xlm-roberta-base), using [this  script](https://github.com/huggingface/transformers/blob/master/examples/legacy/token-classification/run_ner.py).
+
+Non-default training parameters are:
+* training batch size: 16
+* max sequence length: 256
+* number of epochs: 4 -- loading the best checkpoint model by loss at the end, with checkpoints every 200 steps
+* (seed: 1)
+
+
+## Evaluation 
+### Metric
+* entity-level F1
+
+### Results
+
+| overall | 92.7 |
+| --- | ----------- | 
+| LOC | 95.8   | 
+| LOCderiv | 92.7 | 
+| ORG | 92.5 |
+| PER | 86.2 | 
+| RELderiv | 90.7 |
+| SHP | 81.6 |
+
+
+## Authors and references
+### Authors
+Sophie Arnoult, Lodewijk Petram and Piek Vossen
+
+### Reference
+This model was fine-tuned as part of experiments for a paper accepted at [LaTeCH-CLfL 2021](https://sighum.wordpress.com/events/latech-clfl-2021/accepted-papers/): *Batavia asked for advice. Pretrained language models for Named Entity Recognition in historical texts.* 
+

config.json

@@ -0,0 +1,57 @@
+{
+  "_name_or_path": "/archive/transformers-hub/gm-ner-xlmrbase",
+  "architectures": [
+    "XLMRobertaForTokenClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "eos_token_id": 2,
+  "finetuning_task": "ner",
+  "gradient_checkpointing": false,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "id2label": {
+    "0": "B-LOC",
+    "1": "B-LOCderiv",
+    "2": "B-ORG",
+    "3": "B-PER",
+    "4": "B-RELderiv",
+    "5": "B-SHP",
+    "6": "I-LOC",
+    "7": "I-LOCderiv",
+    "8": "I-ORG",
+    "9": "I-PER",
+    "10": "I-SHP",
+    "11": "O"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "label2id": {
+    "B-LOC": 0,
+    "B-LOCderiv": 1,
+    "B-ORG": 2,
+    "B-PER": 3,
+    "B-RELderiv": 4,
+    "B-SHP": 5,
+    "I-LOC": 6,
+    "I-LOCderiv": 7,
+    "I-ORG": 8,
+    "I-PER": 9,
+    "I-SHP": 10,
+    "O": 11
+  },
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 514,
+  "model_type": "xlm-roberta",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "output_past": true,
+  "pad_token_id": 1,
+  "position_embedding_type": "absolute",
+  "torch_dtype": "float32",
+  "transformers_version": "4.9.0.dev0",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 250002
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d524916636607f6f8c7a37b0ff10890fb43c68509fdb84ab3c1f8821d305c13
+size 1109938871

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "sep_token": "</s>", "pad_token": "<pad>", "cls_token": "<s>", "mask_token": {"content": "<mask>", "single_word": false, "lstrip": true, "rstrip": false, "normalized": false}}

tf_model.h5

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c17b309ac994b42d28cdca95c2bed88a8a7625faeebe35af7d3a81653a4087e
+size 1110136976

tokenizer_config.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "sep_token": "</s>", "cls_token": "<s>", "unk_token": "<unk>", "pad_token": "<pad>", "mask_token": {"content": "<mask>", "single_word": false, "lstrip": true, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "model_max_length": 512, "special_tokens_map_file": null, "name_or_path": "xlm-roberta-base", "tokenizer_class": "XLMRobertaTokenizer"}