magistermilitum's picture
Update README.md
098aa7e
metadata
language:
  - lat
  - fra
  - spa
  - multilingual
license: cc-by-nc-4.0
tags:
  - text
  - named entity recognition
  - roberta
  - historical languages
  - precision
  - recall
model-index:
  - name: roberta-multilingual-medieval-ner
    results:
      - task:
          type: named entity recognition
        metrics:
          - type: precision
            value: 98.01
          - type: Recall
            value: 97.08
inference:
  parameters:
    aggregation_strategy: simple
widget:
  - text: >-
      In nomine sanctæ et individuæ Trinitatis. Ego Guido, Dei gratia
      Cathalaunensis episcopus, propter inevitabilem temporum mutationem et
      casum decedentium quotidie personarum, necesse habemus litteris annotare
      quod dampnosa delere non possit oblivio. Eapropter notum fieri volumus tam
      futuris quam presentibus quod, pro remedio animæ meæ et predecessorum
      nostrorum, abbati et fratribus de Insula altare de Hattunmaisnil dedimus
      et perpetuo habendum concessimus, salvis custumiis nostris et archidiaconi
      loci illius. Ne hoc ergo malignorum hominum perversitate aut temporis
      alteratur incommodo presentem paginam sigilli nostri impressione
      firmavimus, testibus subnotatis : S. Raynardy capellani, Roberti Armensis,
      Mathei de Waisseio, Michaeli decani, Hugonis de Monasterio, Hervaudi de
      Panceio. Data per manum Gerardi cancellarii, anno ab incarnatione Domini
      millesimo centesimo septuagesimo octavo. 

Model Details

This is a Fine-tuned version of the multilingual Roberta model on medieval charters. The model is intended to recognize Locations and persons in medieval texts in a Flat and nested manner. The train dataset entails 8k annotated texts on medieval latin, french and Spanish from a period ranging from 11th to 15th centuries.

How to Get Started with the Model

The model is intended to be used in a simple way manner:

import torch
from transformers import pipeline

pipe = pipeline("token-classification", model="magistermilitum/roberta-multilingual-medieval-ner")

results = list(map(pipe, list_of_sentences))
results =[[[y["entity"],y["word"], y["start"], y["end"]] for y in x] for x in results]
print(results)

Model Description

The following snippet can transforms model inferences to CONLL format using the BIO format.

class TextProcessor:
    def __init__(self, filename):
        self.filename = filename
        self.sent_detector = nltk.data.load("tokenizers/punkt/english.pickle") #sentence tokenizer
        self.sentences = []
        self.new_sentences = []
        self.results = []
        self.new_sentences_token_info = []
        self.new_sentences_bio = []
        self.BIO_TAGS = []
        self.stripped_BIO_TAGS = []

    def read_file(self):
        #Reading a txt file with one document per line.
        with open(self.filename, 'r') as f:
            text = f.read()
        self.sentences = self.sent_detector.tokenize(text.strip())

    def process_sentences(self): #We split long sentences as encoder has a 256 max-lenght. Sentences with les of 40 words will be merged.
        for sentence in self.sentences:
            if len(sentence.split()) < 40 and self.new_sentences:
                self.new_sentences[-1] += " " + sentence
            else:
                self.new_sentences.append(sentence)

    def apply_model(self, pipe):
        self.results = list(map(pipe, self.new_sentences))
        self.results=[[[y["entity"],y["word"], y["start"], y["end"]] for y in x] for x in self.results]

    def tokenize_sentences(self):
        for n_s in self.new_sentences:
            tokens=n_s.split() # Basic tokenization
            token_info = []

            # Initialize a variable to keep track of character index
            char_index = 0
            # Iterate through the tokens and record start and end info
            for token in tokens:
                start = char_index
                end = char_index + len(token)  # Subtract 1 for the last character of the token
                token_info.append((token, start, end))

                char_index += len(token) + 1  # Add 1 for the whitespace
            self.new_sentences_token_info.append(token_info)

    def process_results(self): #merge subwords and BIO tags
        for result in self.results:
            merged_bio_result = []
            current_word = ""
            current_label = None
            current_start = None
            current_end = None
            for entity, subword, start, end in result:
                if subword.startswith("▁"):
                    subword = subword[1:]
                    merged_bio_result.append([current_word, current_label, current_start, current_end])
                    current_word = "" ; current_label = None ; current_start = None ; current_end = None
                if current_start is None:
                    current_word = subword ; current_label = entity ; current_start = start+1 ; current_end= end
                else:
                    current_word += subword ; current_end = end
            if current_word:
                merged_bio_result.append([current_word, current_label, current_start, current_end])
            self.new_sentences_bio.append(merged_bio_result[1:])

    def match_tokens_with_entities(self): #match BIO tags with tokens
        for i,ss in enumerate(self.new_sentences_token_info):
            for word in ss:
                for ent in self.new_sentences_bio[i]:
                    if word[1]==ent[2]:
                        if ent[1]=="L-PERS":
                            self.BIO_TAGS.append([word[0], "I-PERS", "B-LOC"])
                            break
                        else:
                            if "LOC" in ent[1]:
                                self.BIO_TAGS.append([word[0], "O", ent[1]])
                            else:
                                self.BIO_TAGS.append([word[0], ent[1], "O"])
                            break
                else:
                    self.BIO_TAGS.append([word[0], "O", "O"])

    def separate_dots_and_comma(self): #optional
        signs=[",", ";", ":", "."]
        for bio in self.BIO_TAGS:
            if any(bio[0][-1]==sign for sign in signs) and len(bio[0])>1:
                self.stripped_BIO_TAGS.append([bio[0][:-1], bio[1], bio[2]]); 
                self.stripped_BIO_TAGS.append([bio[0][-1], "O", "O"])
            else:
                self.stripped_BIO_TAGS.append(bio)

    def save_BIO(self):
        with open('output_BIO_a.txt', 'w', encoding='utf-8') as output_file:
            output_file.write("TOKEN\tPERS\tLOCS\n"+"\n".join(["\t".join(x) for x in self.stripped_BIO_TAGS]))

# Usage:
processor = TextProcessor('my_docs_file.txt')
processor.read_file()
processor.process_sentences()
processor.apply_model(pipe)
processor.tokenize_sentences()
processor.process_results()
processor.match_tokens_with_entities()
processor.separate_dots_and_comma()
processor.save_BIO()
  • Developed by: [Sergio Torres Aguilar]
  • Model type: [XLM-Roberta]
  • Language(s) (NLP): [Medieval Latin, Spanish, French]
  • Finetuned from model [optional]: [Named Entity Recognition]

Direct Use

A sentence as : "Ego Radulfus de Francorvilla miles, notum facio tam presentibus cum futuris quod, cum Guillelmo Bateste militi de Miliaco"

Will be annotated in BIO format as:

('Ego', 'O', 'O')
('Radulfus', 'B-PERS')
('de', 'I-PERS', 'O')
('Francorvilla', 'I-PERS', 'B-LOC')
('miles', 'O')
(',', 'O', 'O')
('notum', 'O', 'O')
('facio', 'O', 'O')
('tam', 'O', 'O')
('presentibus', 'O', 'O')
('quam', 'O', 'O')
('futuris', 'O', 'O')
('quod', 'O', 'O')
(',', 'O', 'O')
('cum', 'O', 'O')
('Guillelmo', 'B-PERS', 'O')
('Bateste', 'I-PERS', 'O')
('militi', 'O', 'O')
('de', 'O', 'O')
('Miliaco', 'O', 'B-LOC')

Training Procedure

The model was fine-tuned during 5 epoch on the XML-Roberta-Large using a 5e-5 Lr and a batch size of 16.

BibTeX:

@inproceedings{aguilar2022multilingual,
  title={Multilingual Named Entity Recognition for Medieval Charters Using Stacked Embeddings and Bert-based Models.},
  author={Aguilar, Sergio Torres},
  booktitle={Proceedings of the second workshop on language technologies for historical and ancient languages},
  pages={119--128},
  year={2022}
}

Model Card Contact

[[email protected]]