README.md

metadata

tags:
  - BERT
  - Text Classification
  - relation
language: Arabic
license: mit
datasets:
  - ACE2005

Arabic Relation Extraction Model

• Github repo
• Relation Extraction model based on GigaBERTv4.
• Model detail: mark two entities in the sentence with special markers (e.g., XXXX <PER> entity1 </PER> XXXXXXX <ORG> entity2 </ORG> XXXXX). Then we use the BERT [CLS] representation to make a prediction.
• ACE2005 Training data: Arabic
• Relation tags including: Physical, Part-whole, Personal-Social, ORG-Affiliation, Agent-Artifact, Gen-Affiliation

Hyperparameters

• learning_rate=2e-5
• num_train_epochs=10
• weight_decay=0.01

ACE2005 Evaluation results (F1)

Language	Arabic
	74.7

How to use

Workflow of a relation extraction model:

1. Input --> NER model --> Entities
2. Input sentence + Entity 1 + Entity 2 --> Relation Classification Model --> Relation Type

>>> from transformers import pipeline, AutoModelForTokenClassification, AutoTokenizer, AuotoModelForSequenceClassification

>>> ner_model = AutoModelForTokenClassification.from_pretrained("ychenNLP/arabic-ner-ace")
>>> ner_tokenizer = AutoTokenizer.from_pretrained("ychenNLP/arabic-ner-ace")
>>> ner_pip = pipeline("ner", model=ner_model, tokenizer=ner_tokenizer, grouped_entities=True)

>>> re_model = AutoModelForSequenceClassification.from_pretrained("ychenNLP/arabic-relation-extraction")
>>> re_tokenizer = AutoTokenizer.from_pretrained("ychenNLP/arabic-relation-extraction")
>>> re_pip = pipeline("text-classification", model=re_model, tokenizer=re_tokenizer)

def process_ner_output(entity_mention, inputs):
    re_input = []
    for idx1 in range(len(entity_mention) - 1):
        for idx2 in range(idx1 + 1, len(entity_mention)):
            ent_1 = entity_mention[idx1]
            ent_2 = entity_mention[idx2]

            ent_1_type = ent_1['entity_group']
            ent_2_type = ent_2['entity_group']
            ent_1_s = ent_1['start']
            ent_1_e = ent_1['end']
            ent_2_s = ent_2['start']
            ent_2_e = ent_2['end']
            new_re_input = ""
            for c_idx, c in enumerate(inputs):
                if c_idx == ent_1_s:
                    new_re_input += "<{}>".format(ent_1_type)
                elif c_idx == ent_1_e:
                    new_re_input += "</{}>".format(ent_1_type)
                elif c_idx == ent_2_s:
                    new_re_input += "<{}>".format(ent_2_type)
                elif c_idx == ent_2_e:
                    new_re_input += "</{}>".format(ent_2_type)
                new_re_input += c
            re_input.append({"re_input": new_re_input, "arg1": ent_1, "arg2": ent_2, "input": inputs})
    return re_input
    
def post_process_re_output(re_output, text_input, ner_output):
    final_output = []
    for idx, out in enumerate(re_output):
        if out["label"] != 'O':
            tmp = re_input[idx]
            tmp['relation_type'] = out
            tmp.pop('re_input', None)
            final_output.append(tmp)

    template = {"input": text_input,
                "entity": ner_output,
                "relation": final_output}

    return template

text_input = 'قال وزير العدل التركي بكير بوزداغ إن أنقرة تريد 12 مشتبهاً بهم من فنلندا و 21 من السويد'
ner_output = ner_pip(text_input) # inference NER tags

re_input = process_ner_output(ner_output, text_input) # prepare a pair of entity and predict relation type

re_output = []
for idx in range(len(re_input)):
    tmp_re_output = re_pip(re_input[idx]["re_input"]) # for each pair of entity, predict relation
    re_output.append(tmp_re_output[0])



re_ner_output = post_process_re_output(re_output, text_input, ner_output) # post process NER and relation predictions
print("Sentence: ",re_ner_output["input"])
print('====Entity====')
for ent in re_ner_output["entity"]:
  print('{}--{}'.format(ent["word"], ent["entity_group"]))
print('====Relation====')
for rel in re_ner_output["relation"]:
  print('{}--{}:{}'.format(rel['arg1']['word'], rel['arg2']['word'], rel['relation_type']['label']))

Sentence:  قال وزير العدل التركي بكير بوزداغ إن أنقرة تريد 12 مشتبهاً بهم من فنلندا و 21 من السويد
====Entity====
وزير--PER
العدل--ORG
التركي--GPE
بكير بوزداغ--PER
انقرة--GPE
مشتبها بهم--PER
فنلندا--GPE
21--PER
السويد--GPE
====Relation====
وزير--العدل:ORG-AFF
مشتبها بهم--فنلندا:PHYS
21--السويد:PHYS

BibTeX entry and citation info

@inproceedings{lan2020gigabert,
  author     = {Lan, Wuwei and Chen, Yang and Xu, Wei and Ritter, Alan},
    title      = {Giga{BERT}: Zero-shot Transfer Learning from {E}nglish to {A}rabic},
    booktitle  = {Proceedings of The 2020 Conference on Empirical Methods on Natural Language Processing (EMNLP)},
    year       = {2020}
  }