--- language: - en license: mit tags: - token-classification - distilbert - pytorch - ner - conll2003 datasets: - conll2003 metrics: - f1 pipeline_tag: token-classification widget: - text: "Alice works at Microsoft in London." - text: "Sara works at Google in Cairo." - text: "Mohamed studies at MIT in USA." - text: "Apple opened an office in Paris." - text: "An Egyptian student moved to Paris." --- # 🏷️ Named Entity Recognition — Fine-Tuned DistilBERT This is the fine-tuned **cased DistilBERT** model for English Named Entity Recognition (NER) trained on the **CoNLL-2003** corpus. It detects and classifies tokens into four entity categories: - **PER** (Person) - **ORG** (Organization) - **LOC** (Location / Place) - **MISC** (Miscellaneous, e.g. nationalities, events) It achieved the winning score in an architectural comparison against LSTM, BiLSTM, and BiLSTM + CRF models. ## 📊 Performance (Strict CoNLL-2003 Evaluation) Evaluated with `seqeval` in strict IOB2 mode (requiring exact boundary and category match): | Entity Type | Precision | Recall | Strict F1 | |---|---:|---:|---:| | **Overall** | **89.39%** | **90.98%** | **90.18%** | | Person (`PER`) | 95.34% | 95.03% | **95.18%** | | Organization (`ORG`) | 86.81% | 88.75% | **87.77%** | | Location (`LOC`) | 92.51% | 92.23% | **92.37%** | | Miscellaneous (`MISC`) | 78.43% | 80.57% | **79.49%** | - **Validation Macro F1:** 93.92% - **Test Strict F1:** 90.18% ## 🚀 Quick Usage ### Using Hugging Face Pipeline ```python from transformers import pipeline ner = pipeline("ner", model="Solayman-Sameh/named-entity-recognition", aggregation_strategy="simple") text = "Sara works at Google in Cairo." entities = ner(text) for ent in entities: print(f"{ent['word']} -> {ent['entity_group']} ({ent['score']:.2%})") ``` ### Using PyTorch Directly ```python import torch from transformers import AutoTokenizer, AutoModelForTokenClassification model_id = "Solayman-Sameh/named-entity-recognition" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForTokenClassification.from_pretrained(model_id).eval() text = "Mohamed studies at MIT in USA." inputs = tokenizer(text, return_tensors="pt") with torch.inference_mode(): outputs = model(**inputs) predictions = outputs.logits.argmax(dim=-1)[0].tolist() tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0]) for token, pred in zip(tokens, predictions): if token not in [tokenizer.cls_token, tokenizer.sep_token, tokenizer.pad_token]: print(f"{token}: {model.config.id2label[pred]}") ``` ## 💻 Full Project & Web App The complete source code, exploratory data analysis, architectural comparison (LSTM, BiLSTM, BiLSTM+CRF vs. DistilBERT), test reports, and an interactive Gradio web demo are available at: 👉 [GitHub: Solayman-Sameh/named-entity-recognition](https://github.com/Solayman-Sameh/named-entity-recognition)