honghanhh / ner-combining-contextual-and-global-features

[ICADL] Named entity recognition architecture combining contextual and global features

Named entity recognition architecture combining contextual and global features

Description

Named entity recognition (NER) is an information extraction technique that aims to locate and classify named entities within a document into predefined categories. However, it is still a difficult task because named entities (NEs) have multiple forms and they are context dependent. We propose the combination of contextual features from XLNet and global features from Graph Convolution Network (GCN) to enhance NER performance. Experiments over a widely-useddataset, CoNLL 2003, show the benefits of our strategy, with results competitive with the state of the art.

---

Implementation

1. Requirements

All the necessary library are noted in requirements.txt, please run the following command to install:

pip install -r requirements.txt

2. Training models

Run the following command to train the model:

btslen.sh

The train.py contains XLNet model whereas the combined_train.py includes the combination of XLNet and GCN, which is also our finalized model. Feel free to change the hyperparameters inside btslen.sh for further experiments and model tuning.

---

Results

1. Performance between joint architecture vs standalone one.

Embeddings	F1-score
Global features	88.63
Contextual features	93.28
Global + contextual features	93.82

2. Performance evaluation per entity type.

Entity types	Precision	Recall	F1-score
LOC	94.15	93.53	93.83
MISC	81.33	81.89	81.62
ORG	88.97	92.29	90.60
PER	96.67	97.09	96.88

3. Comparison of our proposal against SOTA techniques.

---

References

Hanh, Tran Thi Hong, et al. "Named Entity Recognition Architecture Combining Contextual and Global Features." International Conference on Asian Digital Libraries. Springer, Cham, 2021.

---

Contributors:

• Hồng Hạnh
• Prof. dr. Nicolas SIDERE
• Prof. dr. Antoine DOUCET
• Prof. dr. Jose MORENO
• Asst. prof. dr. Senja POLLAK.