tuongtranngoc / language-independent-entity-linking

A official implementation of Language-independent Entity Linking using ML Classifier

Language-independent-Entity-Linking

A implementation of languague-independent Entity Linking using ML Classifier

Introduction

The state-of-the-art VIE, KIE models as examples LayoutXLM, LayoutLMv3, VI-LayoutXLM, ... that they would make misstakes in certain relatively simple scenarios, where the geometric relations between entities are not complicated. They seems to link two entities dependending more in the semantics than geometric layout. To futher verify my conjecture, I conduct an experiment by creating the linkings between entities using features: distance, direction, angle, coordinates based on bounding boxes of text.

Environment

The dependencies are listed in requirements.txt. Please install follow the command as below:

pip install -r requirements.txt

Dataset and model checkpoints

You can download raw dataset from here, and put the file xfund&funsd under the root folder.

You can download pre-processed dataset and model checkpoints from here include features, weights and put the folders under src/ directory.

Training

Before training, please modify configuration in src/config/cfg.py file and getting started training

cd src/
python train --lang <language-specific>

Evaluation

To evaluate for languague specific and one-shot task (training and evaluate on language X).

cd src
python evaluate.py --lang <languague-specific>

To evaluate for zero-shot task (training on English language and evaluate on language X).

cd src
python evaluate.py --task zero-shot

Experiments

F1-score relation extraction on tasks

Languague-specific task

	EN(FUNSD)	ZH	JA	ES	FR	IT	DE	PT	Avg
$LayoutXLM_{Large}$	64.1	78.9	72.5	76.7	71.1	76.9	68.4	67.9	72.1
Our approach	92.5	72.5	78.6	75.8	82.2	84.5	76.1	71.6	79.2

Zero-shot task

	EN(FUNSD)	ZH	JA	ES	FR	IT	DE	PT	Avg
$LayoutXLM_{Large}$	64.1	55.3	56.9	57.8	56.1	51.8	48.9	47.9	54.8
Our approach	92.5	68.1	72.5	72.8	83.5	84.4	73.3	68.1	74.6