dangvantuan/vietnamese-embedding

Model Description:

vietnamese-embedding is the Embedding Model for Vietnamese language. This model is a specialized sentence-embedding trained specifically for the Vietnamese language, leveraging the robust capabilities of PhoBERT, a pre-trained language model based on the RoBERTa architecture. The model utilizes PhoBERT to encode Vietnamese sentences into a 768-dimensional vector space, facilitating a wide range of applications from semantic search to text clustering. The embeddings capture the nuanced meanings of Vietnamese sentences, reflecting both the lexical and contextual layers of the language.

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, 'do_lower_case': False}) with Transformer model: RobertaModel 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
)

Training and Fine-tuning process

The model underwent a rigorous four-stage training and fine-tuning process, each tailored to enhance its ability to generate precise and contextually relevant sentence embeddings for the Vietnamese language. Below is an outline of these stages:

Stage 1: Initial Training

• Dataset: ViNLI-SimCSE-supervised
• Method: Trained using the SimCSE approach which employs a supervised contrastive learning framework. The model was optimized using Triplet Loss to effectively learn from high-quality annotated sentence pairs.

Stage 2: Continued Fine-tuning

• Dataset: XNLI-vn
• Method: Continued fine-tuning using Multi-Negative Ranking Loss. This stage focused on improving the model's ability to discern and rank nuanced differences in sentence semantics.

Stage 3: Continued Fine-tuning for Semantic Textual Similarity on STS Benchmark

• Dataset: STSB-vn
• Method: Fine-tuning specifically for the semantic textual similarity benchmark using Siamese BERT-Networks configured with the 'sentence-transformers' library. This stage honed the model's precision in capturing semantic similarity across various types of Vietnamese texts.

Stage 4: Advanced Augmentation Fine-tuning

• Dataset: STSB-vn with generate silver sample from gold sample
• Method: Employed an advanced strategy using Augmented SBERT with Pair Sampling Strategies, integrating both Cross-Encoder and Bi-Encoder models. This stage further refined the embeddings by enriching the training data dynamically, enhancing the model's robustness and accuracy in understanding and processing complex Vietnamese language constructs.

Usage:

Using this model becomes easy when you have sentence-transformers installed:

pip install -U sentence-transformers
pip install -q pyvi

Then you can use the model like this:

from sentence_transformers import SentenceTransformer
from pyvi.ViTokenizer import tokenize

sentences = ["Hà Nội là thủ đô của Việt Nam", "Đà Nẵng là thành phố du lịch"]
tokenizer_sent = [tokenize(sent) for sent in sentences]

model = SentenceTransformer('dangvantuan/vietnamese-embedding')
embeddings = model.encode(tokenizer_sent)
print(embeddings)

Evaluation

The model can be evaluated as follows on the Vienamese data of stsb.

from sentence_transformers import SentenceTransformer
from sentence_transformers import SentenceTransformer
from sentence_transformers.readers import InputExample
from datasets import load_dataset
from pyvi.ViTokenizer import tokenize
def convert_dataset(dataset):
    dataset_samples=[]
    for df in dataset:
        score = float(df['score'])/5.0  # Normalize score to range 0 ... 1
        inp_example = InputExample(texts=[tokenize(df['sentence1']), 
                                    tokenize(df['sentence2'])], label=score)
        dataset_samples.append(inp_example)
    return dataset_samples

# Loading the dataset for evaluation
vi_sts = load_dataset("doanhieung/vi-stsbenchmark")["train"]
df_dev = vi_sts.filter(lambda example: example['split'] == 'dev')
df_test = vi_sts.filter(lambda example: example['split'] == 'test')

# Convert the dataset for evaluation

# For Dev set:
dev_samples = convert_dataset(df_dev)
val_evaluator = EmbeddingSimilarityEvaluator.from_input_examples(dev_samples, name='sts-dev')
val_evaluator(model, output_path="./")

# For Test set:
test_samples = convert_dataset(df_test)
test_evaluator = EmbeddingSimilarityEvaluator.from_input_examples(test_samples, name='sts-test')
test_evaluator(model, output_path="./")

Test Result:

The performance is measured using Pearson and Spearman correlation:

• On dev

  Model	Pearson correlation	Spearman correlation	#params
  dangvantuan/vietnamese-embedding	88.33	88.20	135M
  VoVanPhuc/sup-SimCSE-VietNamese-phobert-base	84.65	84.59	135M
  keepitreal/vietnamese-sbert	84.51	84.44	135M
  bkai-foundation-models/vietnamese-bi-encoder	78.05	77.94	135M

Metric for all dataset of Semantic Textual Similarity on STS Benchmark

You can run an evaluation on this Colab

Pearson score

Spearman score

Citation

@article{reimers2019sentence,
   title={Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks},
   author={Nils Reimers, Iryna Gurevych},
   journal={https://arxiv.org/abs/1908.10084},
   year={2019}
}


@article{martin2020camembert,
   title={CamemBERT: a Tasty French Language Mode},
   author={Martin, Louis and Muller, Benjamin and Su{\'a}rez, Pedro Javier Ortiz and Dupont, Yoann and Romary, Laurent and de la Clergerie, {\'E}ric Villemonte and Seddah, Djam{\'e} and Sagot, Beno{\^\i}t},
   journal={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
   year={2020}
}
@article{thakur2020augmented,
  title={Augmented SBERT: Data Augmentation Method for Improving Bi-Encoders for Pairwise Sentence Scoring Tasks},
  author={Thakur, Nandan and Reimers, Nils and Daxenberger, Johannes and Gurevych, Iryna},
  journal={arXiv e-prints},
  pages={arXiv--2010},
  year={2020}