NaSE

NaSE (News-adapted Sentence Encoder)

This model is a news-adapted sentence encoder, domain-specialized starting from the pretrained massively mulitlingual sentence encoder LaBSE.

Model Details

Model Description

NaSE is a domain-adapted multilingual sentence encoder, initialized from LaBSE. It was specialized to the news domain using two multilingual corpora, namely Polynews and PolyNewsParallel. More specifically, NaSE was pretrained with two objectives: denoising auto-encoding and sequence-to-sequence machine translation.

Usage (HuggingFace Transformers)

Here is how to use this model to get the sentence embeddings of a given text in PyTorch:

    from transformers import BertModel, BertTokenizerFast

    tokenizer = BertTokenizerFast.from_pretrained('aiana94/NaSE')
    model = BertModel.from_pretrained('aiana94/NaSE')

    # pepare input
    sentences = ["This is an example sentence", "Dies ist auch ein Beispielsatz in einer anderen Sprache."]
    encoded_input = tokenizer(sentences, return_tensors='pt', padding=True)

    # forward pass
    with torch.no_grad():
        output = model(**encoded_input)

    # to get the sentence embeddings, use the pooler output
    sentence_embeddings = output.pooler_output

and in Tensorflow:

    from transformers import TFBertModel, BertTokenizerFast

    tokenizer = BertTokenizerFast.from_pretrained('aiana94/NaSE')
    model = TFBertModell.from_pretrained('aiana94/NaSE')

    # pepare input
    sentences = ["This is an example sentence", "Dies ist auch ein Beispielsatz in einer anderen Sprache."]
    encoded_input = tokenizer(sentences, return_tensors='tf', padding=True)

    # forward pass
    with torch.no_grad():
        output = model(**encoded_input)

    # to get the sentence embeddings, use the pooler output
    sentence_embeddings = output.pooler_output

For similarity between sentences, an L2-norm is recommended before calculating the similarity:

  import torch
  import torch.nn.functional as F

  def cos_sim(a: torch.Tensor, b: torch.Tensor):
    a_norm = F.normalize(a, p=2, dim=1)
    b_norm = F.normalize(b, p=2, dim=1)

    return torch.mm(a_norm, b_norm.transpose(0, 1))

Intended Uses

Our model is intended to be used as a sentence, and in particular, news encoder. Given an input text, it outputs a vector which captures its semantic information. The sentence vector may be used for sentence similarity, information retrieval or clustering tasks.

Training Details

Training Data

NaSE was domain-adapted using two multilingual datasets: Polynews and the parallel PolyNewsParallel.

We use the following procedure to smoothen the per-language distribution when sampling for model training:

• We sample only languages and language-pairs that contain at least 100 texts in PolyNews and PolyNewsParallel, respectively;
• We sample texts from language L by sampling from the modified distribution p(L) ~ |L| * alpha, where |L| is the number of examples and L. We use a smooting rate alpha=0.3 (i.e., we upsample low-resource languages and downsample high-resource languages).

Training Procedure

We initialize NaSE with the pretrained weights of the mulitlingual sentenece encoder LaBSE. Please refer to its model card or the corresponding paper for more detaled information about the pre-training procedure.

We adapt the multilingual sentence encoder to the news domain using two objectives:

• Denoising auto-encoding (DAE): reconstructs the original input sentence from its corrupted version obtained by adding discrete noise (see TSDAE for details);
• Machine translation (MT): generates the taget-language translation from the source-language input sentence (i.e., the source language constitutes the corruption of the target sentence x in the target language, which is to be reconstructed).

NaSE is trained sequentially, first on reconstruction, and then on translation, i.e., we continue training the NaSE encoder obtained with the DAE objective for translation on parallel data.

Training Hyperparameters

• Training regime: fp16 mixed precision
• Training steps: 100k (50K per objective), validating every 5K steps
• Learning rate: 3e-5
• Optimizer: AdamW

The full training scripts is accessible in the training code.

Technical Specifications

The model was pretrained on 1 40GB NVIDIA A100 GPU for a total of 100k steps.

Citation

BibTeX:

@misc{iana2024news,
      title={News Without Borders: Domain Adaptation of Multilingual Sentence Embeddings for Cross-lingual News Recommendation}, 
      author={Andreea Iana and Fabian David Schmidt and Goran Glavaš and Heiko Paulheim},
      year={2024},
      eprint={2406.12634},
      archivePrefix={arXiv},
      url={https://arxiv.org/abs/2406.12634}
}