Edit model card

The SwissBERT model was finetuned via self-supervised SimCSE (Gao et al., EMNLP 2021) for sentence embeddings, using ~1.5 million Swiss news articles from up to 2023 (retrieved via Swissdox@LiRI). Following the Sentence Transformers approach (Reimers and Gurevych, 2019), the average of the last hidden states (pooler_type=avg) is used as sentence representation.


Model Details

Model Description

  • Developed by: Juri Grosjean
  • Model type: XMOD
  • Language(s) (NLP): de_CH, fr_CH, it_CH, rm_CH
  • License: Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)
  • Finetuned from model: SwissBERT


import torch

from transformers import AutoModel, AutoTokenizer

# Load swissBERT for sentence embeddings model
model_name = "jgrosjean-mathesis/sentence-swissbert"
model = AutoModel.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

def generate_sentence_embedding(sentence, language):

    # Set adapter to specified language
    if "de" in language:
    if "fr" in language:
    if "it" in language:
    if "rm" in language:

    # Tokenize input sentence
    inputs = tokenizer(sentence, padding=True, truncation=True, return_tensors="pt", max_length=512)

    # Take tokenized input and pass it through the model
    with torch.no_grad():
        outputs = model(**inputs)

    # Extract sentence embeddings via mean pooling
    token_embeddings = outputs.last_hidden_state
    attention_mask = inputs['attention_mask'].unsqueeze(-1).expand(token_embeddings.size()).float()
    sum_embeddings = torch.sum(token_embeddings * attention_mask, 1)
    sum_mask = torch.clamp(attention_mask.sum(1), min=1e-9)
    embedding = sum_embeddings / sum_mask

    return embedding

# Try it out
sentence_0 = "Wir feiern am 1. August den Schweizer Nationalfeiertag."
sentence_0_embedding = generate_sentence_embedding(sentence_0, language="de")


tensor([[ 5.6306e-02, -2.8375e-01, -4.1495e-02,  7.4393e-02, -3.1552e-01,
          1.5213e-01, -1.0258e-01,  2.2790e-01, -3.5968e-02,  3.1769e-01,
          1.9354e-01,  1.9748e-02, -1.5236e-01, -2.2657e-01,  1.3345e-02,

Semantic Textual Similarity

from sklearn.metrics.pairwise import cosine_similarity

# Define two sentences
sentence_1 = ["Der Zug kommt um 9 Uhr in Zürich an."]
sentence_2 = ["Le train arrive à Lausanne à 9h."]

# Compute embedding for both
embedding_1 = generate_sentence_embedding(sentence_1, language="de")
embedding_2 = generate_sentence_embedding(sentence_2, language="fr")

# Compute cosine-similarity
cosine_score = cosine_similarity(embedding_1, embedding_2)

# Output the score
print("The cosine score for", sentence_1, "and", sentence_2, "is", cosine_score)


The cosine score for ['Der Zug kommt um 9 Uhr in Zürich an.'] and ['Le train arrive à Lausanne à 9h.'] is [[0.85555995]]

Bias, Risks, and Limitations

The sentence swissBERT model has been trained on news articles only. Hence, it might not perform as well on other text classes. Furthermore, it is specific to a Switzerland-related context, which means it probably does not perform as well on text that does not fall in that category. Additionally, the model has neither been trained nor evaluated for machine translation tasks.

Training Details

Training Data

German, French, Italian and Romansh documents in the Swissdox@LiRI database up to 2023.

Training Procedure

This model was finetuned via self-supervised SimCSE. The positive sequence pairs consist of the article body vs. its title and lead, wihout any hard negatives.

The fine-tuning script can be accessed here.

Training Hyperparameters

  • Number of epochs: 1
  • Learning rate: 1e-5
  • Batch size: 512
  • Temperature: 0.05


Testing Data

The two evaluation tasks make use of the 20 Minuten dataset compiled by Kew et al. (2023), which contains Swiss news articles with topic tags and summaries. Parts of the dataset were automatically translated to French, Italian using a Google Cloud API and to Romash via a Textshuttle API.

Evaluation via Document Retrieval

Embeddings are computed for the summary and content of each document. Subsequently, the embeddings are matched by maximizing cosine similarity scores between each summary and content embedding pair.

The performance is measured via accuracy, i.e. the ratio of correct vs. total matches. The script can be found here.

Evaluation via Text Classification

A number of articles with defined topic tags are mapped to 10 categories, filtered from the corpus and split into training data (80%) and test data (20%). Subsequently, embeddings are set up for the train and test data. The test data is then classified using the training data via a k-nearest neighbors approach. The script can be found here.

Note: For French, Italian and Romansh, the training data remains in German, while the test data comprises of translations. This provides insights in the model's abilities in cross-lingual transfer.


Sentence SwissBERT achieves comparable or better results as the best-performing multilingual Sentence-BERT model in these tasks (distiluse-base-multilingual-cased). It outperforms it in all evaluation task, except for the text classification in Italian.

Evaluation task Swissbert Sentence Swissbert Sentence-BERT
accuracy f1-score accuracy f1-score accuracy f1-score
Document Retrieval DE 87.20 % -- 93.40 % -- 91.80 % --
Document Retrieval FR 84.97 % -- 93.99 % -- 93.19 % --
Document Retrieval IT 84.17 % -- 92.18 % -- 91.58 % --
Document Retrieval RM 83.17 % -- 91.58 % -- 73.35 % --
Text Classification DE -- 77.93 % -- 78.49 % -- 77.23 %
Text Classification FR -- 69.62 % -- 77.18 % -- 76.83 %
Text Classification IT -- 67.09 % -- 76.65 % -- 76.90 %
Text Classification RM -- 43.79 % -- 77.20 % -- 65.35 %


The baseline uses mean pooling embeddings from the last hidden state of the original swissbert model and (in these tasks) best-performing Sentence-BERT model distiluse-base-multilingual-cased-v1.

Downloads last month
Inference API
Inference API (serverless) does not yet support transformers models for this pipeline type.

Space using jgrosjean-mathesis/sentence-swissbert 1