metadata

language: tr
tags:
  - roberta
  - language-model
  - scientific
  - turkish
  - fill-mask
license: mit
model_author: Serdar ÇAĞLAR
widget:
  - text: >-
      Sekiz hastada <mask> çıkarılması gerekti (n=5 kateter trombozu, n=2
      kateter enfeksiyonu ve n=1 büyük hematom).
  - text: >-
      Santral <mask> kateterizasyona bağlı süperior vena kava perforasyonun
      video yardımlı torakoskopik cerrahi ile tedavisi
  - text: >-
      Akut lenfoblastik <mask> tanısı sırasında yapılan genetik çalışmalar,
      tedavi yoğunluğunun belirlenmesini ve hedefe yönelik tedavilerin
      planlanmasını sağlar."

🇹🇷

Roberta-Based Language Model Trained on Turkish Scientific Article Abstracts

This model is a powerful natural language processing model trained on Turkish scientific article abstracts. It focuses on scientific content in the Turkish language and excels in tasks related to text comprehension. The model can be used for understanding scientific texts, summarization, and various other natural language processing tasks.

Model Details

Data Source: This model is trained on a custom dataset consisting of Turkish scientific article summaries. The data was collected using web scraping methods from various sources in Turkey, including databases like "trdizin," "yöktez," and "türkiyeklinikleri."
Dataset Preprocessing: The data underwent preprocessing to facilitate better learning. Texts were segmented into sentences, and improperly divided sentences were cleaned. The texts were processed meticulously.
Tokenizer: The model utilizes a BPE (Byte Pair Encoding) tokenizer to process the data effectively, breaking down the text into subword tokens.
Training Details: The model was trained on a large dataset of Turkish sentences. The training spanned 2M Steps, totaling 3+ days, and the model was built from scratch. No fine-tuning was applied.

Usage

Load transformers library with:

from transformers import AutoTokenizer, AutoModelForMaskedLM
  
tokenizer = AutoTokenizer.from_pretrained("serdarcaglar/roberta-base-turkish-scientific-abstract")
model = AutoModelForMaskedLM.from_pretrained("serdarcaglar/roberta-base-turkish-scientific-abstract")

Fill Mask Usage

from transformers import pipeline

fill_mask = pipeline(
    "fill-mask",
    model="serdarcaglar/roberta-base-turkish-scientific-abstract",
    tokenizer="serdarcaglar/roberta-base-turkish-scientific-abstract"
)

fill_mask("İnterarteriyel seyirli anormal <mask> arter hastaları ne zaman ameliyat edilmeli ve hangi cerrahi teknik kullanılmalıdır?")

[{'score': 0.7180812954902649,
  'token': 6252,
  'token_str': ' koroner',
  'sequence': 'İnterarteriyel seyirli anormal koroner arter hastaları ne zaman ameliyat edilmeli ve hangi cerrahi teknik kullanılmalıdır?'},
 {'score': 0.09322144836187363,
  'token': 9978,
  'token_str': ' pulmoner',
  'sequence': 'İnterarteriyel seyirli anormal pulmoner arter hastaları ne zaman ameliyat edilmeli ve hangi cerrahi teknik kullanılmalıdır?'},
 {'score': 0.03268029913306236,
  'token': 16407,
  'token_str': ' uterin',
  'sequence': 'İnterarteriyel seyirli anormal uterin arter hastaları ne zaman ameliyat edilmeli ve hangi cerrahi teknik kullanılmalıdır?'},
 {'score': 0.012145915068686008,
  'token': 12969,
  'token_str': ' renal',
  'sequence': 'İnterarteriyel seyirli anormal renal arter hastaları ne zaman ameliyat edilmeli ve hangi cerrahi teknik kullanılmalıdır?'},
 {'score': 0.011508156545460224,
  'token': 26256,
  'token_str': ' karotis',
  'sequence': 'İnterarteriyel seyirli anormal karotis arter hastaları ne zaman ameliyat edilmeli ve hangi cerrahi teknik kullanılmalıdır?'}]

Disclaimer

The use of this model is subject to compliance with specific copyright and legal regulations, which are the responsibility of the users. The model owner or provider cannot be held liable for any issues arising from the use of the model.

Contact information

For further information, send an email to [email protected]

Serdar ÇAĞLAR.

Important information: MSI RTX 3090 GPU was used for model training. Due to lack of computational resources, the dataset was downsized and I had to interrupt the training process to continue with other projects. I would be grateful if researchers testing the MLM could provide feedback on the success of the model. If I can find the necessary GPU support, I will complete the development of the model.