Model Card for RobeCzech

If you are having issues with the tokenizer, please see https://huggingface.co/ufal/robeczech-base/discussions/4#64b8f6a7f1f8e6ea5860b314.

Model Details

Model Description

RobeCzech is a monolingual RoBERTa language representation model trained on Czech data.

Uses

Direct Use

Fill-Mask tasks.

Downstream Use

Morphological tagging and lemmatization, dependency parsing, named entity recognition, and semantic parsing.

Bias, Risks, and Limitations

Significant research has explored bias and fairness issues with language models (see, e.g., Sheng et al. (2021) and Bender et al. (2021)). Predictions generated by the model may include disturbing and harmful stereotypes across protected classes; identity characteristics; and sensitive, social, and occupational groups.

Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.

Training Details

Training Data

The model creators note in the associated paper:

We trained RobeCzech on a collection of the following publicly available texts:

  • SYN v4, a large corpus of contemporary written Czech, 4,188M tokens;
  • Czes, a collection of Czech newspaper and magazine articles, 432M tokens;
  • documents with at least 400 tokens from the Czech part of the web corpus.W2C , tokenized with MorphoDiTa, 16M tokens;
  • plain texts extracted from Czech Wikipedia dump 20201020 using WikiEx-tractor, tokenized with MorphoDiTa, 123M tokens

All these corpora contain whole documents, even if the SYN v4 is block-shuffled (blocks with at most 100 words respecting sentence boundaries are permuted in a document) and in total contain 4,917M tokens.

Training Procedure

Preprocessing

The texts are tokenized into subwords with a byte-level BPE (BBPE) tokenizer, which was trained on the entire corpus and we limit its vocabulary size to 52,000 items.

Speeds, Sizes, Times

The model creators note in the associated paper:

The training batch size is 8,192 and each training batch consists of sentences sampled contiguously, even across document boundaries, such that the total length of each sample is at most 512 tokens (FULL-SENTENCES setting). We use Adam optimizer with β1 = 0.9 and β2 = 0.98 to minimize the masked language-modeling objective.

Software Used

The Fairseq implementation was used for training.

Evaluation

Testing Data, Factors & Metrics

Testing Data

The model creators note in the associated paper:

We evaluate RobeCzech in five NLP tasks, three of them leveraging frozen contextualized word embeddings, two approached with fine-tuning:

  • morphological analysis and lemmatization: frozen contextualized word embeddings,
  • dependency parsing: frozen contextualized word embeddings,
  • named entity recognition: frozen contextualized word embeddings,
  • semantic parsing: fine-tuned,
  • sentiment analysis: fine-tuned.

Results

Model Morphosynt PDT3.5 (POS) (LAS) Morphosynt UD2.3 (XPOS) (LAS) NER CNEC1.1 (nested) (flat) Semant. PTG (Avg) (F1)
RobeCzech 98.50 91.42 98.31 93.77 87.82 87.47 92.36 80.13

Environmental Impact

  • Hardware Type: 8 QUADRO P5000 GPU
  • Hours used: 2190 (~3 months)

Citation

@InProceedings{10.1007/978-3-030-83527-9_17,
  author={Straka, Milan and N{\'a}plava, Jakub and Strakov{\'a}, Jana and Samuel, David},
  editor={Ek{\v{s}}tein, Kamil and P{\'a}rtl, Franti{\v{s}}ek and Konop{\'i}k, Miloslav},
  title={{RobeCzech: Czech RoBERTa, a Monolingual Contextualized Language Representation Model}},
  booktitle="Text, Speech, and Dialogue",
  year="2021",
  publisher="Springer International Publishing",
  address="Cham",
  pages="197--209",
  isbn="978-3-030-83527-9"
}

How to Get Started with the Model

Use the code below to get started with the model.

Click to expand
from transformers import AutoTokenizer, AutoModelForMaskedLM

tokenizer = AutoTokenizer.from_pretrained("ufal/robeczech-base")

model = AutoModelForMaskedLM.from_pretrained("ufal/robeczech-base")
Downloads last month
9
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.