PhilipMay's picture
Update README.md
55dba4b
metadata
license: cc-by-4.0
language:
  - en
  - de
multilinguality:
  - multilingual
source_datasets:
  - extended|deutsche-telekom/NLU-Evaluation-Data-en-de
size_categories:
  - 1K<n<10K
task_categories:
  - text-classification
task_ids:
  - intent-classification

NLU Few-shot Benchmark - English and German

This is a few-shot training dataset from the domain of human-robot interaction. It contains texts in German and English language with 64 different utterances (classes). Each utterance (class) has exactly 20 samples in the training set. This leads to a total of 1280 different training samples.

The dataset is intended to benchmark the intent classifiers of chat bots in English and especially in German language. We are building on our deutsche-telekom/NLU-Evaluation-Data-en-de data set

Creator

This data set was compiled and open sourced by Philip May of Deutsche Telekom.

Processing Steps

  • drop NaN values
  • drop duplicates in answer_de and answer
  • delete all rows where answer_de has more than 70 characters
  • add column label: df["label"] = df["scenario"] + "_" + df["intent"]
  • remove classes (label) with less than 25 samples:
    • audio_volume_other
    • cooking_query
    • general_greet
    • music_dislikeness
  • random selection for train set - exactly 20 samples for each class (label)
  • rest for test set

Copyright

Copyright (c) the authors of xliuhw/NLU-Evaluation-Data
Copyright (c) 2022 Philip May, Deutsche Telekom AG

All data is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).