metadata

title: distinct
datasets:
  - None
tags:
  - evaluate
  - measurement
description: 'TODO: add a description here'
sdk: gradio
sdk_version: 3.19.1
app_file: app.py
pinned: false

Measurement Card for distinct

Module Card Instructions:

Measurement Description

This metric is used to calculate the diversity of a group of sentences. It can be used to either evaluate the diversity of generated responses of the testset (i.e., corpus level diversity), or calculate diversity of a group of sampled responses given one context (i.e., utterence level diversity). The original paper (Li et al. 2022) used it as corpus-level while some may use it as utterance-level. However, we don't recommend to calculate Distinct on a small group as it is sensitive to the sentence length and number.

How to Use

>>> import evaluate
>>> results = my_new_module.compute(predictions=["Hi.", "I am sorry to hear that", "I don't know", "Do you know who that person is?"], vocab
_size=50257)
>>> my_new_module = evaluate.load("lsy641/distinct")
Downloading builder script: 100%|██████████████████████████████████████████████████████████████████████| 8.62k/8.62k [00:00<00:00, 4.19MB/s]
>>> results = my_new_module.compute(predictions=["Hi.", "I am sorry to hear that", "I don't know", "Do you know who that person is?"], vocab_size=50257)
>>> print(results)
{'Expectation-Adjusted-Distinct': 0.8236605104867569, 'Distinct-1': 0.8235294117647058, 'Distinct-2': 0.9411764705882353, 'Distinct-3': 0.9411764705882353}


>>> dataset = ["This is my friend jack", "I'm sorry to hear that", "But you know I am the one who always support you", "Welcome to our family","Hi.", "I am sorry to hear that", "I don't know", "Do you know who that person is?"]
>>> results = my_new_module.compute(predictions=["But you know I am the one who always support you", "Hi.", "I am sorry to hear that", "I don't know", "I'm sorry to hear that"], dataForVocabCal=dataset)
>>> print(results)
{'Expectation-Adjusted-Distinct': 0.9928137111900845, 'Distinct-1': 0.6538461538461539, 'Distinct-2': 0.8076923076923077, 'Distinct-3': 0.8846153846153846}

Inputs

List all input arguments in the format below

predictions (list of strings): list of sentences to test diversity. Each prediction should be a string.
mode (string): 'Expectation-Adjusted-Distinct' or 'Distinct' for diversity calculation. If 'Expectation-Adjusted-Distinct', the scores for both modes will be returned. The default value is 'Expectation-Adjusted-Distinct'
vocab_size (int): For calculating 'Expectation-Adjusted-Distinct', either vocab_size or dataForVocabCal should not be None. Default value is None
dataForVocabCal (list of string): dataForVocabCal for calculating the vocab_size for 'Expectation-Adjusted-Distinct'. Typically, it should be a list of sentences consisting the task dataset. For calculating 'Expectation-Adjusted-Distinct', either vocab_size or dataForVocabCal should not be None. Default value is None
tokenizer (string or tokenizer class): tokenizer for splitting sentences into words. Default value is "white_space". NLTK tokenizer is available.

Output Values

Expectation-Adjusted-Distinct: Normally it should stay in range 0-1. But it can be more than 1. See the formula property in the Expectation-Adjusted-Distinct paper (Liu and Sabour et al. 2022)
Distinct-1: Range 0-1
Distinct-2: Range 0-1
Distinct-3: Range 0-1

Values from Popular Papers

The Expectation-Adjusted-Distinct paper (Liu and Sabour et al. 2022) compares Expectation-Adjusted-Distinct scores of ten different methods with the original Distinct. These scores get higher human correlation from 0.56 to 0.65.

Examples

Example of calculate Expectation-Adjusted-Distinct byy giving voab_size or data for calculating vocab_size. This will also return Distinct-1,2,and 3.

    >>> my_new_module = evaluate.load("lsy641/distinct")
    >>> results = my_new_module.compute(references=["Hi.", "I'm sorry to hear that", "I don't know"], vocab_size=50257)
    >>> print(results)
    
    
    >>> dataset = ["This is my friend jack", "I'm sorry to hear that", "But you know I am the one who always support you", "Welcome to our family"]
    >>> results = my_new_module.compute(references=["Hi.", "I'm sorry to hear that", "I don't know"], dataForVocabCal = dataset)
    >>> print(results)

Example of calculate original Distinct. This will return Distinct-1,2,and 3.

    >>> my_new_module = evaluate.load("lsy641/distinct")   
    >>> results = my_new_module.compute(references=["Hi.", "I'm sorry to hear that", "I don't know"], mode="Distinct")
    >>> print(results)

Limitations and Bias

TODO

Citation

@inproceedings{liu-etal-2022-rethinking,
    title = "Rethinking and Refining the Distinct Metric",
    author = "Liu, Siyang  and
      Sabour, Sahand  and
      Zheng, Yinhe  and
      Ke, Pei  and
      Zhu, Xiaoyan  and
      Huang, Minlie",
    booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)",
    year = "2022",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.acl-short.86",
    doi = "10.18653/v1/2022.acl-short.86",
    }

@inproceedings{li-etal-2016-diversity,
    title = "A Diversity-Promoting Objective Function for Neural Conversation Models",
    author = "Li, Jiwei  and
      Galley, Michel  and
      Brockett, Chris  and
      Gao, Jianfeng  and
      Dolan, Bill",
    booktitle = "Proceedings of the 2016 Conference of the North {A}merican Chapter of the Association for Computational Linguistics: Human Language Technologies",
    year = "2016",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/N16-1014",
    doi = "10.18653/v1/N16-1014",
}

Further References

TODO