metadata

title: relation_extraction
datasets:
  - none
tags:
  - evaluate
  - metric
description: >-
  This metric is used for evaluating the F1 accuracy of input references and
  predictions.
sdk: gradio
sdk_version: 3.19.1
app_file: app.py
pinned: false
license: apache-2.0

Metric Card for relation_extraction evalutation

This metric is used for evaluating the quality of relation extraction output. By calculating the Micro and Macro F1 score of every relation extraction outputs to ensure the quality.

Metric Description

This metric can be used in relation extraction evaluation.

How to Use

This metric takes 2 inputs, prediction and references(ground truth). Both of them are a list of list of dictionary of entity's name and entity's type:

import evaluate
metric_path = "Ikala-allen/relation_extraction"
module = evaluate.load(metric_path)
references = [
  [
    {"head": "phip igments", "head_type": "brand", "type": "sell", "tail": "國際認證之色乳", "tail_type": "product"},
    {"head": "tinadaviespigments", "head_type": "brand", "type": "sell", "tail": "國際認證之色乳", "tail_type": "product"},
  ]
]
predictions = [
  [
    {"head": "phipigments", "head_type": "product", "type": "sell", "tail": "國際認證之色乳", "tail_type": "product"},
    {"head": "tinadaviespigments", "head_type": "brand", "type": "sell", "tail": "國際認證之色乳", "tail_type": "product"},
  ]
]
evaluation_scores = module.compute(predictions=predictions, references=references, mode = "strict")

Inputs

predictions (list of list of dictionary): a list of list of dictionary with every element's relation_type and their entity name
references (list of list of dictionary): a list of list of dictionary with every element's relation_type and their entity name
mode (str): define strict or boundaries mode for evaluation, strict mode consider "head_type" and "tail_type", boundaries mode doesn't consider "head_type" and "tail_type"
only_all (bool): True for only output ["ALL"] relation_type score. False for output every relation_type score, default True
relation_types (list): define relation type that need to be evaluate, if not given, it will construct relation_types from ground truth, default []

Output Values

output (dictionary of dictionarys) with multiple key-value pairs

sell (dictionary): score of type sell
- tp : true positive count
- fp : false positive count
- fn : false negative count
- p : precision
- r : recall
- f1 : micro f1 score
ALL (dictionary): score of all of the type (sell and belongs to)
- tp : true positive count
- fp : false positive count
- fn : false negative count
- p : precision
- r : recall
- f1 : micro f1 score
- Macro_f1 : macro f1 score
- Macro_p : macro precision
- Macro_r : macro recall

Output Example:

{'tp': 1, 'fp': 1, 'fn': 1, 'p': 50.0, 'r': 50.0, 'f1': 50.0, 'Macro_f1': 50.0, 'Macro_p': 50.0, 'Macro_r': 50.0}

Remind : Macro_f1、Macro_p、Macro_r、p、r、f1 are always a number between 0 and 1. And tp、fp、fn depend on how many data inputs.

Examples

Example1 : only one prediction and reference, mode = strict, consider only output ALL relation score

metric_path = "Ikala-allen/relation_extraction"
module = evaluate.load(metric_path)
references = [
  [
    {"head": "phipigments", "head_type": "brand", "type": "sell", "tail": "國際認證之色乳", "tail_type": "product"},
    {"head": "tinadaviespigments", "head_type": "brand", "type": "sell", "tail": "國際認證之色乳", "tail_type": "product"},
    {'head': 'A醛賦活緊緻精華', 'tail': 'Serum', 'head_type': 'product', 'tail_type': 'category', 'type': 'belongs_to'},
  ]
]

# Example references (ground truth)
predictions = [
  [
    {"head": "phipigments", "head_type": "product", "type": "sell", "tail": "國際認證之色乳", "tail_type": "product"},
    {"head": "tinadaviespigments", "head_type": "brand", "type": "sell", "tail": "國際認證之色乳", "tail_type": "product"},
  ]
]

# Calculate evaluation scores using the loaded metric
evaluation_scores = module.compute(predictions=predictions, references=references, mode = "strict", only_all=True,relation_types = [])

print(evaluation_scores)
>>> {'tp': 1, 'fp': 1, 'fn': 2, 'p': 50.0, 'r': 33.333333333333336, 'f1': 40.0, 'Macro_f1': 25.0, 'Macro_p': 25.0, 'Macro_r': 25.0}

Example2 : only one prediction and reference, mode = boundaries, consider only output ALL relation score

metric_path = "Ikala-allen/relation_extraction"
module = evaluate.load(metric_path)
references = [
  [
    {"head": "phipigments", "head_type": "brand", "type": "sell", "tail": "國際認證之色乳", "tail_type": "product"},
    {"head": "tinadaviespigments", "head_type": "brand", "type": "sell", "tail": "國際認證之色乳", "tail_type": "product"},
    {'head': 'A醛賦活緊緻精華', 'tail': 'Serum', 'head_type': 'product', 'tail_type': 'category', 'type': 'belongs_to'},
  ]
]

# Example references (ground truth)
predictions = [
  [
    {"head": "phipigments", "head_type": "product", "type": "sell", "tail": "國際認證之色乳", "tail_type": "product"},
    {"head": "tinadaviespigments", "head_type": "brand", "type": "sell", "tail": "國際認證之色乳", "tail_type": "product"},
  ]
]

# Calculate evaluation scores using the loaded metric
evaluation_scores = module.compute(predictions=predictions, references=references, mode = "strict", only_all=True,relation_types = [])

print(evaluation_scores)
>>> {'tp': 2, 'fp': 0, 'fn': 1, 'p': 100.0, 'r': 66.66666666666667, 'f1': 80.0, 'Macro_f1': 50.0, 'Macro_p': 50.0, 'Macro_r': 50.0}

Example with two or more prediction and reference:

>>> metric_path = "Ikala-allen/relation_extraction"
>>> module = evaluate.load(metric_path)
>>> references = [
...     [
...         {"head": "phip igments", "head_type": "brand", "type": "sell", "tail": "國際認證之色乳", "tail_type": "product"},
...         {"head": "tinadaviespigments", "head_type": "brand", "type": "sell", "tail": "國際認證之色乳", "tail_type": "product"},
...     ],[
...           {'head': 'SABONTAIWAN', 'tail': '大馬士革玫瑰有機光燦系列', 'head_type': 'brand', 'tail_type': 'product', 'type': 'sell'}
...     ]
...   ]
>>> predictions = [
...    [
...        {"head": "phipigments", "head_type": "product", "type": "sell", "tail": "國際認證之色乳", "tail_type": "product"},
...        {"head": "tinadaviespigments", "head_type": "brand", "type": "sell", "tail": "國際認證之色乳", "tail_type": "product"},
...    ],[
...          {'head': 'SABONTAIWAN', 'tail': '大馬士革玫瑰有機光燦系列', 'head_type': 'brand', 'tail_type': 'product', 'type': 'sell'},
...          {'head': 'SNTAIWAN', 'tail': '大馬士革玫瑰有機光燦系列', 'head_type': 'brand', 'tail_type': 'product', 'type': 'sell'}  
...    ]
...  ]
>>> evaluation_scores = module.compute(predictions=predictions, references=references)  
>>> print(evaluation_scores)  
{'sell': {'tp': 2, 'fp': 2, 'fn': 1, 'p': 50.0, 'r': 66.66666666666667, 'f1': 57.142857142857146}, 'ALL': {'tp': 2, 'fp': 2, 'fn': 1, 'p': 50.0, 'r': 66.66666666666667, 'f1': 57.142857142857146, 'Macro_f1': 57.142857142857146, 'Macro_p': 50.0, 'Macro_r': 66.66666666666667}}

Limitations and Bias

This metric has strict filter mechanism, if any of the prediction's entity names, such as head, head_type, type, tail, or tail_type, is not exactly the same as the reference one. It will count as fp or fn.

Citation

@Paper{
    author = {Bruno Taillé, Vincent Guigue, Geoffrey Scoutheeten, Patrick Gallinari},
    title = {Let's Stop Incorrect Comparisons in End-to-end Relation Extraction!},
    year = {2020},
    link = https://arxiv.org/abs/2009.10684
}

Further References

This evaluation metric implementation uses https://github.com/btaille/sincere/blob/master/code/utils/evaluation.py