nervaluate

nervaluate is a python module for evaluating Named Entity Recognition (NER) models as defined in the SemEval 2013 - 9.1 task.

The evaluation metrics output by nervaluate go beyond a simple token/tag based schema, and consider different scenarios based on weather all the tokens that belong to a named entity were classified or not, and also whether the correct entity type was assigned.

This full problem is described in detail in the original blog post by David Batista, and extends the code in the original repository which accompanied the blog post.

The code draws heavily on:

• Segura-bedmar, I., & Mart, P. (2013). 2013 SemEval-2013 Task 9 Extraction of Drug-Drug Interactions from. Semeval, 2(DDIExtraction), 341–350. link
• https://www.cs.york.ac.uk/semeval-2013/task9/data/uploads/semeval_2013-task-9_1-evaluation-metrics.pdf

The problem

Token level evaluation for NER is too simplistic

When running machine learning models for NER, it is common to report metrics at the individual token level. This may not be the best approach, as a named entity can be made up of multiple tokens, so a full-entity accuracy would be desirable.

When comparing the golden standard annotations with the output of a NER system different scenarios might occur:

I. Surface string and entity type match

II. System hypothesized an incorrect entity

III. System misses an entity

Based on these three scenarios we have a simple classification evaluation that can be measured in terms of false positives, true positives, false negatives and false positives, and subsequently compute precision, recall and F1-score for each named-entity type.

However, this simple schema ignores the possibility of partial matches or other scenarios when the NER system gets the named-entity surface string correct but the type wrong, and we might also want to evaluate these scenarios again at a full-entity level.

For example:

IV. System assigns the wrong entity type

V. System gets the boundaries of the surface string wrong

VI. System gets the boundaries and entity type wrong

How can we incorporate these described scenarios into evaluation metrics? See the original blog for a great explanation, a summary is included here:

We can use the following five metrics to consider difference categories of errors:

These five metrics can be measured in four different ways:

These five errors and four evaluation schema interact in the following ways:

Then precision/recall/f1-score are calculated for each different evaluation schema. In order to achieve data, two more quantities need to be calculated:

POSSIBLE (POS) = COR + INC + PAR + MIS = TP + FN
ACTUAL (ACT) = COR + INC + PAR + SPU = TP + FP

Then we can compute precision/recall/f1-score, where roughly describing precision is the percentage of correct named-entities found by the NER system, and recall is the percentage of the named-entities in the golden annotations that are retrieved by the NER system. This is computed in two different ways depending on whether we want an exact match (i.e., strict and exact ) or a partial match (i.e., partial and type) scenario:

Exact Match (i.e., strict and exact )

Precision = (COR / ACT) = TP / (TP + FP)
Recall = (COR / POS) = TP / (TP+FN)

Partial Match (i.e., partial and type)

Precision = (COR + 0.5 × PAR) / ACT = TP / (TP + FP)
Recall = (COR + 0.5 × PAR)/POS = COR / ACT = TP / (TP + FN)

Putting all together:

Notes:

In scenarios IV and VI the entity type of the true and pred does not match, in both cases we only scored against the true entity, not the predicted one. You can argue that the predicted entity could also be scored as spurious, but according to the definition of spurious:

• Spurious (SPU) : system produces a response which does not exist in the golden annotation;

In this case there exists an annotation, but with a different entity type, so we assume it's only incorrect.

Installation

pip install nervaluate

Example:

The main Evaluator class will accept a number of formats:

• prodi.gy style lists of spans.
• Nested lists containing NER labels.
• CoNLL style tab delimited strings.

Prodigy spans

true = [
    [{"label": "PER", "start": 2, "end": 4}],
    [{"label": "LOC", "start": 1, "end": 2},
     {"label": "LOC", "start": 3, "end": 4}]
]

pred = [
    [{"label": "PER", "start": 2, "end": 4}],
    [{"label": "LOC", "start": 1, "end": 2},
     {"label": "LOC", "start": 3, "end": 4}]
]

from nervaluate import Evaluator

evaluator = Evaluator(true, pred, tags=['LOC', 'PER'])

# Returns overall metrics and metrics for each tag

results, results_per_tag, result_indices, result_indices_by_tag = evaluator.evaluate()

print(results)

{
    'ent_type':{
        'correct':3,
        'incorrect':0,
        'partial':0,
        'missed':0,
        'spurious':0,
        'possible':3,
        'actual':3,
        'precision':1.0,
        'recall':1.0
    },
    'partial':{
        'correct':3,
        'incorrect':0,
        'partial':0,
        'missed':0,
        'spurious':0,
        'possible':3,
        'actual':3,
        'precision':1.0,
        'recall':1.0
    },
    'strict':{
        'correct':3,
        'incorrect':0,
        'partial':0,
        'missed':0,
        'spurious':0,
        'possible':3,
        'actual':3,
        'precision':1.0,
        'recall':1.0
    },
    'exact':{
        'correct':3,
        'incorrect':0,
        'partial':0,
        'missed':0,
        'spurious':0,
        'possible':3,
        'actual':3,
        'precision':1.0,
        'recall':1.0
    }
}

print(results_by_tag)

{
    'LOC':{
        'ent_type':{
            'correct':2,
            'incorrect':0,
            'partial':0,
            'missed':0,
            'spurious':0,
            'possible':2,
            'actual':2,
            'precision':1.0,
            'recall':1.0
        },
        'partial':{
            'correct':2,
            'incorrect':0,
            'partial':0,
            'missed':0,
            'spurious':0,
            'possible':2,
            'actual':2,
            'precision':1.0,
            'recall':1.0
        },
        'strict':{
            'correct':2,
            'incorrect':0,
            'partial':0,
            'missed':0,
            'spurious':0,
            'possible':2,
            'actual':2,
            'precision':1.0,
            'recall':1.0
        },
        'exact':{
            'correct':2,
            'incorrect':0,
            'partial':0,
            'missed':0,
            'spurious':0,
            'possible':2,
            'actual':2,
            'precision':1.0,
            'recall':1.0
        }
    },
    'PER':{
        'ent_type':{
            'correct':1,
            'incorrect':0,
            'partial':0,
            'missed':0,
            'spurious':0,
            'possible':1,
            'actual':1,
            'precision':1.0,
            'recall':1.0
        },
        'partial':{
            'correct':1,
            'incorrect':0,
            'partial':0,
            'missed':0,
            'spurious':0,
            'possible':1,
            'actual':1,
            'precision':1.0,
            'recall':1.0
        },
        'strict':{
            'correct':1,
            'incorrect':0,
            'partial':0,
            'missed':0,
            'spurious':0,
            'possible':1,
            'actual':1,
            'precision':1.0,
            'recall':1.0
        },
        'exact':{
            'correct':1,
            'incorrect':0,
            'partial':0,
            'missed':0,
            'spurious':0,
            'possible':1,
            'actual':1,
            'precision':1.0,
            'recall':1.0
        }
    }
}

from nervaluate import summary_report_overall_indices

print(summary_report_overall_indices(evaluation_indices=result_indices, error_schema='partial', preds=pred))

Indices for error schema 'partial':

Correct indices:
  - Instance 0, Entity 0: Label=PER, Start=2, End=4
  - Instance 1, Entity 0: Label=LOC, Start=1, End=2
  - Instance 1, Entity 1: Label=LOC, Start=3, End=4

Incorrect indices:
  - None

Partial indices:
  - None

Missed indices:
  - None

Spurious indices:
  - None

Nested lists

true = [
    ['O', 'O', 'B-PER', 'I-PER', 'O'],
    ['O', 'B-LOC', 'I-LOC', 'B-LOC', 'I-LOC', 'O'],
]

pred = [
    ['O', 'O', 'B-PER', 'I-PER', 'O'],
    ['O', 'B-LOC', 'I-LOC', 'B-LOC', 'I-LOC', 'O'],
]

evaluator = Evaluator(true, pred, tags=['LOC', 'PER'], loader="list")

results, results_by_tag, result_indices, result_indices_by_tag = evaluator.evaluate()

CoNLL style tab delimited

true = "word\tO\nword\tO\B-PER\nword\tI-PER\n"

pred = "word\tO\nword\tO\B-PER\nword\tI-PER\n"

evaluator = Evaluator(true, pred, tags=['PER'], loader="conll")

results, results_by_tag, result_indices, result_indices_by_tag = evaluator.evaluate()

Extending the package to accept more formats

Additional formats can easily be added to the module by creating a conversion function in nervaluate/utils.py, for example conll_to_spans(). This function must return the spans in the prodigy style dicts shown in the prodigy example above.

The new function can then be added to the list of loaders in nervaluate/nervaluate.py, and can then be selection with the loader argument when instantiating the Evaluator class.

A list of formats we intend to include is included in https://github.com/ivyleavedtoadflax/nervaluate/issues/3.

Contributing to the nervaluate package

Improvements, adding new features and bug fixes are welcome. If you wish to participate in the development of nervaluate please read the following guidelines.

The contribution process at a glance

1. Preparing the development environment
2. Code away!
3. Continuous Integration
4. Submit your changes by opening a pull request

Small fixes and additions can be submitted directly as pull requests, but larger changes should be discussed in an issue first. You can expect a reply within a few days, but please be patient if it takes a bit longer.

Preparing the development environment

Make sure you have Python3.8 installed on your system

macOs

brew install python@3.8
python3.8 -m pip install --user --upgrade pip
python3.8 -m pip install virtualenv

Clone the repository and prepare the development environment:

git clone git@github.com:MantisAI/nervaluate.git
cd nervaluate
python3.8 -m virtualenv venv
source venv/bin/activate
pip install -r requirements_dev.txt
pip install -e .

Continuous Integration

nervaluate runs a continuous integration (CI) on all pull requests. This means that if you open a pull request (PR), a full test suite is run on your PR:

• The code is formatted using black
• Linting is done using pyling and flake8
• Type checking is done using mypy
• Tests are run using pytest

Nevertheless, if you prefer to run the tests & formatting locally, it's possible too.

make all

Opening a Pull Request

Every PR should be accompanied by short description of the changes, including:

• Impact and motivation for the changes
• Any open issues that are closed by this PR

Give a ⭐️ if this project helped you!