Training a Stanford NER Model in Python


April 7, 2022

Stanford NER is a good implementation of a Named Entity Recognizer (NER) using Conditional Random Fields (CRFs). CRFs are no longer near state of the art for NER, having been overtaken by LSTM-CRFs, which have since been overtaken by Transformer models. However CRFs are still a reasonable baseline, and Stanford NER is used in many papers which is good for reproducibility. It doesn’t have a Python binding (a CRF library that does is CRFsuite), but with some work we can train and test a model in Python.

I replicated the benchmark in A Named Entity Based Approach to Model Recipes, by Diwan, Batra, and Bagler using Stanford NER, and check it using seqeval. Evaluating NER is surprisingly tricky, as David Batista explains, and I want to check that the results in the paper are the same as what seqeval gives, so that I’m giving a fair comparison to other models. Thanks to the authors sharing the training data on github I was able to do this, as you can see in the Jupyter Notebook.

The rest of this article goes through how to train and evaluate a Stanford NER model using Python, and that the scores output by Stanford NLP on the test set match those produced by seqeval.

Setting up Stanford NLP

The stanza library has both great neural network based models for linguistic analysis (see my previous writeup), but also an interface to Stanford Core NLP. Unfortunately it doesn’t provide a direct way of training an NER model using Core NLP, however we can do it ourselves using the stanford-corenlp JAR it installs. The first step is to install the models and find the path to the Core NLP JAR.

import os
from pathlib import Path

import stanza

# Reimplement the logic to find the path where stanza_corenlp is installed.
core_nlp_path = os.getenv('CORENLP_HOME', str(Path.home() / 'stanza_corenlp'))

# A heuristic to find the right jar file
classpath = [str(p) for p in Path(core_nlp_path).iterdir()
             if re.match(r"stanford-corenlp-[0-9.]+\.jar",][0]

Train NER Model

The Stanford NER model requires data where each line is a token, followed by a tab, followed by the NER tag. A blank line represents a sentence break. In this case I could get the relevant training and test repository in this format already.

The NER model has to be configured; but there’s no information on the paper on what features and hyperparameters are used. I copied the template configuration out of the FAQ, which happened to work well. The template can be saved to a file and then referred to when training.

def ner_prop_str(train_files: List[str],
                 test_files: List[str],
                 output: str) -> str:
    """Returns configuration string to train NER model"""
    train_file_str = ','.join(train_files)
    test_file_str = ','.join(test_files)
    return f"""
trainFileList = {train_file_str}
testFiles = {test_file_str}
serializeTo = {output}
map = word=0,answer=1


For more information on the parameters you can check the NERFeatureFactory documentation or the source. We need to write this to a file so we can write a wrapper to do this

def write_ner_prop_file(ner_prop_file: str,
                        train_files: List[str],
                        test_files: List[str],
                        output_file: str) -> None:
    with open(ner_prop_file, 'wt') as f:
        props = ner_prop_str(train_files, test_files, output_file)

The actual NER Training process is in Java, so we’ll run a Java process to train a model and return the path to the model file. We’ll also print out the report from stderr summarising the training.

import subprocess
from typing import List

def train_model(model_name: str,
                train_files: List[str],
                test_files: List[str],
                print_report: bool = True,
                classpath: str = classpath) -> str:
    """Trains CRF NER Model using StanfordNLP"""
    model_file = f'{model_name}.model.ser.gz'
    ner_prop_filename = f'{model_name}.model.props'
    write_ner_prop_file(ner_prop_filename, train_files, test_files, model_file)

    result =
                 '-cp', classpath,
                 '-prop', ner_prop_filename],

    if print_report:
        print(*result.stderr.decode('utf-8').split('\n')[-11:], sep='\n')

    return model_file

Running training on the train and test set produced an output like this.

The summary report shows for each model and entity type:

  • True Positives (TP): The number of times that entity was predicted correctly
  • False Positives (FP): The number of times that entity in the text but not predicted correctly
  • False Negative (FN): The number of times that entity was not in the text and predicted
  • Precision (P): Probability a predicted entity is correct, TP/(TP+FP)
  • Recall (R): Probability a correct entity is predicted, TP/(TP+FN)
  • F1 Score (F1): Harmonic mean of precision and recall, 2/(1/P + 1/R).
CRFClassifier tagged 2788 words in 483 documents at 9992.83 words per second.
         Entity P   R   F1  TP  FP  FN
             DF 1.0000  0.9608  0.9800  49  0   2
           NAME 0.9297  0.9279  0.9288  463 35  36
       QUANTITY 1.0000  0.9962  0.9981  522 0   2
           SIZE 1.0000  1.0000  1.0000  20  0   0
          STATE 0.9601  0.9633  0.9617  289 12  11
           TEMP 0.8750  0.7000  0.7778  7   1   3
           UNIT 0.9819  0.9841  0.9830  434 8   7
         Totals 0.9696  0.9669  0.9682  1784    56  61

The Totals F1 score of 0.9682 exactly matched what was reported in the paper. Now let’s try to manually evaluate the test set using seqeval.

Running the model

Stanza has a robust way of running CoreNLP and annotating texts, as per the documentation. We can configure the NER model used to the one that we just trained. Because the text we’ve used is pre-tokenized I’m just going to join them with a space and tokenize on whitespace; for ingredients we want quantities like 1/2 to be treated as a single token but the default tokenizer will split them. When I first ran the annotations it would sometime output NUMBER, which wasn’t an input entity; it turns out this is hardcoded and we have to diable the numeric classifiers.

from stanza.server import CoreNLPClient

def annotate_ner(ner_model_file: str,
                 texts: List[str],
                 tokenize_whitespace: bool = True):
    properties = {"ner.model": ner_model_file,
                  "tokenize.whitespace": tokenize_whitespace,
                  "ner.applyNumericClassifiers": False}

    annotated = []
    with CoreNLPClient(
        memory='6G') as client:

        for text in texts:
    return annotated

The annotated data will have many attributes, but we’re just interested in the input words and named entities so we’ll extract them into a dictionary. Note that we extract the coarseNER; sometimes another default NER model predicts a fine grained NER (like NATIONALITY) which writes into the ner attribute if it’s empty. Using coarseNER means we only get tags from our training set.

def extract_ner_data(annotation) -> Dict[str, List[str]]:
    tokens = [token for sentence in annotation.sentence
                    for token in sentence.token]
    return {'tokens': [t.word for t in tokens],
            'ner': [t.coarseNER for t in tokens]}

def ner_extract(ner_model_file: str,
                texts: List[str],
                tokenize_whitespace: bool = True) -> List[Dict[str, List[str]]]:
    annotations = annotate_ner(ner_model_file, texts, tokenize_whitespace)
    return [extract_ner_data(ann) for ann in annotations]

Now if we’ve got the test tokens as a list containing lists of words, and the test labels as a list containing corresponding lists of NER tags we can run them through the model.

test_texts = [' '.join(text) for text in test_tokens]
pred_labels = [text['ner'] for text in ner_extract(modelfile, texts)]

Evaluating with seqeval

The library seqeval provides robust sequence labelling metrics. In particular scores should be at an entity level; you don’t get it right unless you predict exactly the tokens in an entity. I wanted to check seqeval gave similar results to the sumamary report above.

Seqeval expects the tags to be in one of the standard tagging formats, but the data I had just had labels (like NAME, QUANTITY, and UNIT). It is impossible to disambiguate adjacent tags of the same entity type, but the annotations mostly assume there can only be one of each kind of entity in an ingredient. The simplest way to convert it is into IOB-1, which only adds a B tag when there are two adjacent tags of the same entity type. Since we’re assuming this doesn’t happen we just need to prepend I- to all tags other than O.

def convert_to_iob1(tokens):
    return ['I-' + label if label != 'O' else 'O' for label in tokens]

result = convert_to_iob1(['QUANTITY', 'SIZE', 'NAME', 'NAME', 'O', 'STATE'])
expected = ['I-QUANTITY', 'I-SIZE', 'I-NAME', 'I-NAME', 'O', 'I-STATE']
assert result == expected

Then we can get the classification report using seqeval

from seqeval.metrics import classification_report

actual_labels = [convert_to_iob1(text) for text in actual_labels]
pred_labels = [convert_to_iob1(text) for text in pred_labels]

print(classification_report(actual_labels, pred_labels, digits=4))

The output report matches the report from Stanford NLP precisely. Note that it uses the support (total number of actual entities) instead of the True Positives, False Positives, and False Negatives, but actually they are equivalent.

  • support = TP + FN
  • TP = R * support
  • FP = TP (1/P - 1)
  • FN = support - TP
              precision    recall  f1-score   support

          DF     1.0000    0.9608    0.9800        51
        NAME     0.9297    0.9279    0.9288       499
    QUANTITY     1.0000    0.9962    0.9981       524
        SIZE     1.0000    1.0000    1.0000        20
       STATE     0.9601    0.9633    0.9617       300
        TEMP     0.8750    0.7000    0.7778        10
        UNIT     0.9819    0.9841    0.9830       441

   micro avg     0.9696    0.9669    0.9682      1845
   macro avg     0.9638    0.9332    0.9471      1845
weighted avg     0.9695    0.9669    0.9682      1845

Using seqeval on all the training and test sets in the paper I could reproduce their f1-scores within 0.01. Below shows my results (left) and the results on the paper (right).


The Stanford NER library is a bit under-documented and has some surprising features, but with some work we can get it to run in Python. The metrics it produces line up with those from seqeval.