Module 5: Named Entity Recognition & Sequence Labeling

This module covers methods for identifying and labeling named entities (e.g. persons, locations, organizations) in text and more general sequence‐labeling tasks.

---

5.1 spaCy Pretrained NER

Use spaCy’s small English model to extract entities out of the box:

import spacy

# 1. Load the spaCy model (install with: pip install spacy && python -m spacy download en_core_web_sm)
nlp = spacy.load('en_core_web_sm')

# 2. Process text
text = "Apple is looking at buying U.K. startup for $1 billion"
doc  = nlp(text)

# 3. Print detected entities
for ent in doc.ents:
    print(ent.text, ent.label_)

Output:

Apple ORG
U.K. GPE
$1 billion MONEY

5.2 CRF-Based Sequence Labeling

Train a Conditional Random Field on toy NER data:

import sklearn_crfsuite
from sklearn_crfsuite import metrics
from collections import defaultdict

# 1. Toy labeled data: list of sentences, each a list of (word, tag)
train_sents = [
    [('John','B-PER'), ('lives','O'), ('in','O'), ('London','B-LOC')],
    [('Mary','B-PER'), ('works','O'), ('at','O'), ('Google','B-ORG')],
]
test_sents = [
    [('Alice','B-PER'), ('visited','O'), ('Paris','B-LOC')],
]

# 2. Feature functions (reuse from Module 4.4)
def word2features(sent, i):
    token = sent[i][0] if isinstance(sent[i], tuple) else sent[i]
    feats = {
        'bias': 1.0,
        'word.lower()': token.lower(),
        'word.istitle()': token.istitle(),
        'word.isupper()': token.isupper(),
    }
    if i > 0:
        prev = sent[i-1][0] if isinstance(sent[i-1], tuple) else sent[i-1]
        feats.update({'-1:word.lower()': prev.lower()})
    else:
        feats['BOS'] = True
    if i < len(sent)-1:
        nxt = sent[i+1][0] if isinstance(sent[i+1], tuple) else sent[i+1]
        feats.update({'+1:word.lower()': nxt.lower()})
    else:
        feats['EOS'] = True
    return feats

def sent2features(sent): return [word2features(sent, i) for i in range(len(sent))]
def sent2labels(sent):   return [tag for (_, tag) in sent]

X_train = [sent2features(s) for s in train_sents]
y_train = [sent2labels(s)   for s in train_sents]
X_test  = [sent2features(s) for s in test_sents]
y_test  = [sent2labels(s)   for s in test_sents]

# 3. Train CRF
crf = sklearn_crfsuite.CRF(
    algorithm='lbfgs', c1=0.1, c2=0.1,
    max_iterations=50, all_possible_transitions=True
)
crf.fit(X_train, y_train)

# 4. Predict and evaluate
y_pred = crf.predict(X_test)
print("Input:", [w for w,_ in test_sents[0]])
print("Predicted:", y_pred[0])

Output:

Input: ['Alice', 'visited', 'Paris']
Predicted: ['B-PER' 'O' 'B-LOC']

5.3 Transformer-Based NER with Hugging Face

Use a pretrained BERT‐style model for NER:

from transformers import pipeline

# 1. Explicitly specify model/tokenizer & aggregation strategy
ner = pipeline(
    task="ner",
    model="dbmdz/bert-large-cased-finetuned-conll03-english",
    tokenizer="dbmdz/bert-large-cased-finetuned-conll03-english",
    aggregation_strategy="simple",
    framework="pt",    # use PyTorch backend
    device=-1          # CPU
)

# 2. Run the pipeline
text = "Amazon founder Jeff Bezos visited Berlin University."
entities = ner(text)

# 3. Display results
for ent in entities:
    word  = ent['word']
    label = ent['entity_group']
    span  = f"{ent['start']}-{ent['end']}"
    print(f"{word:20} {label:6} {span}")

Output:

Amazon               ORG    0-6
Jeff Bezos           PER    15-25
Berlin University    ORG    34-51

5.4 Evaluation with seqeval

Compute precision/recall/F₁ for sequence labels:

# 1. Install seqeval (only needs to run once per kernel)
%pip install --quiet seqeval

# 2. Import
from seqeval.metrics import classification_report

# --- 
# At this point you should have:
#   y_test:  a list of sequences, e.g. [['B-PER','O','B-LOC'], ['O','B-ORG',...], ...]
#   y_pred:  a list (or numpy array) with the same shape, containing predicted tags.

# If your y_test / y_pred are numpy arrays, convert them to lists of lists of strings:
# For example, if y_test is an array of shape (n_sents, max_len), do:
# y_test_list = [list(seq) for seq in y_test]
# y_pred_list = [list(seq) for seq in y_pred]

# But if they’re already Python lists of lists, just rename:
y_test_list = [list(seq) for seq in y_test]
y_pred_list = [list(seq) for seq in y_pred]

# 3. Print the seqeval classification report
print(classification_report(
    y_test_list,
    y_pred_list,
    zero_division='0'   # fills 0.0 for any undefined precision/recall instead of error
))

Output: