spaCy

spaCy is "Industrial Strength" NLP. Unlike NLTK (academic), spaCy focuses on providing the best single algorithm for a task. v3.8 supports Python 3.13.

When to Use

• Production-Ready Natural Language Processing: Fast, industrial-grade tokenization, lemmatization, and dependency parsing.
• Named Entity Recognition (NER): Extracting organizations, people, locations, dates, and domain-specific entities.
• Rule-Based Matching & Pattern Extraction: Combining linguistic rules with regex and token attributes using Matcher and EntityRuler.
• Large-Scale Text Processing Pipelines: Batch processing millions of documents with nlp.pipe() and GPU acceleration.

Quick Start

import spacy

# Load transformer or core English pipeline
nlp = spacy.load("en_core_web_sm")
doc = nlp("Apple is acquiring a London startup for $1 billion.")

# Extract named entities
for ent in doc.ents:
    print(f"{ent.text:<15} {ent.label_:<10}")
# Output:
# Apple           ORG
# London          GPE
# $1 billion      MONEY

Core Concepts

High-Throughput Batch Processing with nlp.pipe

Processing document streams efficiently using multiprocessing and batching:

import spacy

# Load transformer or optimized small model
nlp = spacy.load("en_core_web_sm")

documents = [
    "Apple Inc. acquired a machine learning startup based in Zurich for $200 million.",
    "Dr. Maria Santos published research on genomic sequencing at Oxford University.",
    "Amazon Web Services launched new European cloud regions in Frankfurt."
]

# Process documents in batches, disabling unused pipeline stages
for doc in nlp.pipe(documents, batch_size=64, disable=["parser"]):
    print(f"--- Document ({len(doc)} tokens) ---")
    for ent in doc.ents:
        print(f"Entity: {ent.text:25s} Label: {ent.label_}")

Rule-Based EntityRuler & Pattern Matching

Combining custom dictionary rules with statistical NER:

nlp = spacy.load("en_core_web_sm")

# Add EntityRuler before the statistical ner component
ruler = nlp.add_pipe("entity_ruler", before="ner")
patterns = [
    {"label": "PRODUCT_SKU", "pattern": [{"TEXT": {"REGEX": r"^SKU-\d{4}-[A-Z]{2}$"}}]},
    {"label": "SECURITY_FLAG", "pattern": [{"LOWER": "unauthorized"}, {"LOWER": "access"}]}
]
ruler.add_patterns(patterns)

doc = nlp("Alert: unauthorized access detected for asset SKU-8492-EU on production server.")
for ent in doc.ents:
    print(f"Match: {ent.text} -> {ent.label_}")

Syntactic Dependency Parsing & Semantic Traversal

Extracting subject-verb-object relationships:

doc = nlp("Autonomous vehicles navigate complex city intersections safely.")

for token in doc:
    print(f"{token.text:12s} Dep: {token.dep_:10s} Head: {token.head.text:12s} POS: {token.pos_}")

# Extract grammatical subject and direct object
for token in doc:
    if "subj" in token.dep_:
        print(f"Subject: {token.text}")
    elif "obj" in token.dep_:
        print(f"Object: {token.text}")

Common Patterns

Efficient Batch Processing with nlp.pipe

Problem: Calling nlp(text) one by one on millions of records is slow and underutilizes CPU cores.

Solution: Process text in batches with disabled unused pipeline components:

texts = ["Invoice #1234 from Vendor A...", "Contract dated 2025..."]

# Disable tagger and parser if only NER is required
for doc in nlp.pipe(texts, batch_size=50, disable=["tagger", "parser"]):
    entities = [(ent.text, ent.label_) for ent in doc.ents]
    print(entities)

Best Practices

Do:

• Always use nlp.pipe(texts, batch_size=...) instead of iterating with [nlp(t) for t in texts].
• Disable pipeline components not required for the task (nlp.pipe(texts, disable=["parser", "ner"])) to gain up to 5x speedups.
• Call spacy.require_gpu() before loading models when running on CUDA-enabled servers.
• Train custom spaCy components using the declarative config.cfg system and spacy train.

Don't:

• Load the heavy transformer model (en_core_web_trf) if simple tokenization/POS with en_core_web_sm suffices.
• Modify doc.ents directly without handling token boundary overlaps; use spacy.util.filter_spans.
• Reload spacy.load() inside request handlers; load models once as global singletons.

Troubleshooting

References

• spaCy Documentation