NLP_Parsing_Tutorial

ipynb6 kBPython · 13 cellsExtra material of chapter 3

Tutorial: Syntactic Parsing with spaCy

NLP Course - Lab 4: Constituency & Dependency Analysis

Introduction

This tutorial is designed to explore the structural analysis of natural language. We will cover:

  1. Dependency Parsing: Focusing on the relationships between individual words.
  2. Constituency Parsing: Focusing on the hierarchical phrase structure.

Prerequisites

You will need spacy and benepar (Berkeley Neural Parser).

1. Setup and Environment

We will use a transformer-optimized model for spaCy and integrate benepar for constituency parsing, as spaCy does not provide this out-of-the-box.

In [ ]
!pip install spacy benepar
import spacy
from spacy import displacy
import benepar

# Download models
spacy.cli.download('en_core_web_md')
benepar.download('benepar_en3')

# Load the spaCy model and add the benepar component to the pipeline
nlp = spacy.load('en_core_web_md')
if 'benepar' not in nlp.pipe_names:
    nlp.add_pipe('benepar', config={'model': 'benepar_en3'})

print('\nPipeline components:', nlp.pipe_names)

2. Dependency Parsing

Dependency parsing identifies the head words and their dependents. In Computer Science terms, this results in a Directed Acyclic Graph (DAG).

2.1 Visualization

We use displacy to view the grammatical dependencies.

In [ ]
doc = nlp('The experienced computer science students are analyzing complex linguistic structures.')

# Visualize the dependency tree
displacy.render(doc, style='dep', jupyter=True, options={'distance': 120})
2.2 Navigating the Dependency Graph

Every Token in a spaCy Doc has attributes that allow you to traverse the tree: .head, .children, .lefts, and .rights.

In [ ]
print(f"{'Text':<12} | {'Dep':<10} | {'Head':<12} | {'Children'}")
print('-' * 60)

for token in doc:
    children = [child.text for child in token.children]
    print(f"{token.text:<12} | {token.dep_:<10} | {token.head.text:<12} | {children}")

3. Constituency (Phrase Structure) Parsing

Constituency parsing breaks sentences into sub-phrases (constituents). This follows Context-Free Grammar (CFG) rules.

3.1 Inspecting the Tree Structure

With benepar, we can access labels like S (Sentence), NP (Noun Phrase), and VP (Verb Phrase).

In [ ]
sentence = 'The student solved the problem with a clever algorithm.'
doc = nlp(sentence)
sent = list(doc.sents)[0]

def print_tree(node, depth=0):
    # Extract labels from the benepar extension
    label = node._.labels[0] if node._.labels else 'TOKEN'
    print('  ' * depth + f'[{label}] {node.text}')
    for child in node._.children:
        print_tree(child, depth + 1)

print_tree(sent)
3.2 Handling Ambiguity

Syntactic ambiguity is a major challenge in NLP. Observe how the parser handles Prepositional Phrase (PP) attachment.

In [ ]
ambig_1 = nlp('He ate the salad with a fork.')
ambig_2 = nlp('He ate the salad with croutons.')

print('--- Sentence 1 Tree ---')
print_tree(list(ambig_1.sents)[0])
print('\n--- Sentence 2 Tree ---')
print_tree(list(ambig_2.sents)[0])

4. Comparison Summary

Feature Dependency Constituency
Focus Word-to-word relations Hierarchical grouping
Application Information Extraction Grammar Checking
Logic Head/Dependent Part/Whole
In [ ]