NLP Challenges — Ambiguity, Context & Linguistic Complexity

import spacy
nlp = spacy.load("en_core_web_sm")

# ── Coreference resolution challenge ──
sentences = [
    "Alice told Bob she was tired. He was also exhausted.",
    "The trophy did not fit in the suitcase because it was too big.",  # Winograd schema
    "The council denied the demonstrators a permit because they feared violence.",  # "they" = council or demonstrators?
]

for sent in sentences:
    doc = nlp(sent)
    print(f"Text: {sent}")
    print(f"Named entities: {[(ent.text, ent.label_) for ent in doc.ents]}")
    # Basic spaCy cannot resolve coreference — requires specialised models
    print()

# ── Word sense disambiguation challenge ──
ambiguous = ["I went to the bank to deposit money.",
             "The bank of the river was muddy.",
             "The children played on the bank of sand."]

print("'bank' in different contexts:")
for sent in ambiguous:
    doc = nlp(sent)
    bank = [t for t in doc if t.text.lower() == 'bank']
    if bank:
        print(f"  Context: '{sent[:50]}...'")
        print(f"  Dependency: {bank[0].dep_}, Head: {bank[0].head.text}")
        # spaCy gives grammatical role but not semantic disambiguation

# ── Negation handling ──
sentiments = [
    "This is not bad at all",     # double negation = positive
    "I would not say it is good", # = negative
    "It is not without merit",    # = has merit (double neg)
]
from transformers import pipeline
clf = pipeline("sentiment-analysis", model="distilbert-base-uncased-finetuned-sst-2-english")
for text in sentiments:
    result = clf(text)[0]
    print(f"{result['label']:8} ({result['score']:.2f}): {text}")

from transformers import pipeline

# Word embedding bias test (Word Association Test)
unmasker = pipeline('fill-mask', model='bert-base-uncased')

# Occupational bias
templates = [
    "The doctor told the nurse that [MASK] needed to update the charts.",
    "The engineer finished [MASK] project on time.",
    "The nurse greeted [MASK] patient with a smile.",
]

print("Occupational gender bias in BERT:")
for template in templates:
    results = unmasker(template)[:3]
    pronoun_results = [(r['token_str'], round(r['score'],3))
                       for r in results if r['token_str'].lower() in
                       ['he', 'she', 'his', 'her', 'they', 'their']]
    print(f"  '{template[:60]}...'")
    print(f"  Pronouns: {pronoun_results}")
    # "The engineer finished [MASK] project" → "his" (0.81), "the" (0.05), "her" (0.04)

# Sentiment bias test
test_sentences = [
    "The {group} doctor examined the patient carefully.",
    "The {group} lawyer argued the case effectively.",
    "The {group} person was walking down the street.",
]
groups = ['Black', 'White', 'Asian', 'male', 'female']

print("
Sentiment consistency test:")
from transformers import pipeline as clf_pipeline
clf = clf_pipeline('sentiment-analysis')
for template in test_sentences[:1]:
    for group in groups:
        text   = template.format(group=group)
        result = clf(text)[0]
        print(f"  {result['label']:8} ({result['score']:.3f}): {text}")