Skip to content

Sentiment Analysis

Sentimatrix provides multiple sentiment analysis modes for different use cases.

Analysis Modes

Mode Output Speed Use Case
Quick Label only Fastest High-volume filtering
Structured Label + scores Fast Analytics
Fine-Grained 5 classes Medium Nuanced analysis
Aspect-Based Per-aspect Medium Product feedback
Comparative Comparison Slower Competition analysis
Temporal Over time Medium Trend tracking

Quick Sentiment

Fast classification into positive/negative/neutral:

async with Sentimatrix() as sm:
    result = await sm.analyze("This product is amazing!")

    print(result.sentiment)   # "positive"
    print(result.confidence)  # 0.967

Batch Processing

texts = [
    "Great quality!",
    "Terrible experience",
    "It's okay",
]

results = await sm.analyze_batch(texts)

for text, result in zip(texts, results):
    print(f"{result.sentiment:>10} ({result.confidence:.0%}): {text}")

Output:

  positive (97%): Great quality!
  negative (94%): Terrible experience
   neutral (82%): It's okay

Structured Analysis

Get detailed scores for each class:

result = await sm.analyze("Good product but shipping was slow", structured=True)

print(f"Label: {result.sentiment}")
print(f"Confidence: {result.confidence:.2%}")
print(f"Scores:")
print(f"  Positive: {result.scores['positive']:.3f}")
print(f"  Negative: {result.scores['negative']:.3f}")
print(f"  Neutral: {result.scores['neutral']:.3f}")

Output:

Label: positive
Confidence: 54.23%
Scores:
  Positive: 0.542
  Negative: 0.312
  Neutral: 0.146

Fine-Grained Sentiment

5-class classification for nuanced understanding:

result = await sm.analyze(
    "I absolutely love this product!",
    mode="fine_grained"
)

print(result.sentiment)  # "very_positive"

Classes: - very_negative - negative - neutral - positive - very_positive

Aspect-Based Sentiment

Analyze sentiment for specific product aspects:

result = await sm.analyze_aspects(
    "The camera is excellent but battery life is disappointing",
    aspects=["camera", "battery", "design", "price"]
)

for aspect, sentiment in result.items():
    print(f"{aspect}: {sentiment}")

Output:

camera: positive
battery: negative
design: neutral
price: neutral

With Confidence Scores

result = await sm.analyze_aspects(
    "Great camera, terrible battery",
    aspects=["camera", "battery"],
    include_scores=True
)

# {
#   'camera': {'sentiment': 'positive', 'confidence': 0.92},
#   'battery': {'sentiment': 'negative', 'confidence': 0.88}
# }

Comparative Sentiment

Compare sentiment across products or time periods:

products = {
    "Product A": [
        "Great quality product",
        "Love this item",
        "Some issues with durability"
    ],
    "Product B": [
        "Okay product",
        "Not worth the price",
        "Returned it"
    ]
}

comparison = await sm.compare_sentiment(products)

print(comparison)
# {
#   'Product A': {'positive': 0.67, 'negative': 0.33, 'avg_score': 0.45},
#   'Product B': {'positive': 0.00, 'negative': 0.67, 'avg_score': -0.52},
#   'winner': 'Product A'
# }

Temporal Sentiment

Track sentiment changes over time:

reviews_over_time = [
    {"text": "Great product!", "date": "2024-01-01"},
    {"text": "Quality seems worse", "date": "2024-03-01"},
    {"text": "Terrible now", "date": "2024-06-01"},
]

timeline = await sm.analyze_temporal(reviews_over_time)

for period in timeline:
    print(f"{period.date}: {period.sentiment} ({period.avg_score:.2f})")

Domain-Specific Analysis

Use domain-specific models for better accuracy:

# Financial sentiment
result = await sm.analyze(
    "Stock prices fell sharply today",
    domain="financial"
)
# Correctly identifies as negative (financial context)

# Social media
result = await sm.analyze(
    "This is so bad it's good! ",
    domain="social_media"
)
# Handles sarcasm and emojis better

Available domains: - general (default) - financial - social_media - product_reviews - healthcare

Model Configuration

Default Model

# Uses cardiffnlp/twitter-roberta-base-sentiment-latest
async with Sentimatrix() as sm:
    result = await sm.analyze("Hello world")

Custom Model

from sentimatrix.config import SentimatrixConfig, ModelConfig

config = SentimatrixConfig(
    model=ModelConfig(
        sentiment_model="nlptown/bert-base-multilingual-uncased-sentiment"
    )
)

async with Sentimatrix(config) as sm:
    result = await sm.analyze("Bonjour le monde")

Available Models (7 Implemented)

Model Languages Classes Best For
cardiffnlp/twitter-roberta-base-sentiment-latest English 3 Default, general use
cardiffnlp/twitter-roberta-base-sentiment English 3 Social media
nlptown/bert-base-multilingual-uncased-sentiment 100+ 5 Multi-language
distilbert-base-uncased-finetuned-sst-2-english English 2 Fast binary
finiteautomata/bertweet-base-sentiment-analysis English 3 Twitter/social
siebert/sentiment-roberta-large-english English 2 High accuracy
lxyuan/distilbert-base-multilingual-cased-sentiments-student 100+ 3 Fast multi-language

Domain-Specific Models (4 Implemented)

Model Domain Use Case
ProsusAI/finbert Financial Stock news, earnings
yiyanghkust/finbert-tone Financial Sentiment tone
nlpaueb/legal-bert-base-uncased Legal Legal documents
allenai/scibert_scivocab_uncased Scientific Research papers

Performance Tips

  1. Use Batch Processing 

    # Slower
for text in texts:
    result = await sm.analyze(text)

# Faster (3-10x)
results = await sm.analyze_batch(texts)

  2. Enable GPU Acceleration 

    config = SentimatrixConfig(
    model=ModelConfig(device="cuda")  # or "mps" for Apple Silicon
)

  3. Use Appropriate Mode

    • Quick: High volume, don't need scores
    • Structured: Need confidence scores
    • Fine-grained: Need nuanced classes

  4. Cache Results 

    config = SentimatrixConfig(
    cache=CacheConfig(enabled=True)
)