AI Web FeedsAI Web FeedsOpen web AI reader
    • Features
    Documentation

    Sentiment Analysis

    Transformer-based sentiment classification and trend tracking

    Source: apps/web/content/docs/features/sentiment-analysis.mdx

    Sentiment Analysis

    Sentiment Analysis classifies article sentiment using transformer models (DistilBERT) and tracks sentiment trends over time by topic.

    Overview

    The sentiment analyzer:

    1. Classifies article sentiment: positive, neutral, or negative
    2. Computes sentiment scores (-1.0 to +1.0)
    3. Aggregates daily sentiment by topic
    4. Detects sentiment shifts using moving averages

    Architecture

    Sentiment Classification

    Model

    Uses Hugging Face's distilbert-base-uncased-finetuned-sst-2-english:

    • Model Size: 67MB
    • Accuracy: ~92% on SST-2 benchmark
    • Inference Time: ~50ms per article (CPU)
    • Context Window: 512 tokens (truncates longer articles)

    Sentiment Score Mapping

    # Model output → Sentiment score
"POSITIVE" (confidence 0.85) → +0.85
"NEGATIVE" (confidence 0.92) → -0.92
"NEUTRAL" → 0.0

    Classification Thresholds

    if sentiment_score > 0.3:
    classification = "positive"
elif sentiment_score < -0.3:
    classification = "negative"
else:
    classification = "neutral"

    Usage

    CLI Commands

    Analyze Sentiment

    ai-web-feeds nlp sentiment

    Options:

    • --batch-size: Number of articles (default: 100)
    • --force: Reprocess all articles
    # Process 50 articles
ai-web-feeds nlp sentiment --batch-size 50
    # 30-day sentiment trend for "AI Safety"
ai-web-feeds nlp sentiment-trend "AI Safety" --days 30

    Output:

    AI Safety - Sentiment Trend (30 days)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Date       Avg Sentiment  Articles  Positive  Neutral  Negative
2023-10-01    +0.45         24        18        4         2
2023-10-02    +0.32         19        12        5         2
2023-10-03    -0.15         28         8       12         8  ⚠️  Shift

    Detect Sentiment Shifts

    # Show topics with sentiment shifts (>0.3 change in 7-day MA)
ai-web-feeds nlp sentiment-shifts

    Output:

    Recent Sentiment Shifts
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Topic          Previous  Current  Change  Status
AI Safety        +0.25    -0.18    -0.43   🔴 Major shift
AI Regulation    -0.10    +0.35    +0.45   🟢 Improving

    Compare Topics

    ai-web-feeds nlp sentiment-compare "AI Safety" "AI Capabilities"

    Shows side-by-side sentiment trends for two topics.

    Python API

    from ai_web_feeds.nlp import SentimentAnalyzer
from ai_web_feeds.config import Settings

analyzer = SentimentAnalyzer(Settings())

article = {
    "id": 1,
    "title": "RLHF Concerns",
    "content": "Critics have raised serious concerns about RLHF..."
}

sentiment = analyzer.analyze_sentiment(article)
# Returns: {
#     "sentiment_score": -0.65,
#     "classification": "negative",
#     "confidence": 0.89,
#     "model_name": "distilbert-base-uncased-finetuned-sst-2-english"
# }

    Batch Processing

    Sentiment analysis runs hourly:

    from ai_web_feeds.nlp.scheduler import NLPScheduler

nlp_scheduler = NLPScheduler(scheduler)
nlp_scheduler.register_jobs()
# Registers:
# - Sentiment analysis (every hour)
# - Sentiment aggregation (15 min after analysis)

    Database Schema

    article_sentiment Table

    CREATE TABLE article_sentiment (
    article_id INTEGER PRIMARY KEY,
    sentiment_score REAL NOT NULL CHECK(sentiment_score BETWEEN -1.0 AND 1.0),
    classification TEXT NOT NULL CHECK(classification IN ('positive', 'neutral', 'negative')),
    model_name TEXT NOT NULL,
    confidence REAL NOT NULL CHECK(confidence BETWEEN 0 AND 1),
    computed_at DATETIME DEFAULT CURRENT_TIMESTAMP,
    FOREIGN KEY (article_id) REFERENCES articles(id)
);

    topic_sentiment_daily Table

    Aggregated daily sentiment by topic:

    CREATE TABLE topic_sentiment_daily (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    topic TEXT NOT NULL,
    date DATE NOT NULL,
    avg_sentiment REAL NOT NULL,
    article_count INTEGER NOT NULL,
    positive_count INTEGER DEFAULT 0,
    neutral_count INTEGER DEFAULT 0,
    negative_count INTEGER DEFAULT 0,
    UNIQUE(topic, date)
);

    Sentiment Aggregation

    Daily Aggregation

    Runs 15 minutes after sentiment analysis:

    # Group sentiment scores by (topic, date)
aggregates = {}
for article in recent_articles:
    for topic in article.topics:
        key = (topic, article.date)
        aggregates[key]["scores"].append(article.sentiment_score)
        aggregates[key][article.classification] += 1

# Compute average
for (topic, date), data in aggregates.items:
    avg_sentiment = sum(data["scores"]) / len(data["scores"])
    storage.upsert_topic_sentiment_daily(
        topic=topic,
        date=date,
        avg_sentiment=avg_sentiment,
        article_count=len(data["scores"]),
        positive_count=data["positive"],
        neutral_count=data["neutral"],
        negative_count=data["negative"]
    )

    Shift Detection

    7-day moving average:

    def detect_shift(topic: str, threshold: float = 0.3) -> bool:
    """Detect sentiment shift using 7-day moving average"""
    trend = storage.get_topic_sentiment_trend(topic, days=14)

    # Compute 7-day MA for last 2 weeks
    ma_recent = mean([day.avg_sentiment for day in trend[:7]])
    ma_previous = mean([day.avg_sentiment for day in trend[7:14]])

    shift = abs(ma_recent - ma_previous)
    return shift > threshold

    Configuration

    class Phase5Settings(BaseSettings):
    sentiment_batch_size: int = 100
    sentiment_cron: str = "0 * * * *"  # Every hour
    sentiment_model: str = "distilbert-base-uncased-finetuned-sst-2-english"
    sentiment_shift_threshold: float = 0.3

    Environment Variables:

    PHASE5_SENTIMENT_BATCH_SIZE=100
PHASE5_SENTIMENT_SHIFT_THRESHOLD=0.3
PHASE5_SENTIMENT_MODEL=distilbert-base-uncased-finetuned-sst-2-english

    Performance

    • Throughput: ~100 articles/hour (CPU)
    • Memory: ~500MB (model loaded)
    • Storage: ~50 bytes per sentiment record

    Use Cases

    Monitor Topic Sentiment

    Track sentiment for specific topics:

    # Daily check for "AI Safety" sentiment
ai-web-feeds nlp sentiment-trend "AI Safety" --days 7

    Detect Controversies

    Identify topics with negative sentiment spikes:

    # Topics with sentiment < -0.5 in last 7 days
ai-web-feeds nlp sentiment-shifts --threshold -0.5

    Compare Competing Approaches

    # Compare sentiment for competing techniques
ai-web-feeds nlp sentiment-compare "RLHF" "Constitutional AI"

    Model Details

    DistilBERT Architecture

    • Base Model: BERT distilled to 66M parameters (40% smaller)
    • Training: Fine-tuned on SST-2 (Stanford Sentiment Treebank)
    • Input: Max 512 tokens (articles truncated to ~2000 chars)
    • Output: Binary classification (positive/negative) with confidence

    Limitations

    1. Context Window: Only first 512 tokens considered
    2. Binary Classification: Model trained for binary sentiment (positive/negative), neutral inferred
    3. Domain Shift: SST-2 is movie reviews; AI articles may differ
    4. No Fine-tuning: Pre-trained model used as-is (no domain adaptation)

    Troubleshooting

    Low Confidence Scores

    Symptom: All sentiment predictions have low confidence (<0.6).

    Cause: Articles too long, model only sees truncated beginning.

    Solution: Increase truncation window or use extractive summarization before analysis.

    Model Download Fails

    Symptom: OSError: Can't find model

    Solution:

    # Models auto-download to ~/.cache/huggingface/hub
# Ensure internet connection and disk space (~67MB)

# Manual download:
uv run python -c "from transformers import pipeline; pipeline('sentiment-analysis', model='distilbert-base-uncased-finetuned-sst-2-english')"

    Sentiment Shifts Not Detected

    Symptom: No shifts reported despite obvious sentiment changes.

    Cause: Threshold too high.

    Solution:

    # Lower threshold to 0.2
export PHASE5_SENTIMENT_SHIFT_THRESHOLD=0.2

    Future Enhancements

    1. Domain-Specific Fine-tuning: Train on AI article sentiment labels
    2. Aspect-Based Sentiment: Sentiment for specific entities/topics within articles
    3. Multilingual Support: Add models for non-English content
    4. Real-Time Alerts: Webhook notifications for sentiment shifts

    See Also

    On this page

    Sentiment AnalysisOverviewArchitectureSentiment ClassificationModelSentiment Score MappingClassification ThresholdsUsageCLI CommandsAnalyze SentimentView Sentiment TrendsDetect Sentiment ShiftsCompare TopicsPython APIBatch ProcessingDatabase Schemaarticle_sentiment Tabletopic_sentiment_daily TableSentiment AggregationDaily AggregationShift DetectionConfigurationPerformanceUse CasesMonitor Topic SentimentDetect ControversiesCompare Competing ApproachesModel DetailsDistilBERT ArchitectureLimitationsTroubleshootingLow Confidence ScoresModel Download FailsSentiment Shifts Not DetectedFuture EnhancementsSee Also
    Sentiment Analysis | AI Web Feeds