Skip to main content

Open in Colab

A comprehensive demonstration of Fenic’s semantic classification capabilities for detecting editorial bias and analyzing news articles across multiple sources.

​
Overview

This pipeline performs sophisticated news analysis using Fenic’s semantic operations:
  • Language Analysis: Uses semantic.extract() to identify biased, emotional, or sensationalist language patterns
  • Political Bias Classification: Uses semantic.classify() grounded in extracted data for accurate bias detection
  • Topic Classification: Categorizes articles by subject (politics, technology, business, climate, healthcare)
  • AI-Powered Source Profiling: Uses semantic.reduce() to create comprehensive media profiles
Available in both Python script (news_analysis.py) and Jupyter notebook (news_analysis.ipynb) formats for different learning preferences.

​
Key Features

​
Two-Stage Analysis Pipeline

Stage 1 - Information Extraction: Uses semantic.extract() with Pydantic models to identify bias indicators, emotional language, and opinion markers. Stage 2 - Grounded Classification: Uses extracted information as context for semantic.classify() to achieve more accurate political bias detection.

​
Multi-Dimensional Classification

Simultaneously classifies articles across:
  • Topics: politics, technology, business, climate, healthcare
  • Political Bias: far_left, left_leaning, neutral, right_leaning, far_right
  • Journalistic Style: sensationalist vs informational

​
Source Consistency Analysis

Analyzes bias patterns across multiple articles per source to identify editorial consistency.

​
AI-Generated Media Profiles

Uses semantic.reduce() to synthesize extracted information into comprehensive, natural language profiles for each news source.

​
Dataset

The example includes 25 news articles from 8 sources covering diverse topics:
  • Politics: Federal Reserve policy, climate agreements, Supreme Court cases
  • Technology: AI developments, content moderation, privacy concerns
  • Business: Corporate earnings, market analysis, economic trends
  • Healthcare: Medical breakthroughs, drug pricing, treatment access

​
Implementation

​
Session Configuration

import fenic as fc
from pydantic import BaseModel, Field

config = fc.SessionConfig(
    app_name="news_analysis",
    semantic=fc.SemanticConfig(
        language_models={
            "openai": fc.OpenAILanguageModel(
                model_name="gpt-4o-mini",
                rpm=500,
                tpm=200_000
            ),
            # Alternative models:
            # "gemini": fc.GoogleDeveloperLanguageModel(
            #     model_name="gemini-2.0-flash",
            #     rpm=500,
            #     tpm=1_000_000
            # ),
            # "anthropic": fc.AnthropicLanguageModel(
            #     model_name="claude-3-5-haiku-latest",
            #     rpm=500,
            #     input_tpm=80_000,
            #     output_tpm=32_000,
            # )
        }
    )
)

session = fc.Session.get_or_create(config)
The script includes commented configurations for Google Gemini and Anthropic models if you prefer different providers.

​
Stage 1: Information Extraction

# Define Pydantic model for article analysis
class ArticleAnalysis(BaseModel):
    """Comprehensive analysis of news article content and bias."""
    bias_indicators: str = Field(
        description="Key words or phrases that indicate political bias"
    )
    emotional_language: str = Field(
        description="Emotionally charged words or neutral descriptive language"
    )
    opinion_markers: str = Field(
        description="Words or phrases that signal opinion vs. factual reporting"
    )

# Create combined content for analysis
combined_content = fc.text.concat(
    fc.col("headline"),
    fc.lit(" | "),
    fc.col("content")
)

# Extract metadata and classify primary topic
enriched_df = df.with_column("combined_content", combined_content).select(
    fc.col("source"),
    fc.col("headline"),
    fc.col("content"),
    # Topic classification
    fc.semantic.classify(
        fc.col("combined_content"),
        ["politics", "technology", "business", "climate", "healthcare"]
    ).alias("primary_topic"),
    # Extract structured analysis using Pydantic
    fc.semantic.extract(
        fc.col("combined_content"),
        ArticleAnalysis,
        max_output_tokens=512,
    ).alias("analysis_metadata"),
).unnest("analysis_metadata")

​
Stage 2: Grounded Classification

# Combine extracted information for context-aware classification
combined_extracts = fc.text.jinja(
    (
        "Primary Topic: {{primary_topic}}\n"
        "Political Bias Indicators: {{bias_indicators}}\n"
        "Emotional Language Summary: {{emotional_language}}\n"
        "Opinion Markers: {{opinion_markers}}"
    ),
    primary_topic=fc.col("primary_topic"),
    bias_indicators=fc.col("bias_indicators"),
    emotional_language=fc.col("emotional_language"),
    opinion_markers=fc.col("opinion_markers")
)

enriched_df = enriched_df.with_column("combined_extracts", combined_extracts)

# Classify bias using extracted context
results_df = enriched_df.select(
    "*",
    fc.semantic.classify(
        fc.col("combined_extracts"),
        ["far_left", "left_leaning", "neutral", "right_leaning", "far_right"]
    ).alias("content_bias"),
    fc.semantic.classify(
        fc.col("combined_extracts"),
        ["sensationalist", "informational"]
    ).alias("journalistic_style")
).cache()
Grounded classification improves accuracy by first extracting relevant context, then using that information to make more informed classifications.

​
AI-Powered Source Profiling

# Prepare article attributes for profiling
results_df = results_df.with_column("article_attributes", fc.text.jinja(
    (
        "Primary Topics: {{primary_topic}}\n"
        "Detected Political Bias: {{content_bias}}\n"
        "Detected Bias Indicators: {{bias_indicators}}\n"
        "Opinion Indicators: {{opinion_markers}}\n"
        "Emotional Language: {{emotional_language}}\n"
        "Journalistic Style: {{journalistic_style}}"
    ),
    primary_topic=fc.col("primary_topic"),
    content_bias=fc.col("content_bias"),
    bias_indicators=fc.col("bias_indicators"),
    opinion_markers=fc.col("opinion_markers"),
    emotional_language=fc.col("emotional_language"),
    journalistic_style=fc.col("journalistic_style")
))

# Generate semantic summaries for each source
source_language_profiles = results_df.group_by("source").agg(
    fc.semantic.reduce(
        """
        You are given a set of article analyses from {{news_outlet}}.
        Create a concise (3-5 sentence) media profile for {{news_outlet}}.
        Summarize the information provided without explicitly referencing it.
        """,
        column=fc.col("article_attributes"),
        group_context={
            "news_outlet": fc.col("source"),
        },
        max_output_tokens=1024,
    ).alias("source_profile"),
).select(fc.col("source"), fc.col("source_profile")).cache()

print("AI-Generated Media Profiles:")
source_language_profiles.show()

​
Analytics

​
Distribution Analysis

# Source bias distribution
source_bias_distribution = results_df.group_by("source", "content_bias").agg(
    fc.count("*").alias("count")
).order_by(["source", fc.desc("count")])

print("Source Bias Distribution:")
source_bias_distribution.show()

# Topic distribution
print("Topic Distribution:")
results_df.group_by("primary_topic").agg(
    fc.count("*").alias("count")
).order_by(fc.desc("count")).show()

# Bias level distribution
print("Content Bias Distribution:")
results_df.group_by("content_bias").agg(
    fc.count("*").alias("count")
).order_by(fc.desc("count")).show()

# Topic vs Bias cross-analysis
print("Topic vs Bias Analysis:")
results_df.group_by("primary_topic", "content_bias").agg(
    fc.count("*").alias("count")
).order_by(cols=[fc.col("primary_topic"), fc.desc("count")]).show()

​
Language Pattern Analysis

# Show examples of neutral vs biased language
print("Neutral Articles - Language Patterns:")
results_df.filter(
    fc.col("content_bias") == "neutral"
).select(
    "source",
    "headline",
    "bias_indicators",
    "opinion_markers"
).show(5)

print("Biased Articles - Language Patterns:")
results_df.filter(
    fc.col("content_bias") != "neutral"
).select(
    "source",
    "headline",
    "content_bias",
    "bias_indicators",
    "emotional_language",
    "opinion_markers"
).show(6)

​
Expected Results

​
Generated Source Profile Example

The Balanced Tribune presents a diverse range of topics, primarily focusing on business, technology, climate, and healthcare. It exhibits a right-leaning bias in its business and technology coverage, emphasizing themes like Wall Street stability and American free enterprise, while adopting a far-left perspective on climate issues, critiquing fossil fuel companies. The publication often employs sensationalist and informational journalistic styles, utilizing emotional language to evoke strong reactions.

​
Use Cases

Media Organizations

Content quality assessment, bias detection in reporter training, and audience analytics.

News Aggregators

Content categorization, bias warnings for balanced consumption, and source diversity.

Research Applications

Media bias studies, information quality research, and comparative analysis.

Educational Tools

Media literacy training, critical thinking exercises, and journalism education.

​
Running the Example

# Set your API key (OpenAI, Google, or Anthropic)
export OPENAI_API_KEY="your-api-key-here"
# export GOOGLE_API_KEY="your-api-key-here"
# export ANTHROPIC_API_KEY="your-api-key-here"

python news_analysis.py

​
Advanced Features

​
Grounded Classification Pipeline

Shows how to improve classification accuracy by first extracting relevant information with semantic.extract(), then using that context for more informed semantic.classify() operations.

​
Pydantic Integration

Demonstrates structured data extraction using type-safe Pydantic models with automatic field validation.

​
Multi-Model Support

Includes configurations for Google Gemini (default), OpenAI, and Anthropic models.

​
Semantic Reduction for Profiling

Uses semantic.reduce() to synthesize multiple data points into coherent natural language profiles.

​
Key Insights Demonstrated

  • Content-based bias detection without relying on source name predictions
  • Source consistency analysis across multiple articles
  • Language pattern identification for bias indicators
  • Topic-agnostic bias detection (same source biased across different topics)
  • Quality assessment with confidence scoring

Build docs developers (and LLMs) love

Get started for freeTalk to us