Skip to content

CASE_NAME_EXTRACTION_FIXES.md

Latest commit

 

History

History
80 lines (52 loc) · 3.64 KB

CASE_NAME_EXTRACTION_FIXES.md

File metadata and controls

80 lines (52 loc) · 3.64 KB

Case Name Extraction Fixes - Summary

Problem Statement

The case name extraction logic was picking up incorrect or truncated case names from document headers and captions, leading to contamination in citation extraction.

Root Causes Identified

  1. Document headers like "CARTER, Respondent, v. MARY E. JONES, Appellant" were not being properly filtered
  2. Sentence fragments like "This was later cited in Smith v. Johnson" were being extracted instead of just "Smith v. Johnson"
  3. Role words (Respondent, Appellant, etc.) in case names were not being cleaned properly
  4. Context window was too small (40 chars) to capture full case names

Fixes Implemented

1. Enhanced Document Primary Case Name Detection (unified_clustering_master.py)

  • Strategy 2 & 3: Added robust cleaning of role words from both ends of party names
  • Handles patterns like "CARTER, Respondent, v. MARY E. JONES, Appellant" → "CARTER v. MARY E. JONES"
  • Cleans up double commas and extra spaces

2. Improved Header Contamination Filtering (unified_case_extraction_master.py)

  • Lines 1185-1186: Added pattern to detect case caption headers with role words
  • Lines 1196-1216: Enhanced logic to filter case captions while preserving legitimate case discussion
  • Distinguishes between headers ("CARTER, Respondent, v. MARY E. JONES, Appellant") and discussion ("In Smith v. Jones, the court held...")

3. Enhanced Case Name Cleaning

  • Lines 2989-2999: Added role word removal from extracted case names
  • Removes patterns like ", Respondent" and "Appellant, " from case names
  • Cleans up resulting punctuation

4. Improved Sentence-Level Extraction

  • Lines 1735-1754: Added logic to extract case names from longer sentences
  • Detects signal words (see, cited, established, etc.) to identify sentence contamination
  • Uses precise regex pattern to extract just the case name: r"\b[A-Z][a-z]+(?:\s+[A-Z][a-z]+)*\s+v\.\s+[A-Z][a-z]+(?:\s+[A-Z][a-z]+)*\b"
  • Example: "As established in Davis v. Wilson" → "Davis v. Wilson"

5. Increased Context Window

  • Line 1448: Increased default context window from 40 to 100 characters
  • Ensures enough context to capture complete case names

6. Refined Contamination Indicators

  • Lines 1949-1963: Made contamination indicators more specific
  • Removed overly broad terms like "established" that caused false positives
  • Now only rejects clear contamination patterns like "held that", "the court held that", etc.

Test Results

All test cases now pass:

  • ✅ Simple case: "Smith v. Johnson" from "The court ruled in Smith v. Johnson, 123 F.3d 456"
  • ✅ Signal words: "Davis v. Wilson" from "As established in Davis v. Wilson, 456 U.S. 789"
  • ✅ Header contamination: Properly filters "CARTER, Respondent, v. MARY E. JONES, Appellant"
  • ✅ Role word cleaning: Removes "Respondent" and "Appellant" from case names

Impact

  1. More accurate citation extraction - No longer picks up document headers as case names
  2. Cleaner case names - Role words and signal phrases properly removed
  3. Better contamination detection - Distinguishes between headers and legitimate case discussion
  4. Improved clustering - Correct case names lead to better citation clustering

Files Modified

  • src/unified_clustering_master.py: Enhanced primary case name extraction
  • src/unified_case_extraction_master.py: Multiple improvements to extraction and filtering logic

Testing

Created comprehensive test suites:

  • test_case_name_extraction_fixes.py: Full pipeline testing
  • test_simple_case_extraction.py: Focused testing of specific scenarios

All tests pass successfully, confirming the fixes work as expected.