Skip to main content

claim_extraction

Claim extraction from video summaries using LLMs.

This module provides functions for extracting atomic factual claims from summary text using language models. Supports multiple extraction strategies, contextual enrichment from ontology and annotations, and hierarchical claim decomposition.

json​

logging​

re​

Any​

GenerationConfig​

LLMLoader​

ExtractedClaim​

logger​

extract_claims_from_summary​

async def extract_claims_from_summary(
    summary_text: str,
    sentences: list[str] | None,
    strategy: str,
    max_claims: int,
    min_confidence: float,
    llm_loader: LLMLoader,
    ontology_context: dict[str, Any] | None = None,
    annotation_context: list[dict[str, Any]] | None = None
) -> list[ExtractedClaim]

Extract atomic claims from summary text.

Parameters​

summary_text : str Full summary text to extract claims from. sentences : list[str] | None Pre-split sentences (if None, will split automatically). strategy : str Extraction strategy: "sentence-based", "semantic-units", or "hierarchical". max_claims : int Maximum number of claims to extract. min_confidence : float Minimum confidence threshold. llm_loader : LLMLoader Loaded LLM for generation. ontology_context : dict[str, Any] | None Ontology types and glosses for context. annotation_context : list[dict[str, Any]] | None Annotation data for context.

Returns​

list[ExtractedClaim] List of extracted claims with subclaims.

build_extraction_prompt​

def build_extraction_prompt(summary_text: str, sentences: list[str],
                            strategy: str,
                            ontology_context: dict[str, Any] | None,
                            annotation_context: list[dict[str, Any]] | None,
                            max_claims: int) -> str

Build LLM prompt for claim extraction.

Parameters​

summary_text : str Full summary text. sentences : list[str] Split sentences. strategy : str Extraction strategy. ontology_context : dict[str, Any] | None Ontology types and glosses. annotation_context : list[dict[str, Any]] | None Annotation data. max_claims : int Maximum claims to extract.

Returns​

str Formatted prompt for LLM.

parse_claims_response​

def parse_claims_response(response: str, summary_text: str,
                          sentences: list[str],
                          min_confidence: float) -> list[ExtractedClaim]

Parse LLM response into structured claims.

Parameters​

response : str Raw LLM response text. summary_text : str Original summary text. sentences : list[str] Split sentences. min_confidence : float Minimum confidence threshold.

Returns​

list[ExtractedClaim] Parsed and validated claims.

parse_single_claim​

def parse_single_claim(claim_data: dict[str, Any],
                       min_confidence: float) -> ExtractedClaim | None

Parse single claim recursively.

Parameters​

claim_data : dict[str, Any] Claim data dictionary. min_confidence : float Minimum confidence threshold.

Returns​

ExtractedClaim | None Parsed claim or None if below threshold.

split_into_sentences​

def split_into_sentences(text: str) -> list[str]

Split text into sentences using simple heuristics.

Parameters​

text : str Text to split.

Returns​

list[str] List of sentences.