corp-extractor 0.2.7__py3-none-any.whl

statement_extractor/scoring.py

@@ -0,0 +1,419 @@
+"""
+Scoring module for statement extraction quality assessment.
+
+Provides:
+- TripleScorer: Score individual triples for groundedness
+- BeamScorer: Score and select/merge beams based on quality metrics
+"""
+
+from typing import Optional
+
+from .models import ScoringConfig, Statement
+
+
+class TripleScorer:
+    """
+    Score individual triples for groundedness in source text.
+
+    Groundedness is measured by checking:
+    - Subject text appears in source
+    - Object text appears in source
+    - Subject and object are in proximity (same/nearby sentences)
+    - Evidence span exists and is valid
+    """
+
+    def __init__(self, config: Optional[ScoringConfig] = None):
+        self.config = config or ScoringConfig()
+
+    def score_triple(self, statement: Statement, source_text: str) -> float:
+        """
+        Score a triple's groundedness (0-1).
+
+        Higher scores indicate better grounding in source text.
+        """
+        if not source_text:
+            return 0.5  # Neutral score if no source text
+
+        score = 0.0
+        weights_sum = 0.0
+
+        # Check subject appears in source (weight: 0.3)
+        subject_found = self._text_appears_in(statement.subject.text, source_text)
+        score += 0.3 * (1.0 if subject_found else 0.0)
+        weights_sum += 0.3
+
+        # Check object appears in source (weight: 0.3)
+        object_found = self._text_appears_in(statement.object.text, source_text)
+        score += 0.3 * (1.0 if object_found else 0.0)
+        weights_sum += 0.3
+
+        # Check predicate has lexical trigger (weight: 0.2)
+        predicate_grounded = self._predicate_has_trigger(statement.predicate, source_text)
+        score += 0.2 * (1.0 if predicate_grounded else 0.0)
+        weights_sum += 0.2
+
+        # Check proximity - subject and object in same/nearby region (weight: 0.2)
+        if subject_found and object_found:
+            proximity_score = self._compute_proximity(
+                statement.subject.text,
+                statement.object.text,
+                source_text
+            )
+            score += 0.2 * proximity_score
+        weights_sum += 0.2
+
+        return score / weights_sum if weights_sum > 0 else 0.0
+
+    def find_evidence_span(
+        self,
+        statement: Statement,
+        source_text: str
+    ) -> Optional[tuple[int, int]]:
+        """
+        Find character offsets where the triple is grounded in source text.
+
+        Returns (start, end) tuple or None if not found.
+        """
+        if not source_text:
+            return None
+
+        # If statement has source_text field, try to find it
+        if statement.source_text:
+            pos = source_text.lower().find(statement.source_text.lower())
+            if pos >= 0:
+                return (pos, pos + len(statement.source_text))
+
+        # Otherwise, find the region containing both subject and object
+        subject_lower = statement.subject.text.lower()
+        object_lower = statement.object.text.lower()
+        source_lower = source_text.lower()
+
+        subj_pos = source_lower.find(subject_lower)
+        obj_pos = source_lower.find(object_lower)
+
+        if subj_pos >= 0 and obj_pos >= 0:
+            start = min(subj_pos, obj_pos)
+            end = max(
+                subj_pos + len(subject_lower),
+                obj_pos + len(object_lower)
+            )
+            # Extend to sentence boundaries
+            start, end = self._extend_to_sentence(source_text, start, end)
+            return (start, end)
+
+        return None
+
+    def _text_appears_in(self, text: str, source: str) -> bool:
+        """Check if text appears in source (case-insensitive)."""
+        return text.lower() in source.lower()
+
+    def _predicate_has_trigger(self, predicate: str, source: str) -> bool:
+        """Check if predicate has a lexical trigger in source."""
+        # Extract main verb/word from predicate
+        words = predicate.lower().split()
+        source_lower = source.lower()
+
+        # Check if any predicate word appears in source
+        for word in words:
+            if len(word) > 2 and word in source_lower:
+                return True
+        return False
+
+    def _compute_proximity(
+        self,
+        subject_text: str,
+        object_text: str,
+        source: str
+    ) -> float:
+        """
+        Compute proximity score (0-1) based on distance between subject and object.
+
+        Returns 1.0 if same sentence, decreasing with distance.
+        """
+        source_lower = source.lower()
+        subj_pos = source_lower.find(subject_text.lower())
+        obj_pos = source_lower.find(object_text.lower())
+
+        if subj_pos < 0 or obj_pos < 0:
+            return 0.0
+
+        # Check if in same sentence
+        start = min(subj_pos, obj_pos)
+        end = max(subj_pos, obj_pos)
+        region = source[start:end]
+
+        # If no sentence boundary between them, high proximity
+        if '.' not in region and '!' not in region and '?' not in region:
+            return 1.0
+
+        # Otherwise, score decreases with distance
+        # Assume ~100 chars per sentence on average
+        sentence_distance = region.count('.') + region.count('!') + region.count('?')
+        return max(0.0, 1.0 - (sentence_distance * 0.2))
+
+    def _extend_to_sentence(
+        self,
+        source: str,
+        start: int,
+        end: int
+    ) -> tuple[int, int]:
+        """Extend span to sentence boundaries."""
+        # Find sentence start
+        sentence_start = start
+        while sentence_start > 0:
+            char = source[sentence_start - 1]
+            if char in '.!?\n':
+                break
+            sentence_start -= 1
+
+        # Find sentence end
+        sentence_end = end
+        while sentence_end < len(source):
+            char = source[sentence_end]
+            if char in '.!?\n':
+                sentence_end += 1
+                break
+            sentence_end += 1
+
+        return (sentence_start, sentence_end)
+
+
+class BeamScorer:
+    """
+    Score and select/merge beams based on quality metrics.
+
+    Implements the scoring function:
+    Score = Σ quality(t) + β×Coverage - γ×Redundancy - δ×Length
+    """
+
+    def __init__(
+        self,
+        config: Optional[ScoringConfig] = None,
+        triple_scorer: Optional[TripleScorer] = None
+    ):
+        self.config = config or ScoringConfig()
+        self.triple_scorer = triple_scorer or TripleScorer(config)
+
+    def score_beam(
+        self,
+        statements: list[Statement],
+        source_text: str
+    ) -> float:
+        """
+        Compute beam score using the quality formula.
+
+        Score = Σ quality(t) + β×Coverage - γ×Redundancy - δ×Length
+        """
+        if not statements:
+            return 0.0
+
+        # Sum of quality scores
+        quality_sum = sum(
+            (stmt.confidence_score or self.triple_scorer.score_triple(stmt, source_text))
+            for stmt in statements
+        )
+        quality_term = self.config.quality_weight * quality_sum
+
+        # Coverage bonus
+        coverage = self.compute_coverage(statements, source_text)
+        coverage_term = self.config.coverage_weight * coverage
+
+        # Redundancy penalty
+        redundancy = self.compute_redundancy(statements)
+        redundancy_term = self.config.redundancy_penalty * redundancy
+
+        # Length penalty (normalized by statement count)
+        length = len(statements)
+        length_term = self.config.length_penalty * (length / 10.0)  # Normalize
+
+        return quality_term + coverage_term - redundancy_term - length_term
+
+    def compute_coverage(
+        self,
+        statements: list[Statement],
+        source_text: str
+    ) -> float:
+        """
+        Compute coverage: % of source text tokens explained by evidence spans.
+        """
+        if not source_text or not statements:
+            return 0.0
+
+        # Track which character positions are covered
+        covered = set()
+
+        for stmt in statements:
+            span = stmt.evidence_span
+            if span is None:
+                span = self.triple_scorer.find_evidence_span(stmt, source_text)
+
+            if span:
+                for i in range(span[0], min(span[1], len(source_text))):
+                    covered.add(i)
+
+        # Calculate coverage as percentage of non-whitespace characters
+        content_chars = sum(1 for c in source_text if not c.isspace())
+        covered_content = sum(1 for i in covered if not source_text[i].isspace())
+
+        return covered_content / content_chars if content_chars > 0 else 0.0
+
+    def compute_redundancy(self, statements: list[Statement]) -> float:
+        """
+        Compute redundancy penalty for near-duplicate triples.
+
+        Only counts exact duplicates (same subject, predicate, and object).
+        Note: Same subject+predicate with different objects is NOT redundant,
+        as it represents distinct relationships (e.g., "Apple announced iPhone and iPad").
+        """
+        if len(statements) < 2:
+            return 0.0
+
+        redundant_pairs = 0
+        total_pairs = 0
+
+        for i, stmt1 in enumerate(statements):
+            for stmt2 in statements[i + 1:]:
+                total_pairs += 1
+
+                # Only count exact duplicates (same subject, predicate, AND object)
+                if (stmt1.subject.text.lower() == stmt2.subject.text.lower() and
+                    stmt1.predicate.lower() == stmt2.predicate.lower() and
+                    stmt1.object.text.lower() == stmt2.object.text.lower()):
+                    redundant_pairs += 1
+
+        return redundant_pairs / total_pairs if total_pairs > 0 else 0.0
+
+    def score_and_rank_statements(
+        self,
+        statements: list[Statement],
+        source_text: str
+    ) -> list[Statement]:
+        """
+        Score each statement and return sorted by confidence (descending).
+        """
+        for stmt in statements:
+            if stmt.confidence_score is None:
+                stmt.confidence_score = self.triple_scorer.score_triple(stmt, source_text)
+            if stmt.evidence_span is None:
+                stmt.evidence_span = self.triple_scorer.find_evidence_span(stmt, source_text)
+
+        return sorted(statements, key=lambda s: s.confidence_score or 0.0, reverse=True)
+
+    def select_best_beam(
+        self,
+        candidates: list[list[Statement]],
+        source_text: str
+    ) -> list[Statement]:
+        """
+        Select the highest-scoring beam from candidates.
+        """
+        if not candidates:
+            return []
+
+        # Score each candidate and add confidence scores
+        scored_candidates = []
+        for beam in candidates:
+            # Score individual statements
+            for stmt in beam:
+                if stmt.confidence_score is None:
+                    stmt.confidence_score = self.triple_scorer.score_triple(stmt, source_text)
+                if stmt.evidence_span is None:
+                    stmt.evidence_span = self.triple_scorer.find_evidence_span(stmt, source_text)
+
+            beam_score = self.score_beam(beam, source_text)
+            scored_candidates.append((beam_score, beam))
+
+        # Select best
+        scored_candidates.sort(key=lambda x: x[0], reverse=True)
+        return scored_candidates[0][1]
+
+    def merge_beams(
+        self,
+        candidates: list[list[Statement]],
+        source_text: str,
+        top_n: Optional[int] = None
+    ) -> list[Statement]:
+        """
+        Merge top-N beams, keeping high-quality unique triples.
+
+        1. Score all beams
+        2. Take top N beams
+        3. Pool all triples
+        4. Filter by confidence threshold
+        5. Deduplicate (keeping highest confidence)
+        6. Resolve conflicts
+        """
+        if not candidates:
+            return []
+
+        top_n = top_n or self.config.merge_top_n
+
+        # Score each beam
+        scored_beams = []
+        for beam in candidates:
+            for stmt in beam:
+                if stmt.confidence_score is None:
+                    stmt.confidence_score = self.triple_scorer.score_triple(stmt, source_text)
+                if stmt.evidence_span is None:
+                    stmt.evidence_span = self.triple_scorer.find_evidence_span(stmt, source_text)
+
+            beam_score = self.score_beam(beam, source_text)
+            scored_beams.append((beam_score, beam))
+
+        # Sort and take top N
+        scored_beams.sort(key=lambda x: x[0], reverse=True)
+        top_beams = [beam for _, beam in scored_beams[:top_n]]
+
+        # Pool all triples
+        all_statements: list[Statement] = []
+        for beam in top_beams:
+            all_statements.extend(beam)
+
+        # Filter by confidence threshold
+        min_conf = self.config.min_confidence
+        filtered = [s for s in all_statements if (s.confidence_score or 0) >= min_conf]
+
+        # Filter out statements where source_text doesn't support the predicate
+        # This catches model hallucinations where predicate doesn't match the evidence
+        consistent = [
+            s for s in filtered
+            if self._source_text_supports_predicate(s)
+        ]
+
+        # Deduplicate - keep highest confidence for each (subject, predicate, object)
+        # Note: Same subject+predicate with different objects is valid (e.g., "Apple announced X and Y")
+        seen: dict[tuple[str, str, str], Statement] = {}
+        for stmt in consistent:
+            key = (
+                stmt.subject.text.lower(),
+                stmt.predicate.lower(),
+                stmt.object.text.lower()
+            )
+            if key not in seen or (stmt.confidence_score or 0) > (seen[key].confidence_score or 0):
+                seen[key] = stmt
+
+        return list(seen.values())
+
+    def _source_text_supports_predicate(self, stmt: Statement) -> bool:
+        """
+        Check if a statement's source_text contains a lexical trigger for its predicate.
+
+        Returns True if:
+        - source_text is None (no requirement to check)
+        - source_text contains at least one significant word from the predicate
+
+        Returns False if:
+        - source_text is set but contains no words from the predicate
+        """
+        if not stmt.source_text:
+            return True  # No source_text to check
+
+        predicate_words = stmt.predicate.lower().split()
+        source_lower = stmt.source_text.lower()
+
+        # Check if any significant predicate word appears in source_text
+        for word in predicate_words:
+            if len(word) > 2 and word in source_lower:
+                return True
+
+        return False