File size: 8,099 Bytes
bf64b03 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 | """
CitationModule: Understands scientific citation structure.
Detects citation spans, tracks provenance, and estimates claim confidence.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import re
from typing import Optional, Tuple, List
class CitationModule(nn.Module):
"""
Understands scientific citation structure.
- Detects citation spans [Author, Year] or (1) style
- Learns that cited claims carry different epistemic weight
- Distinguishes established facts vs recent/contested findings
- Tracks claim provenance through the context window
"""
def __init__(self, d_model: int):
"""
Initialize CitationModule.
Args:
d_model: Model dimension
"""
super().__init__()
self.d_model = d_model
# Citation span detector (3 classes: none, inline, reference)
# Inline: (Author, Year) or [1]
# Reference: full citation at end of paper
self.citation_detector = nn.Linear(d_model, 3)
# Provenance gate: modulates information flow based on citation context
self.provenance_gate = nn.Linear(d_model, d_model)
# Claim confidence head: estimates how well-supported a claim is
self.confidence_head = nn.Linear(d_model, 1)
# Citation type embeddings
self.citation_type_embedding = nn.Embedding(3, d_model)
# Initialize weights
self._initialize_weights()
def _initialize_weights(self):
"""Initialize weights."""
for module in [self.citation_detector, self.provenance_gate, self.confidence_head, self.citation_type_embedding]:
if hasattr(module, 'weight'):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
if hasattr(module, 'bias') and module.bias is not None:
nn.init.zeros_(module.bias)
def detect_citation_spans(
self,
text: str,
) -> List[Tuple[int, int, str]]:
"""
Detect citation spans in text.
Supports: (Author, Year), [1], [Author, Year], et al.
Args:
text: Input text string
Returns:
List of (start_char, end_char, citation_type)
citation_type: "inline" or "reference"
"""
spans = []
# Pattern 1: (Author, Year) or (Author Year)
for match in re.finditer(r'\([A-Za-z\s]+(?:et al\.)?,?\s*\d{4}\)', text):
spans.append((match.start(), match.end(), "inline"))
# Pattern 2: [1] or [1-3] or [1,2,3]
for match in re.finditer(r'\[\d+(?:[-,]\d+)*\]', text):
spans.append((match.start(), match.end(), "inline"))
# Pattern 3: [Author, Year]
for match in re.finditer(r'\[[A-Za-z\s]+,?\s*\d{4}\]', text):
spans.append((match.start(), match.end(), "inline"))
# Pattern 4: et al. (often indicates citation)
for match in re.finditer(r'\bet al\.\b', text):
spans.append((match.start(), match.end(), "inline"))
return spans
def forward(
self,
x: torch.Tensor,
text: Optional[List[str]] = None,
citation_spans: Optional[List[List[Tuple[int, int, str]]]] = None,
) -> torch.Tensor:
"""
Forward pass through citation module.
Args:
x: Input tensor (batch, seq_len, d_model)
text: Optional original text strings
citation_spans: Optional pre-computed citation spans per batch
Returns:
Citation-enhanced representation (batch, seq_len, d_model)
"""
batch, seq_len, d_model = x.shape
# Detect citation spans
if citation_spans is None and text is not None:
citation_spans = []
for b in range(batch):
spans = self.detect_citation_spans(text[b])
# Convert char spans to token spans (approximate)
token_spans = []
for start_char, end_char, ctype in spans:
start_tok = max(0, start_char // 4)
end_tok = min(seq_len, end_char // 4 + 1)
token_spans.append((start_tok, end_tok, ctype))
citation_spans.append(token_spans)
# Compute citation type logits
citation_logits = self.citation_detector(x) # (batch, seq_len, 3)
citation_probs = F.softmax(citation_logits, dim=-1)
# Apply citation-specific transformations
output = x.clone()
if citation_spans:
for b in range(batch):
spans_b = citation_spans[b] if b < len(citation_spans) else []
for start_tok, end_tok, ctype in spans_b:
if end_tok <= start_tok:
continue
# Get citation type embedding
if ctype == "inline":
type_id = 1
elif ctype == "reference":
type_id = 2
else:
type_id = 0
type_emb = self.citation_type_embedding(
torch.tensor(type_id, device=x.device)
)
# Apply provenance gate to citation span
span_slice = x[b, start_tok:end_tok, :]
gated = span_slice * torch.sigmoid(self.provenance_gate(span_slice))
# Add citation type embedding
gated = gated + type_emb.unsqueeze(0).unsqueeze(0)
output[b, start_tok:end_tok, :] = gated
# Compute confidence scores (for auxiliary loss)
confidence = torch.sigmoid(self.confidence_head(x)) # (batch, seq_len, 1)
return output, confidence
def compute_citation_loss(
self,
x: torch.Tensor,
citation_mask: torch.Tensor,
confidence: torch.Tensor,
) -> torch.Tensor:
"""
Compute auxiliary loss for citation detection and confidence.
Args:
x: Input tensor (batch, seq_len, d_model)
citation_mask: Ground truth citation mask (batch, seq_len), 1 if token is in citation
confidence: Predicted confidence scores (batch, seq_len, 1)
Returns:
Combined citation loss
"""
# Citation detection loss
logits = self.citation_detector(x) # (batch, seq_len, 3)
detection_loss = F.cross_entropy(
logits.view(-1, 3),
citation_mask.long().view(-1),
)
# Confidence calibration loss (encourage high confidence for true citations)
confidence_loss = F.mse_loss(
confidence.squeeze(-1),
citation_mask.float(),
)
return detection_loss + 0.1 * confidence_loss
def test_citation_module():
"""Test CitationModule."""
d_model = 512
batch_size = 2
seq_len = 128
module = CitationModule(d_model)
x = torch.randn(batch_size, seq_len, d_model)
text = [
"The theory of relativity (Einstein, 1905) revolutionized physics. See also [1, 2].",
"According to Smith et al., the results are significant. Further reading: [Doe, 2020]."
]
output, confidence = module(x, text=text)
print(f"Input shape: {x.shape}")
print(f"Output shape: {output.shape}")
print(f"Confidence shape: {confidence.shape}")
assert output.shape == x.shape
assert confidence.shape == (batch_size, seq_len, 1)
# Test loss
citation_mask = torch.zeros(batch_size, seq_len)
citation_mask[0, 20:25] = 1.0 # Simulate citation span
citation_mask[1, 10:18] = 1.0
loss = module.compute_citation_loss(x, citation_mask, confidence)
print(f"Citation loss: {loss.item():.4f}")
print("CitationModule test passed!")
if __name__ == "__main__":
test_citation_module()
|