aste-v2 / src /alignment.py
Matthew Franglen
Casefold before calculating indices
4104e84
from dataclasses import asdict
from typing import Optional
import Levenshtein
import pandas as pd
from .types import CharacterIndices, Triplet, WordSpans
def find_closest_text(
*,
original: pd.Series,
replacement: pd.Series,
) -> pd.Series:
# Returns a series of the replacement values aligned to the original values
no_space_replacements = {text.replace(" ", ""): text for text in replacement}
original_text = original.str.replace(" ", "")
result = original_text.map(no_space_replacements)
non_perfect_matches = result.isna().sum()
assert non_perfect_matches / len(original) <= 0.20, (
"Poor alignment with replacement text. "
f"{non_perfect_matches:,} of {len(original),} rows did not match well"
)
def closest(text: str) -> str:
distances = replacement.apply(
lambda comparison: Levenshtein.distance(text, comparison)
)
return replacement.iloc[distances.argmin()]
result.loc[result.isna()] = original_text[result.isna()].apply(closest)
result = result.str.strip()
return result
def to_character_indices_series(row: pd.Series) -> pd.Series:
result = to_character_indices(triplet=row.triples, text=row.text)
return pd.Series(asdict(result))
def to_character_indices(
*,
triplet: Triplet,
text: str,
) -> CharacterIndices:
aspect_span, opinion_span, _ = triplet
assert _is_sequential(aspect_span), f"aspect span not sequential: {aspect_span}"
assert _is_sequential(opinion_span), f"opinion span not sequential: {opinion_span}"
spans = WordSpans.make(text)
aspect_start_index, aspect_end_index = spans.to_indices(aspect_span)
aspect_term = text[aspect_start_index : aspect_end_index + 1]
opinion_start_index, opinion_end_index = spans.to_indices(opinion_span)
opinion_term = text[opinion_start_index : opinion_end_index + 1]
return CharacterIndices(
aspect_start_index=aspect_start_index,
aspect_end_index=aspect_end_index,
aspect_term=aspect_term,
opinion_start_index=opinion_start_index,
opinion_end_index=opinion_end_index,
opinion_term=opinion_term,
)
def to_aligned_character_indices_series(row: pd.Series) -> pd.Series:
indices = to_character_indices(triplet=row.triples, text=row.original)
result = to_aligned_character_indices(
original=row.original,
replacement=row.text,
original_indices=indices,
)
return pd.Series(asdict(result))
def to_aligned_character_indices(
*,
original: str,
replacement: str,
original_indices: CharacterIndices,
) -> CharacterIndices:
indices = _aligned_character_indices(original=original, replacement=replacement)
aspect_start_index = _aligned_start_index(
text=replacement,
original_index=original_indices.aspect_start_index,
indices=indices,
)
aspect_end_index = _aligned_end_index(
text=replacement,
original_index=original_indices.aspect_end_index,
indices=indices,
)
aspect_term = replacement[aspect_start_index : aspect_end_index + 1]
opinion_start_index = _aligned_start_index(
text=replacement,
original_index=original_indices.opinion_start_index,
indices=indices,
)
opinion_end_index = _aligned_end_index(
text=replacement,
original_index=original_indices.opinion_end_index,
indices=indices,
)
opinion_term = replacement[opinion_start_index : opinion_end_index + 1]
return CharacterIndices(
aspect_start_index=aspect_start_index,
aspect_end_index=aspect_end_index,
aspect_term=aspect_term,
opinion_start_index=opinion_start_index,
opinion_end_index=opinion_end_index,
opinion_term=opinion_term,
)
def _is_sequential(span: tuple[int, ...]) -> bool:
return all(span[index + 1] - span[index] == 1 for index in range(len(span) - 1))
def _aligned_character_indices(original: str, replacement: str) -> list[Optional[int]]:
original = original.casefold()
replacement = replacement.casefold()
indices: list[Optional[int]] = list(range(len(original)))
for operation, _source_position, destination_position in Levenshtein.editops(
original, replacement
):
if operation == "replace":
indices[destination_position] = None
elif operation == "insert":
indices.insert(destination_position, None)
elif operation == "delete":
del indices[destination_position]
return indices
def _aligned_start_index(
text: str, original_index: int, indices: list[Optional[int]]
) -> int:
closest_after = min(
index for index in indices if index is not None and index >= original_index
)
index = indices.index(closest_after)
# Not every character in the original text is aligned to a character in the
# replacement text. The replacement text may have deleted it, or replaced
# it. Can step back through each letter until the word boundary is found or
# an aligned character is found.
while index > 0:
if indices[index - 1] is not None:
break
if text[index - 1] == " ":
break
index -= 1
return index
def _aligned_end_index(
text: str, original_index: int, indices: list[Optional[int]]
) -> int:
closest_before = max(
index for index in indices if index is not None and index <= original_index
)
index = indices.index(closest_before)
# Not every character in the original text is aligned to a character in the
# replacement text. The replacement text may have deleted it, or replaced
# it. Can step back through each letter until the word boundary is found or
# an aligned character is found.
while index < len(indices) - 1:
if indices[index + 1] is not None:
break
if text[index + 1] == " ":
break
index += 1
return index