# Copyright The Lightning team. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # referenced from # Library Name: torchtext # Authors: torchtext authors # Date: 2021-11-30 # Link: # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ # Copyright 2020 Memsource # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import re from collections.abc import Iterator, Sequence from functools import lru_cache from typing import List, Optional, Union from torch import Tensor, tensor from torchmetrics.functional.text.helper import ( _flip_trace, _LevenshteinEditDistance, _trace_to_alignment, _validate_inputs, ) # Tercom-inspired limits _MAX_SHIFT_SIZE = 10 _MAX_SHIFT_DIST = 50 # Sacrebleu-inspired limits _MAX_SHIFT_CANDIDATES = 1000 class _TercomTokenizer: """Re-implementation of Tercom Tokenizer in Python 3. See src/ter/core/Normalizer.java in https://github.com/jhclark/tercom Note that Python doesn't support named Unicode blocks so the mapping for relevant blocks was taken from here: https://unicode-table.com/en/blocks/ This implementation follows the implementation from https://github.com/mjpost/sacrebleu/blob/master/sacrebleu/tokenizers/tokenizer_ter.py. """ _ASIAN_PUNCTUATION = r"([\u3001\u3002\u3008-\u3011\u3014-\u301f\uff61-\uff65\u30fb])" _FULL_WIDTH_PUNCTUATION = r"([\uff0e\uff0c\uff1f\uff1a\uff1b\uff01\uff02\uff08\uff09])" def __init__( self, normalize: bool = False, no_punctuation: bool = False, lowercase: bool = True, asian_support: bool = False, ) -> None: """Initialize the tokenizer. Args: normalize: An indication whether a general tokenization to be applied. no_punctuation: An indication whteher a punctuation to be removed from the sentences. lowercase: An indication whether to enable case-insensitivity. asian_support: An indication whether asian characters to be processed. """ self.normalize = normalize self.no_punctuation = no_punctuation self.lowercase = lowercase self.asian_support = asian_support @lru_cache(maxsize=2**16) # noqa: B019 def __call__(self, sentence: str) -> str: """Apply a different tokenization techniques according. Args: sentence: An input sentence to pre-process and tokenize. Return: A tokenized and pre-processed sentence. """ if not sentence: return "" if self.lowercase: sentence = sentence.lower() if self.normalize: sentence = self._normalize_general_and_western(sentence) if self.asian_support: sentence = self._normalize_asian(sentence) if self.no_punctuation: sentence = self._remove_punct(sentence) if self.asian_support: sentence = self._remove_asian_punct(sentence) # Strip extra whitespaces return " ".join(sentence.split()) @staticmethod def _normalize_general_and_western(sentence: str) -> str: """Apply a language-independent (general) tokenization.""" sentence = f" {sentence} " rules = [ (r"\n-", ""), # join lines (r"\n", " "), # handle XML escaped symbols (r""", '"'), (r"&", "&"), (r"<", "<"), (r">", ">"), # tokenize punctuation (r"([{-~[-` -&(-+:-@/])", r" \1 "), # handle possessive (r"'s ", r" 's "), (r"'s$", r" 's"), # tokenize period and comma unless preceded by a digit (r"([^0-9])([\.,])", r"\1 \2 "), # tokenize period and comma unless followed by a digit (r"([\.,])([^0-9])", r" \1 \2"), # tokenize dash when preceded by a digit (r"([0-9])(-)", r"\1 \2 "), ] for pattern, replacement in rules: sentence = re.sub(pattern, replacement, sentence) return sentence @classmethod def _normalize_asian(cls: type["_TercomTokenizer"], sentence: str) -> str: """Split Chinese chars and Japanese kanji down to character level.""" # 4E00—9FFF CJK Unified Ideographs # 3400—4DBF CJK Unified Ideographs Extension A sentence = re.sub(r"([\u4e00-\u9fff\u3400-\u4dbf])", r" \1 ", sentence) # 31C0—31EF CJK Strokes # 2E80—2EFF CJK Radicals Supplement sentence = re.sub(r"([\u31c0-\u31ef\u2e80-\u2eff])", r" \1 ", sentence) # 3300—33FF CJK Compatibility # F900—FAFF CJK Compatibility Ideographs # FE30—FE4F CJK Compatibility Forms sentence = re.sub(r"([\u3300-\u33ff\uf900-\ufaff\ufe30-\ufe4f])", r" \1 ", sentence) # 3200—32FF Enclosed CJK Letters and Months sentence = re.sub(r"([\u3200-\u3f22])", r" \1 ", sentence) # Split Hiragana, Katakana, and KatakanaPhoneticExtensions # only when adjacent to something else # 3040—309F Hiragana # 30A0—30FF Katakana # 31F0—31FF Katakana Phonetic Extensions sentence = re.sub(r"(^|^[\u3040-\u309f])([\u3040-\u309f]+)(?=$|^[\u3040-\u309f])", r"\1 \2 ", sentence) sentence = re.sub(r"(^|^[\u30a0-\u30ff])([\u30a0-\u30ff]+)(?=$|^[\u30a0-\u30ff])", r"\1 \2 ", sentence) sentence = re.sub(r"(^|^[\u31f0-\u31ff])([\u31f0-\u31ff]+)(?=$|^[\u31f0-\u31ff])", r"\1 \2 ", sentence) sentence = re.sub(cls._ASIAN_PUNCTUATION, r" \1 ", sentence) return re.sub(cls._FULL_WIDTH_PUNCTUATION, r" \1 ", sentence) @staticmethod def _remove_punct(sentence: str) -> str: """Remove punctuation from an input sentence string.""" return re.sub(r"[\.,\?:;!\"\(\)]", "", sentence) @classmethod def _remove_asian_punct(cls: type["_TercomTokenizer"], sentence: str) -> str: """Remove asian punctuation from an input sentence string.""" sentence = re.sub(cls._ASIAN_PUNCTUATION, r"", sentence) return re.sub(cls._FULL_WIDTH_PUNCTUATION, r"", sentence) def _preprocess_sentence(sentence: str, tokenizer: _TercomTokenizer) -> str: """Given a sentence, apply tokenization. Args: sentence: The input sentence string. tokenizer: An instance of ``_TercomTokenizer`` handling a sentence tokenization. Return: The pre-processed output sentence string. """ return tokenizer(sentence.rstrip()) def _find_shifted_pairs(pred_words: list[str], target_words: list[str]) -> Iterator[tuple[int, int, int]]: """Find matching word sub-sequences in two lists of words. Ignores sub- sequences starting at the same position. Args: pred_words: A list of a tokenized hypothesis sentence. target_words: A list of a tokenized reference sentence. Return: Yields tuples of ``target_start, pred_start, length`` such that: ``target_words[target_start : target_start + length] == pred_words[pred_start : pred_start + length]`` pred_start: A list of hypothesis start indices. target_start: A list of reference start indices. length: A length of a word span to be considered. """ for pred_start in range(len(pred_words)): for target_start in range(len(target_words)): # this is slightly different from what tercom does but this should # really only kick in in degenerate cases if abs(target_start - pred_start) > _MAX_SHIFT_DIST: continue for length in range(1, _MAX_SHIFT_SIZE): # Check if hypothesis and reference are equal so far if pred_words[pred_start + length - 1] != target_words[target_start + length - 1]: break yield pred_start, target_start, length # Stop processing once a sequence is consumed. _hyp = len(pred_words) == pred_start + length _ref = len(target_words) == target_start + length if _hyp or _ref: break def _handle_corner_cases_during_shifting( alignments: dict[int, int], pred_errors: list[int], target_errors: list[int], pred_start: int, target_start: int, length: int, ) -> bool: """Return ``True`` if any of corner cases has been met. Otherwise, ``False`` is returned. Args: alignments: A dictionary mapping aligned positions between a reference and a hypothesis. pred_errors: A list of error positions in a hypothesis. target_errors: A list of error positions in a reference. pred_start: A hypothesis start index. target_start: A reference start index. length: A length of a word span to be considered. Return: An indication whether any of conrner cases has been met. """ # don't do the shift unless both the hypothesis was wrong and the # reference doesn't match hypothesis at the target position if sum(pred_errors[pred_start : pred_start + length]) == 0: return True if sum(target_errors[target_start : target_start + length]) == 0: return True # don't try to shift within the subsequence return pred_start <= alignments[target_start] < pred_start + length def _perform_shift(words: list[str], start: int, length: int, target: int) -> list[str]: """Perform a shift in ``words`` from ``start`` to ``target``. Args: words: A words to shift. start: An index where to start shifting from. length: A number of how many words to be considered. target: An index where to end shifting. Return: A list of shifted words. """ def _shift_word_before_previous_position(words: list[str], start: int, target: int, length: int) -> list[str]: return words[:target] + words[start : start + length] + words[target:start] + words[start + length :] def _shift_word_after_previous_position(words: list[str], start: int, target: int, length: int) -> list[str]: return words[:start] + words[start + length : target] + words[start : start + length] + words[target:] def _shift_word_within_shifted_string(words: list[str], start: int, target: int, length: int) -> list[str]: shifted_words = words[:start] shifted_words += words[start + length : length + target] shifted_words += words[start : start + length] shifted_words += words[length + target :] return shifted_words if target < start: return _shift_word_before_previous_position(words, start, target, length) if target > start + length: return _shift_word_after_previous_position(words, start, target, length) return _shift_word_within_shifted_string(words, start, target, length) def _shift_words( pred_words: list[str], target_words: list[str], cached_edit_distance: _LevenshteinEditDistance, checked_candidates: int, ) -> tuple[int, list[str], int]: """Attempt to shift words to match a hypothesis with a reference. It returns the lowest number of required edits between a hypothesis and a provided reference, a list of shifted words and number of checked candidates. Note that the filtering of possible shifts and shift selection are heavily based on somewhat arbitrary heuristics. The code here follows as closely as possible the logic in Tercom, not always justifying the particular design choices. The paragraph copied from https://github.com/mjpost/sacrebleu/blob/master/sacrebleu/metrics/lib_ter.py. Args: pred_words: A list of tokenized hypothesis sentence. target_words: A list of lists of tokenized reference sentences. cached_edit_distance: A pre-computed edit distance between a hypothesis and a reference. checked_candidates: A number of checked hypothesis candidates to match a provided reference. Return: best_score: The best (lowest) number of required edits to match hypothesis and reference sentences. shifted_words: A list of shifted words in hypothesis sentences. checked_candidates: A number of checked hypothesis candidates to match a provided reference. """ edit_distance, inverted_trace = cached_edit_distance(pred_words) trace = _flip_trace(inverted_trace) alignments, target_errors, pred_errors = _trace_to_alignment(trace) best: Optional[tuple[int, int, int, int, list[str]]] = None for pred_start, target_start, length in _find_shifted_pairs(pred_words, target_words): if _handle_corner_cases_during_shifting( alignments, pred_errors, target_errors, pred_start, target_start, length ): continue prev_idx = -1 for offset in range(-1, length): if target_start + offset == -1: idx = 0 elif target_start + offset in alignments: idx = alignments[target_start + offset] + 1 # offset is out of bounds => aims past reference else: break # Skip idx if already tried if idx == prev_idx: continue prev_idx = idx shifted_words = _perform_shift(pred_words, pred_start, length, idx) # Elements of the tuple are designed to replicate Tercom ranking of shifts: candidate = ( edit_distance - cached_edit_distance(shifted_words)[0], # highest score first length, # then, longest match first -pred_start, # then, earliest match first -idx, # then, earliest target position first shifted_words, ) checked_candidates += 1 if not best or candidate > best: best = candidate if checked_candidates >= _MAX_SHIFT_CANDIDATES: break if not best: return 0, pred_words, checked_candidates best_score, _, _, _, shifted_words = best return best_score, shifted_words, checked_candidates def _translation_edit_rate(pred_words: list[str], target_words: list[str]) -> Tensor: """Compute translation edit rate between hypothesis and reference sentences. Args: pred_words: A list of a tokenized hypothesis sentence. target_words: A list of lists of tokenized reference sentences. Return: A number of required edits to match hypothesis and reference sentences. """ if len(target_words) == 0: return tensor(0.0) cached_edit_distance = _LevenshteinEditDistance(target_words) num_shifts = 0 checked_candidates = 0 input_words = pred_words while True: # do shifts until they stop reducing the edit distance delta, new_input_words, checked_candidates = _shift_words( input_words, target_words, cached_edit_distance, checked_candidates ) if checked_candidates >= _MAX_SHIFT_CANDIDATES or delta <= 0: break num_shifts += 1 input_words = new_input_words edit_distance, _ = cached_edit_distance(input_words) total_edits = num_shifts + edit_distance return tensor(total_edits) def _compute_sentence_statistics(pred_words: list[str], target_words: list[list[str]]) -> tuple[Tensor, Tensor]: """Compute sentence TER statistics between hypothesis and provided references. Args: pred_words: A list of tokenized hypothesis sentence. target_words: A list of lists of tokenized reference sentences. Return: best_num_edits: The best (lowest) number of required edits to match hypothesis and reference sentences. avg_tgt_len: Average length of tokenized reference sentences. """ tgt_lengths = tensor(0.0) best_num_edits = tensor(2e16) for tgt_words in target_words: num_edits = _translation_edit_rate(tgt_words, pred_words) tgt_lengths += len(tgt_words) if num_edits < best_num_edits: best_num_edits = num_edits avg_tgt_len = tgt_lengths / len(target_words) return best_num_edits, avg_tgt_len def _compute_ter_score_from_statistics(num_edits: Tensor, tgt_length: Tensor) -> Tensor: """Compute TER score based on pre-computed a number of edits and an average reference length. Args: num_edits: A number of required edits to match hypothesis and reference sentences. tgt_length: An average length of reference sentences. Return: A corpus-level TER score or 1 if reference_length == 0. """ if tgt_length > 0 and num_edits > 0: return num_edits / tgt_length if tgt_length == 0 and num_edits > 0: return tensor(1.0) return tensor(0.0) def _ter_update( preds: Union[str, Sequence[str]], target: Sequence[Union[str, Sequence[str]]], tokenizer: _TercomTokenizer, total_num_edits: Tensor, total_tgt_length: Tensor, sentence_ter: Optional[List[Tensor]] = None, ) -> tuple[Tensor, Tensor, Optional[List[Tensor]]]: """Update TER statistics. Args: preds: An iterable of hypothesis corpus. target: An iterable of iterables of reference corpus. tokenizer: An instance of ``_TercomTokenizer`` handling a sentence tokenization. total_num_edits: A total number of required edits to match hypothesis and reference sentences. total_tgt_length: A total average length of reference sentences. sentence_ter: A list of sentence-level TER values Return: total_num_edits: A total number of required edits to match hypothesis and reference sentences. total_tgt_length: A total average length of reference sentences. sentence_ter: (Optionally) A list of sentence-level TER. Raises: ValueError: If length of ``preds`` and ``target`` differs. """ target, preds = _validate_inputs(target, preds) for pred, tgt in zip(preds, target): tgt_words_: list[list[str]] = [_preprocess_sentence(_tgt, tokenizer).split() for _tgt in tgt] pred_words_: list[str] = _preprocess_sentence(pred, tokenizer).split() num_edits, tgt_length = _compute_sentence_statistics(pred_words_, tgt_words_) total_num_edits += num_edits total_tgt_length += tgt_length if sentence_ter is not None: sentence_ter.append(_compute_ter_score_from_statistics(num_edits, tgt_length).unsqueeze(0)) return total_num_edits, total_tgt_length, sentence_ter def _ter_compute(total_num_edits: Tensor, total_tgt_length: Tensor) -> Tensor: """Compute TER based on pre-computed a total number of edits and a total average reference length. Args: total_num_edits: A total number of required edits to match hypothesis and reference sentences. total_tgt_length: A total average length of reference sentences. Return: A corpus-level TER score. """ return _compute_ter_score_from_statistics(total_num_edits, total_tgt_length) def translation_edit_rate( preds: Union[str, Sequence[str]], target: Sequence[Union[str, Sequence[str]]], normalize: bool = False, no_punctuation: bool = False, lowercase: bool = True, asian_support: bool = False, return_sentence_level_score: bool = False, ) -> Union[Tensor, tuple[Tensor, List[Tensor]]]: """Calculate Translation edit rate (`TER`_) of machine translated text with one or more references. This implementation follows the implementations from https://github.com/mjpost/sacrebleu/blob/master/sacrebleu/metrics/ter.py. The `sacrebleu` implementation is a near-exact reimplementation of the Tercom algorithm, produces identical results on all "sane" outputs. Args: preds: An iterable of hypothesis corpus. target: An iterable of iterables of reference corpus. normalize: An indication whether a general tokenization to be applied. no_punctuation: An indication whteher a punctuation to be removed from the sentences. lowercase: An indication whether to enable case-insensitivity. asian_support: An indication whether asian characters to be processed. return_sentence_level_score: An indication whether a sentence-level TER to be returned. Return: A corpus-level translation edit rate (TER). (Optionally) A list of sentence-level translation_edit_rate (TER) if `return_sentence_level_score=True`. Example: >>> preds = ['the cat is on the mat'] >>> target = [['there is a cat on the mat', 'a cat is on the mat']] >>> translation_edit_rate(preds, target) tensor(0.1538) References: [1] A Study of Translation Edit Rate with Targeted Human Annotation by Mathew Snover, Bonnie Dorr, Richard Schwartz, Linnea Micciulla and John Makhoul `TER`_ """ if not isinstance(normalize, bool): raise ValueError(f"Expected argument `normalize` to be of type boolean but got {normalize}.") if not isinstance(no_punctuation, bool): raise ValueError(f"Expected argument `no_punctuation` to be of type boolean but got {no_punctuation}.") if not isinstance(lowercase, bool): raise ValueError(f"Expected argument `lowercase` to be of type boolean but got {lowercase}.") if not isinstance(asian_support, bool): raise ValueError(f"Expected argument `asian_support` to be of type boolean but got {asian_support}.") tokenizer: _TercomTokenizer = _TercomTokenizer(normalize, no_punctuation, lowercase, asian_support) total_num_edits = tensor(0.0) total_tgt_length = tensor(0.0) sentence_ter: Optional[List[Tensor]] = [] if return_sentence_level_score else None total_num_edits, total_tgt_length, sentence_ter = _ter_update( preds, target, tokenizer, total_num_edits, total_tgt_length, sentence_ter, ) ter_score = _ter_compute(total_num_edits, total_tgt_length) if sentence_ter: return ter_score, sentence_ter return ter_score