o .wiD@sZddlZddlZddlmZddlmZddlmZmZm Z ddl m Z m Z m Z ddlmZddlmZded ed efd d Z   d2dedededededed efddZded efddZded efddZdee d e fdd Zd!e eeefd"ee eeefd#ed$d ee eeefee eeefffd%d&Z   d2ded'e eeefdedededed e fd(d)Z *    d3d!e eeefd"ee eeefd#ed$dedededed+eee d ee fd,d-Z * .   d4d!e eeefd"ee eeefd#ed$d/ededededed e e ee e fffd0d1ZdS)5N)Sequence)inf)ListOptionalUnion)Tensorstacktensor)Literal)_validate_inputs preds_word target_wordreturncCs t||kS)a.Distance measure used for substitutions/identity operation. Code adapted from https://github.com/rwth-i6/ExtendedEditDistance/blob/master/EED.py. Args: preds_word: hypothesis word string target_word: reference word string Return: 0 for match, 1 for no match )int)r r r]/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/torchmetrics/functional/text/eed.py_distance_between_wordses r@333333?皙??hyprefalpharhodeletion insertionc sfdgt|d}dgt|d}d|d<tgt|d}tdt|dD]n} tt|dD]2} | dkr\t|| d||| dt|| d|| d|| ||| <q2|| d|| <q2|t|} || d7<|| ddkr||| fdd|D}|}tgt|d}q(|td d |D} td|d| tt|| S) a8Compute extended edit distance score for two lists of strings: hyp and ref. Code adapted from: https://github.com/rwth-i6/ExtendedEditDistance/blob/master/EED.py. Args: hyp: A hypothesis string ref: A reference string alpha: optimal jump penalty, penalty for jumps between characters rho: coverage cost, penalty for repetition of characters deletion: penalty for deletion of character insertion: penalty for insertion or substitution of character Return: Extended edit distance score as float rr csg|]}t|qSr)min.0xjumprr sz!_eed_function..css |] }|dkr |ndVqdS)rrNrr"rrr sz _eed_function..)lenrranger!rindexsumfloat) rrrrrrnumber_of_visitsrownext_rowwi min_indexcoveragerr%r _eed_functionus,$   "r5sentencecCst|tstdt|d|}gd}|D] \}}|||}qgd}|D] \}}t|||}q*gd}|D] \}}|||}qr?r@rA unicodedata normalize)r6rrr_preprocess_jas  rMsentence_level_scorescCs(t|dkr tdSt|tt|S)zReduction for extended edit distance. Args: sentence_level_scores: list of sentence-level scores as floats Return: average of scores as a tensor rr)r)r r,)rNrrr _eed_computes rOpredstargetlanguage)enjacsft||d\}}|dkrtn|dkrtntd|fdd|D}fdd|D}||fS)auPreprocess strings according to language requirements. Args: preds: An iterable of hypothesis corpus. target: An iterable of iterables of reference corpus. language: Language used in sentences. Only supports English (en) and Japanese (ja) for now. Defaults to en Return: Tuple of lists that contain the cleaned strings for target and preds Raises: ValueError: If a different language than ``'en'`` or ``'ja'`` is used ValueError: If length of target not equal to length of preds ValueError: If objects in reference and hypothesis corpus are not strings )hypothesis_corpus ref_corpusrSrTz?Expected argument `language` to either be `en` or `ja` but got cg|]}|qSrr)r#predpreprocess_functionrrr'z)_preprocess_sentences..csg|] }fdd|DqS)crWrr)r#rrYrrr'r[z4_preprocess_sentences...r)r# referencerYrrr's)r rIrMr?)rPrQrRrrYr_preprocess_sentencessr] target_wordsc Cs4t}|D]}t||||||}||kr|}qt|S)aCompute scores for ExtendedEditDistance. Args: target_words: An iterable of reference words preds_word: A hypothesis word alpha: An optimal jump penalty, penalty for jumps between characters rho: coverage cost, penalty for repetition of characters deletion: penalty for deletion of character insertion: penalty for insertion or substitution of character Return: best_score: best (lowest) sentence-level score as a Tensor )rr5r ) r r^rrrr best_scorer\scorerrr_compute_sentence_statistics#srarS sentence_eedc Cslt|||\}}|durg}dt|t|dfvr|St||D]\}} t|| ||||} || q!|S)aCompute scores for ExtendedEditDistance. Args: preds: An iterable of hypothesis corpus target: An iterable of iterables of reference corpus language: Language used in sentences. Only supports English (en) and Japanese (ja) for now. Defaults to en alpha: optimal jump penalty, penalty for jumps between characters rho: coverage cost, penalty for repetition of characters deletion: penalty for deletion of character insertion: penalty for insertion or substitution of character sentence_eed: list of sentence-level scores Return: individual sentence scores as a list of Tensors Nr)r]r)zipraappend) rPrQrRrrrrrb hypothesisr^r`rrr _eed_updateCs rfFreturn_sentence_level_scorec Cs|tgd||||gD]\}} t| trt| tr%| dkr%td|dq t|||||||} t| } |r<| t| fS| S)uXCompute extended edit distance score (`ExtendedEditDistance`_) [1] for strings or list of strings. The metric utilises the Levenshtein distance and extends it by adding a jump operation. Args: preds: An iterable of hypothesis corpus. target: An iterable of iterables of reference corpus. language: Language used in sentences. Only supports English (en) and Japanese (ja) for now. Defaults to en return_sentence_level_score: An indication of whether sentence-level EED score is to be returned. alpha: optimal jump penalty, penalty for jumps between characters rho: coverage cost, penalty for repetition of characters deletion: penalty for deletion of character insertion: penalty for insertion or substitution of character Return: Extended edit distance score as a tensor Example: >>> from torchmetrics.functional.text import extended_edit_distance >>> preds = ["this is the prediction", "here is an other sample"] >>> target = ["this is the reference", "here is another one"] >>> extended_edit_distance(preds=preds, target=target) tensor(0.3078) References: [1] P. Stanchev, W. Wang, and H. Ney, “EED: Extended Edit Distance Measure for Machine Translation”, submitted to WMT 2019. `ExtendedEditDistance`_ )rrrrrz Parameter `z)` is expected to be a non-negative float.)rcr=r-r?rfrOr) rPrQrRrgrrrr param_nameparamrNaveragerrrextended_edit_distancems( rk)rrrr)rSrrrrN)rSFrrrr) rCrKcollections.abcrmathrtypingrrrtorchrrr typing_extensionsr #torchmetrics.functional.text.helperr r>rrr-r5rIrMrOtupler]rarfboolrkrrrrsX     :-& ) #  -