o 5tiY@sdZddlmZmZmZmZmZddlmZddl m Z ddl m Z m Z mZddlmZGd d d e ZGd d d e ZGd ddeZdS)z;The implementation of the TER metric (Snover et al., 2006).)ListDictSequenceOptionalAny)TercomTokenizer) sum_of_lists)Score SignatureMetric)translation_edit_ratecs&eZdZdZdeffdd ZZS) TERSignaturezA convenience class to represent the reproducibility signature for TER. :param args: key-value dictionary passed from the actual metric instance. argscs^t||jdddddd|j|drdnd |d |d |d  |d ddS)z`TERSignature` initializer.ctnrpnas)casetoknormpunctasiancase_sensitivemixedlctokenizer_signature normalizedno_punct asian_supportN)super__init___abbrupdateinfo)selfr __class__I/home/ubuntu/.local/lib/python3.10/site-packages/sacrebleu/metrics/ter.pyr#s  zTERSignature.__init__)__name__ __module__ __qualname____doc__dictr# __classcell__r*r*r(r+rsrcs.eZdZdZdededeffdd ZZS)TERScorezA convenience class to represent TER scores. :param score: The TER score. :param num_edits: The cumulative number of edits. :param ref_length: The cumulative average reference length. score num_edits ref_lengthcs"td|t||_||_dS)z`TERScore` initializer.TERN)r"r#intr4r5)r'r3r4r5r(r*r+r#:s  zTERScore.__init__)r,r-r.r/floatr#r1r*r*r(r+r23s"r2cseZdZdZeZ     ddededededeeee f fd d Z d e d e fd dZ de e d efddZde e e d efddZde ded e e fddZdee d ee effddZZS)r6aTranslation edit rate (TER). A near-exact reimplementation of the Tercom algorithm, produces identical results on all "sane" outputs. Tercom original implementation: https://github.com/jhclark/tercom The beam edit distance algorithm uses a slightly different approach (we stay around the diagonal which is faster, at least in Python) so in some (extreme) corner cases, the output could differ. Caching in the edit distance is based partly on the PyTer package by Hiroyuki Tanaka (MIT license). (https://github.com/aflc/pyter) :param normalized: Enable character normalization. By default, normalizes a couple of things such as newlines being stripped, retrieving XML encoded characters, and fixing tokenization for punctuation. When 'asian_support' is enabled, also normalizes specific Asian (CJK) character sequences, i.e. split them down to the character level. :param no_punct: Remove punctuation. Can be used in conjunction with 'asian_support' to also remove typical punctuation markers in Asian languages (CJK). :param asian_support: Enable special treatment of Asian characters. This option only has an effect when 'normalized' and/or 'no_punct' is enabled. If 'normalized' is also enabled, then Asian (CJK) characters are split down to the character level. If 'no_punct' is enabled alongside 'asian_support', specific unicode ranges for CJK and full-width punctuations are also removed. :param case_sensitive: If `True`, does not lowercase sentences. :param references: A sequence of reference documents with document being defined as a sequence of reference strings. If given, the reference info will be pre-computed and cached for faster re-computation across many systems. FNrr r!r referencescsdt||_||_||_||_t|j|j|j|jd|_|j|_ |dur0| ||_ dSdS)z`TER` initializer.)rr r!rN) r"r#r rr!rr tokenizer signaturer_cache_references _ref_cache)r'rr r!rr9r(r*r+r#`s  z TER.__init__sentreturncCs||S)zGiven a sentence, apply tokenization if enabled. :param sent: The input sentence string. :return: The pre-processed output string. )r:rstrip)r'r>r*r*r+_preprocess_segmentxzTER._preprocess_segmentstatscCsF|d|d}}|dkr||}n |dkrd}nd}td|||S)zComputes the final score from already aggregated statistics. :param stats: A list or numpy array of segment-level statistics. :return: A `TERScore` object. rr g?gd)r2)r'rC total_editssum_ref_lengthsr3r*r*r+_compute_score_from_statss zTER._compute_score_from_statscCs|t|S)zComputes the final TER score given the pre-computed corpus statistics. :param stats: A list of segment-level statistics :return: A `TERScore` instance. )rGr )r'rCr*r*r+_aggregate_and_computerBzTER._aggregate_and_compute hypothesis ref_kwargsc Cs\d}td}|}|d}|D]}t||\}} || 7}||kr#|}q|t|} || gS)aGiven a (pre-processed) hypothesis sentence and already computed reference words, returns the segment statistics required to compute the full TER score. :param hypothesis: Hypothesis sentence. :param ref_kwargs: A dictionary with `ref_words` key which is a list where each sublist contains reference words. :return: A two-element list that contains the 'minimum number of edits' and 'the average reference length'. rg7yAC ref_words)r7splitrlen) r'rIrJ ref_lengthsbest_num_edits words_hyprK words_refr4ref_len avg_ref_lenr*r*r+_compute_segment_statisticss  zTER._compute_segment_statisticsrefscCs*g}|D] }|||qd|iS)a.Given a list of reference segments, applies pre-processing & tokenization and returns list of tokens for each reference. :param refs: A sequence of strings. :return: A dictionary that will be passed to `_compute_segment_statistics()` through keyword arguments. rK)appendrArL)r'rUrKrefr*r*r+_extract_reference_infoszTER._extract_reference_info)FFFFN)r,r-r.r/r_SIGNATURE_TYPEboolrrstrr#rArr8r2rGrHrrTrrXr1r*r*r(r+r6As8 &r6N)r/typingrrrrrtokenizers.tokenizer_terrutilsr baser r r lib_terrrr2r6r*r*r*r+s