Method annotate_span

data_convert/format/text2target.py:174–231

src: et + + source text 为 t5-base 分隔符 traget: ((Tri)) 只对tri进行生成, 如果同一个et存在多个Tri, 则生成 ((Tri)(Tri)...) :param tokens: US President George W. Bush told Canadian Prime Minister Jean Chretien by telephone Monday that he looked forward to see

(tokens, predicate_arguments, mark_tree=False, zh=False, isTest = False)

    
    @staticmethod
    def annotate_span(tokens, predicate_arguments, mark_tree=False, zh=False, isTest = False):
        """
        src: et + </s> + source text    </s> 为 t5-base 分隔符
        traget: ((Tri))   只对tri进行生成, 如果同一个et存在多个Tri, 则生成 ((Tri)(Tri)...)
        :param tokens:
            US President George W. Bush told Canadian Prime Minister Jean Chretien by telephone Monday that he looked forward
            to seeing him at the upcoming summit of major industrialized nations and Russia , the White House said Tuesday .
        :param predicate_arguments:

        :return:
        """

        token_separator = '' if zh else ' '

        event_str_rep_list = list()
        
        source_list = []
        target_list = []

        # 若出现单句多个相同 type 事件, 则将其合并
        et_tri_dict = {} # {et + src: tri_list}
        et_set = set() # 文本已经包含的事件类型

        # 加载 schema文件中的信息, 目前每次都会读取文件, 很低效, 后续将这一步改为传参优化处理
        et_list = EventSchema.read_from_file("data/raw_data/dyiepp_ace2005/event.schema").type_list

        # 针对事件类型遍历制作训练样本
        for et in et_list:
            et_tri_dict[et + " </s> " + token_separator.join(tokens)] = set()


        # 遍历event
        for predicate_argument in predicate_arguments:
            et = predicate_argument['type']
            et_set.add(et)

            tri_text = get_str_from_tokens(predicate_argument['tokens'], tokens, separator=token_separator) # 此处的 predicate_argument['tokens'] 为 [start, end](多个单词), 或者[start] (一个单词)

            et_tri_dict[et + " </s> " + token_separator.join(tokens)].add(tri_text) 

        
        for src, tri_set in et_tri_dict.items():
            if not tri_set: continue # 过滤掉不包含触发词的样本
            source_list.append(src)
            tmp_list = []
            for tri_text in tri_set:
                tmp_list.append(' '.join([type_start, tri_text, type_end]))
            target_text = f'{type_start} ' + " ".join(tmp_list) + f' {type_end}'
            target_list.append(target_text)
        
        # negative sample on tri train data
        if not isTest:
            for tmp_et in random.sample(set(et_list) - et_set, 6):
                source_list.append(tmp_et + " </s> " + token_separator.join(tokens))
                target_list.append(f'{type_start} ' + " ".join([]) + f' {type_end}')


        return source_list, target_list