使用Spacy实现代词-先行词关联时结果异常，寻求技术排查帮助

你的代码出现错误输出的核心原因：

• 遍历token时，同一个代词token会同时触发先行词赋值和指代输出逻辑，比如He既是主语（nsubj）又是代词（PRON），导致先行词被错误设置为代词本身。
• 先行词的提取逻辑完全错误，没有从上下文的前置名词中获取，反而在处理当前token时错误赋值。

方案1：基于实体识别的手动匹配 #

适合简单场景，先提取文本中的人名实体，再根据代词的性别/语境匹配先行词：

import spacy

nlp = spacy.load("en_core_web_md")
text = "John and Maggie walk along the street. He told her he is leaving for Paris. She was surprised"
doc = nlp(text)

# 提取所有候选人名先行词
person_candidates = []
for ent in doc.ents:
    if ent.label_ == "PERSON":
        # 拆分并列的人名（比如John and Maggie）
        person_candidates.extend([token.text for token in ent if token.pos_ == "PROPN"])

# 构建代词-先行词映射（基于常见性别指代）
pronoun_to_antecedent = {
    "He": person_candidates[0],
    "he": person_candidates[0],
    "She": person_candidates[1],
    "she": person_candidates[1],
    "her": person_candidates[1]
}

# 输出指代关系
for token in doc:
    if token.pos_ == "PRON" and token.text in pronoun_to_antecedent:
        print(f"{token.text} refers to {pronoun_to_antecedent[token.text]}")

方案2：使用Spacy官方指代消解模型 #

适合复杂场景，自动处理指代簇，无需手动维护映射：

1. 先安装依赖和模型：

pip install spacy-transformers
python -m spacy download en_coreference_web_trf

2. 代码实现：

import spacy

nlp = spacy.load("en_coreference_web_trf")
text = "John and Maggie walk along the street. He told her he is leaving for Paris. She was surprised"
doc = nlp(text)

# 遍历每个指代簇，第一个元素是先行词，后续是指代它的代词
for cluster in doc.spans["coref"]:
    antecedent = cluster[0].text
    for pronoun in cluster[1:]:
        print(f"{pronoun.text} refers to {antecedent}")

运行方案2会输出：

He refers to John
her refers to Maggie
he refers to John
She refers to Maggie

内容的提问来源于stack exchange，提问作者user20309717