如何配置SpaCy，使歧义的VERB/NOUN优先被标记为VERB？

以下是我的代码：
SpaCy版本：spacy==3.6.1

import spacy

try:
    nlp = spacy.load("en_core_web_sm")
except Exception as e:
    print(f"An error occurred: {str(e)}")
    print(
        print(
            "SpaCy model not found. Please run `python -m spacy download en_core_web_sm`"
        )
    )

doc = nlp("Buy groceries.")
for sent in doc.sents:
    print(f"Sentence: {sent.text}")
    for token in sent:
        print(f"{token.text}: {token.pos_}")

运行后，Buy被标记为NOUN。

我了解这属于词性歧义问题，请问如何设置才能让所有存在VERB/NOUN歧义的词优先被标记为VERB？尤其是当该词位于句首、前面是标点符号，且后面紧跟另一个名词的场景。

1. 用自定义匹配规则修正特定场景 #

针对你描述的句首、后接名词的歧义场景，可以用SpaCy的Matcher组件添加规则，强制将符合条件的词标记为VERB：

import spacy
from spacy.matcher import Matcher

nlp = spacy.load("en_core_web_sm")
matcher = Matcher(nlp.vocab)

# 定义匹配规则：句首词+后续名词，且句首词本身存在VERB/NOUN歧义
pattern = [
    {"IS_SENT_START": True, "POS": {"IN": ["NOUN", "VERB"]}},
    {"POS": "NOUN"}
]
matcher.add("VERB_AMBIGUITY_FIX", [pattern])

def fix_verb_ambiguity(doc):
    matches = matcher(doc)
    for match_id, start, end in matches:
        token = doc[start]
        # 验证该词确实具备动词词性的可能
        if any(morph.pos == "VERB" for morph in token.morph.get("POS")):
            token.pos_ = "VERB"
            token.tag_ = "VB"  # 对应动词原形的Penn Treebank标记
    return doc

# 将自定义处理器添加到管道，放在词性标注器之后
nlp.add_pipe(fix_verb_ambiguity, after="tagger")

# 测试
doc = nlp("Buy groceries.")
for sent in doc.sents:
    print(f"Sentence: {sent.text}")
    for token in sent:
        print(f"{token.text}: {token.pos_}")

2. 针对祈使句场景优化 #

你提到的场景大多是祈使句（无主语、句首动词引导），可以直接针对祈使句特征写规则：

def fix_imperative_verbs(doc):
    for sent in doc.sents:
        # 匹配句首名词+后续名词的结构，且句首词为原形
        if len(sent) >=2 and sent[0].pos_ == "NOUN" and sent[1].pos_ == "NOUN":
            if sent[0].lemma_ == sent[0].text:
                sent[0].pos_ = "VERB"
                sent[0].tag_ = "VB"
    return doc

nlp.add_pipe(fix_imperative_verbs, after="tagger")

3. 微调预训练模型（彻底解决多场景歧义） #

如果需要覆盖更多歧义场景，最彻底的方式是收集类似例句，微调SpaCy的词性标注模型：

• 准备带正确标注的训练数据：

TRAIN_DATA = [
    ("Buy groceries.", {"tags": ["VB", "NNS", "."]}),
    ("Cook dinner.", {"tags": ["VB", "NN", "."]}),
    ("Clean the room.", {"tags": ["VB", "DT", "NN", "."]}),
    # 添加更多祈使句或歧义场景示例
]

• 按照SpaCy的训练流程加载模型、配置训练参数、运行训练循环，更新模型的标注逻辑。

内容的提问来源于stack exchange，提问作者Kim Stacks