如何在spaCy中提取动词短语 解决ROOT子树提取失效问题

原有基于依存子树截取的方法仅适用于主语、宾语、状语从句这类本身对应独立完整依存子树的成分。主句核心动词的依存标签为ROOT，它的子树天然覆盖整句所有内容，直接截取会把主语、从句、宾语等无关成分全部包含，必须通过依存关系做过滤，才能精准提取动词短语。

实现逻辑 #

• 遍历句法分析结果，定位依存标签为ROOT的主句核心谓词
• 以核心谓词为中心，仅收集动词短语内部构成成分：助动词（含被动语态助动词）、修饰动词的副词、否定词，直接排除主语、宾语、状语从句、介词短语等非VP附属成分
• 将收集到的所有成分按原文出现顺序排序后拼接，得到最终动词短语

代码实现 #

新增动词短语提取函数get_vp，完整可运行代码如下：

import spacy

def get_subj(decomp):
    for token in decomp:
        if ("subj" in token.dep_):
            subtree = list(token.subtree)
            start = subtree[0].i
            end = subtree[-1].i + 1
            return str(decomp[start:end])

def get_obj(decomp):
    for token in decomp:
        if ("dobj" in token.dep_ or "pobj" in token.dep_):
            subtree = list(token.subtree)
            start = subtree[0].i
            end = subtree[-1].i + 1
            return str(decomp[start:end])

def get_advcl(decomp):
    for token in decomp:
        if ("advcl" in token.dep_):
            subtree = list(token.subtree)
            start = subtree[0].i
            end = subtree[-1].i + 1
            return str(decomp[start:end])

def get_vp(decomp):
    # 定位主句ROOT核心谓词
    root = None
    for token in decomp:
        if token.dep_ == "ROOT":
            root = token
            break
    if not root:
        return ""
    vp_tokens = {root}
    # 定义动词短语内部允许的依存关系
    allowed_deps = {"aux", "aux:pass", "advmod", "neg"}
    for child in root.children:
        if child.dep_ in allowed_deps:
            vp_tokens.add(child)
            # 递归收集修饰助词/副词的附属成分，适配very/extremely等程度副词修饰场景
            for sub_child in child.children:
                if sub_child.dep_ in allowed_deps:
                    vp_tokens.add(sub_child)
    # 按原文顺序排序后截取
    vp_tokens = sorted(vp_tokens, key=lambda x: x.i)
    start = vp_tokens[0].i
    end = vp_tokens[-1].i + 1
    return str(decomp[start:end])

phrase = "Ultimate Swirly Ice Cream Scoopers are usually overrated when one considers all of the scoopers one could buy."

nlp = spacy.load("en_core_web_sm")
decomp = nlp(phrase)

subj = get_subj(decomp)
obj = get_obj(decomp)
advcl = get_advcl(decomp)
vp = get_vp(decomp)

print("subj: ", subj)
print("obj: ", obj)
print("advcl: ", advcl)
print("vp: ", vp)

运行结果 #

subj:  Ultimate Swirly Ice Cream Scoopers
obj:  all of the scoopers
advcl:  when one considers all of the scoopers one could buy
vp:  are usually overrated

扩展适配 #

如果需要覆盖更复杂的动词短语场景，可在allowed_deps集合中补充对应依存标签：

• 短语动词小品词（如pick up中的up）添加prt标签
• 并列动词结构添加cc、conj标签
• 情态动词（can/should等）本身属于aux标签，无需额外配置

内容的提问来源于stack exchange，提问作者Chris