Phase 05: NLP: Foundations to Advanced

核心提议

"她打电话给他,他没有回答.医生在午餐上".三次提到两个人,没有人被命名.

Type: Learn

Languages: Python

Prerequisites: Phase 5 · 06 (NER), Phase 5 · 07 (POS & Parsing)

Time: ~60 minutes

问题

简单地说"果".很难说"公司","他们","库珀蒂诺的科技巨头",或"就业公司".没有把这些提及解决给同一实体,你的NER管道会错过60-80%.

核心解析度将指同一实体的每个表达式连接到一个集群中. 它是表面级NLP (NER,解析) 和下游语义 (IE,QA,总结,KG) 之间的粘合物.

2026年为什么这很重要:

总结:"首席执行官宣布... "vs"蒂姆·库克宣布... " 总结应该命名首席执行官.
回答"她叫谁?"的问题需要解决"她".
信息提取:一个知识图表,其中"PER1创立了果"和"Jobs创立了果"作为单独的条目是错误的.
多文档IE:将有关同一事件的文章中的提到合并是跨文档核心参考.

概念

!Coreference clustering: mentions → entities

The task.输入:一个文件.输出:一个指名 (范围) 的集群,其中每个集群指的是一个实体.

Mention types.

Named entity."蒂姆·库克"
Nominal."CEO", "公司"
Pronominal."他", "她", "他们", "它"
Appositive."果公司首席执行官蒂姆·库克,

Architectures.

Rule-based (Hobbs, 1978).语法规则的语法定量,基本线程很好,在代名词上难以击败.
Mention-pair classifier.对于每一对提及 $inline$ 根据过渡关闭, 标准在2016年之前.
Mention-ranking.对于每一个提及,排名候选人的前例 (包括"没有前例").
Span-based end-to-end (Lee et al., 2017).变压器编码器. 列出所有候选人跨度到长度限制. 预测提及分数. 预测每个跨度的前例概率. 贪地集群. 现代默认.
Generative (2024+).简单的案例,长文档和罕见的参考文件.

The evaluation metrics.报告前三项的平均值为CoNLL F1.2026年最新情况:CONLL-2012: ~83F1.

Known hard cases.

关于之前引入的页面的实体的确定的描述.
桥梁"轮子" →之前提到的汽车.
在中国和日本等语言中,无法.
形 (引用者之前的代名词):she走进来,玛丽笑了.

建立它

步骤1:预训练的神经核心 (AllenNLP / spaCy-实验)

python

import spacy
nlp = spacy.load("en_coreference_web_trf")   # experimental model
doc = nlp("Apple announced new products. The company said it would ship them soon.")
for cluster in doc._.coref_clusters:
    print(cluster, "->", [m.text for m in cluster])

在更长的文件上,你得到了这样的东西:

集团1: [果,公司,它]
集团2:[新产品,它们]

步骤2:基于规则的代名词解决器 (教学)

看到code/main.py仅用于执行:

提取提及:命名实体 (大写范围),代名词 (直观搜索),明确描述 ("X").
对于每个代词,看看前面的K提及,并以:

- 性别/数量协议 (理性)

- 收获 (收获)

- 语法作用 (优先专题)

连接最高分的前例.

但它显示了搜索空间和一个端到端模型必须做出的决定.

步骤3:使用LLM为核心参考

python

prompt = f"""Text: {text}

List every pronoun and noun phrase that refers to a person or company.
Cluster them by what they refer to. Output JSON:
[{{"entity": "Apple", "mentions": ["Apple", "the company", "it"]}}, ...]
"""

两个失败模式可以观看.第一,LLM过度合并 ("他"和"她"指两个不同的人).第二,LLM默默地放弃长文档中的提及.总是通过跨度抵消检查验证.

步骤4:评估

标准 conll-2012 脚本计算MUC,B3,CEAF-φ4并报告平均值.为了进行内部评估,首先使用跨度级别精度,然后回忆出注释测试集,然后添加引用链接F1.

陷

Singleton explosion.某些系统将每一个提到的数据都视为自己的集群.B3是宽松的.MUC惩罚了这件事.
Pronouns in long context.在2000个代币以上的文件上,性能下降了15 F1.
Gender assumptions.硬编码的性别规则违反了非二进制指标,组织,动物.
LLM drift on long docs.单个API调用不能可靠地集群在50+段落中提到.使用滑动窗口+合并.

用它

现在,我们要做什么?

Situation	Pick
English, single document	`en_coreference_web_trf` (spaCy-experimental) or AllenNLP neural coref
Multilingual	SpanBERT / XLM-R trained on OntoNotes or Multilingual CoNLL
Cross-document event coref	Specialized end-to-end models (2025–26 SOTA)
Quick LLM baseline	GPT-4o / Claude with structured-output coref prompt
Production dialog systems	Rule-based fallback + neural primary + manual review for critical slots

2026年将出现的集成模式:首先运行NER,运行coref,将coref集群合并到NER实体. 下游任务每集群都会看到一个实体,而不是每一个实体.

运送它

保存如outputs/skill-coref-picker.md其他:

markdown

---
name: coref-picker
description: Pick a coreference approach, evaluation plan, and integration strategy.
version: 1.0.0
phase: 5
lesson: 24
tags: [nlp, coref, information-extraction]
---

Given a use case (single-doc / multi-doc, domain, language), output:

1. Approach. Rule-based / neural span-based / LLM-prompted / hybrid. One-sentence reason.
2. Model. Named checkpoint if neural.
3. Integration. Order of operations: tokenize → NER → coref → downstream task.
4. Evaluation. CoNLL F1 (MUC + B³ + CEAF-φ4 average) on held-out set + manual cluster review on 20 documents.

Refuse LLM-only coref for documents over 2,000 tokens without sliding-window merge. Refuse any pipeline that runs coref without a mention-level precision-recall report. Flag gender-heuristic systems deployed in demographically diverse text.

运动

Easy.运行基于规则的解决器code/main.py根据5个手工制成的段落, 根据基础的真相衡量引用链接的准确性.
Medium.根据你自己的手动注释,它失败了哪里?
Hard.建立一个核心增强的NER管道:首先是NER,然后是通过核心集群合并.

关键词

Term	What people say	What it actually means
Mention	A reference	A span of text that refers to an entity (name, pronoun, noun phrase).
Antecedent	What "it" refers to	The earlier mention a later one corefers with.
Cluster	The entity's mentions	Set of mentions that all refer to the same real-world entity.
Anaphora	Backward reference	Later mention refers to earlier ("he" → "John").
Cataphora	Forward reference	Earlier mention refers to later ("When he arrived, John...").
Bridging	Implicit reference	"I bought a car. The wheels were bad." (wheels of THAT car.)
CoNLL F1	The number on leaderboards	Average of MUC, B³, CEAF-φ4 F1 scores.

进一步阅读

Jurafsky & Martin, SLP3 Ch. 26 — Coreference Resolution and Entity Linking经典教科书章节.
Lee et al. (2017). End-to-end Neural Coreference Resolution基于跨度的端到端.
Joshi et al. (2020). SpanBERT预训练,以改善核心.
Pradhan et al. (2012). CoNLL-2012 Shared Task基准指数
Hobbs (1978). Resolving Pronoun References基于规则的经典.

This free lesson is part of the AI Engineering from Scratch curriculum. Read the full explanation, run the lesson code, and verify the result in the interactive reader or from the repository source.

Browse the complete course catalog or open this lesson on GitHub.