Displacy无法识别自定义NER实体的技术求助

算力阿峰

2026-7-16

自定义NER实体Displacy不高亮的解决方案

核心问题与修复步骤

1. 管道加载时机错误

你先调用nlp(cleaned_text)处理文本，之后才添加entity_ruler管道。spaCy的管道只会作用于添加管道之后处理的文本，所以之前生成的doc不会包含自定义实体。

修复： 先添加entity_ruler管道，再处理文本。

2. Entity Ruler模式不匹配测试文本

你的pattern设置不符合实际文本结构：

BUN的pattern要求前面有一个Xxxxx形状的词（首字母大写、后面小写），但测试文本里"BUN"是独立出现的，没有前置词；
SerumUricAcid的pattern多了一个Xxxxx形状的词，实际文本是"Serum Uric Acid"三个词，不需要前置额外词汇。

修复： 调整pattern为匹配实际文本的结构：

ner_patterns = [
    {"label": "BUN", "pattern": [{"LOWER": "bun"}]},
    {"label": "SerumUricAcid", "pattern": [{"LOWER": "serum"}, {"LOWER": "uric"}, {"LOWER": "acid"}]},
    {"label": "SerumPotassium", "pattern": [{"LOWER": "serum"}, {"LOWER": "potassium"}]}
]

3. Displacy配置语法错误

options的ents列表中，"BUN"后面的引号未闭合，导致语法错误，Displacy无法识别自定义实体标签。

修复： 修正引号：

options = {"ents": ["BUN", "SerumUricAcid", "SerumPotassium"], "colors": colors}

修正后的完整代码

import spacy
from spacy import displacy
from IPython.display import HTML

# Load the spaCy English model
nlp = spacy.load("en_core_web_sm")

# Add the custom NER patterns to the entity ruler FIRST
ner_patterns = [
    {"label": "BUN", "pattern": [{"LOWER": "bun"}]},
    {"label": "SerumUricAcid", "pattern": [{"LOWER": "serum"}, {"LOWER": "uric"}, {"LOWER": "acid"}]},
    {"label": "SerumPotassium", "pattern": [{"LOWER": "serum"}, {"LOWER": "potassium"}]}
]

ruler = nlp.add_pipe("entity_ruler", after='ner')
ruler.add_patterns(ner_patterns)

# 测试文本（替换为你的cleaned_text）
cleaned_text = """BIOCHEMISTRY KIDNEY FUNCTION TEST (KFT) TEST VALUE UNIT REFERENCE BUN 10.27 mg/dl 7.9 - 20 Serum Urea 22 mg/dl 13 - 40 Serum Creatinine H 0.9 mg/dl 0.5 - 0.8 Serum Calcium 9.0 mg/dl 8.8 - 10.6 Serum Potassium 3.9 mmol/L 3.5 - 5.1 Serum Sodium L 132 mmol/L 136 - 146 Serum Uric Acid 5 mg/dl 2.6 - 6 Urea / Creatinine Ratio 24.44 BUN / Creatinine Ratio 11.41 ~~~ End of report ~~~ LABSMART SAMPLE REPORT Patient Name: Mrs. Dummy Registered on: 09/08/2022 11:35 AM 1001 Age / Sex: 34 YRS / F Collected on: 09/08/2022 Referred By: Dr. Self Received on: 12/08/2022 Reg. no. / UHID: 1001 / Reported on: 09/08/2022 11:35 AM Investigations: Kidney Function Test (KFT) Page 1 of 1 Mr. Sachin Sharma DMLT, Lab Incharge Dr. A. K. Asthana MBBS, MD Pathologist"""

# Process the cleaned text AFTER adding the pipeline
doc = nlp(cleaned_text)

# Extracted values
extracted_values = {}

# Process the tokens to extract values
for token in doc:
    # 调整匹配逻辑，对应自定义标签
    if token.ent_type_ in ["BUN", "SerumUricAcid", "SerumPotassium"]:
        # 找到实体后的数值
        next_token = token.nbor()
        while next_token and not next_token.like_num:
            next_token = next_token.nbor()
        if next_token and next_token.like_num:
            extracted_values[token.ent_type_] = next_token.text

# Print the extracted values
print("Extracted values:")
for label, value in extracted_values.items():
    print(f"{label}: {value}")

# Prepare entity highlighting with linear gradients
colors = {
    "BUN": "linear-gradient(90deg, #aa9cfc, #fc9ce7)",
    "SerumUricAcid": "linear-gradient(90deg, #ffabab, #ffd8a8)",
    "SerumPotassium": "linear-gradient(90deg, #a8e6cf, #dcedc1)"
}
options = {"ents": ["BUN", "SerumUricAcid", "SerumPotassium"], "colors": colors}

# Render the displaCy visualization with highlighting
html = displacy.render(doc, style="ent", options=options)

# Display the HTML
HTML(html)

额外优化点

提取数值的逻辑改为基于实体类型判断，更贴合自定义NER的结果，避免匹配无关的"Serum"等单独词汇；
添加了循环跳过非数值词，确保找到实体对应的第一个数值。

内容的提问来源于stack exchange，提问作者Ujjwal