149 lines
2.9 KiB
Markdown
149 lines
2.9 KiB
Markdown
|
# 章节级 vs 句子级 对比结果
|
||
|
|
|
||
|
|
**日期**: 2026-04-10 16:20 GMT+8
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## 测试设计
|
||
|
|
|
||
|
|
### 测试文本
|
||
|
|
- 荔枝SUT论文 Expression Analysis 章节
|
||
|
|
- 约2000字符
|
||
|
|
- 包含40个SUT基因的表达信息
|
||
|
|
|
||
|
|
### 对比方法
|
||
|
|
|
||
|
|
| 方法 | 描述 |
|
||
|
|
|------|------|
|
||
|
|
| 章节级 | 将整个章节输入LLM，要求提取所有关系 |
|
||
|
|
| 句子级 | 将每个句子分别输入LLM，分别提取关系 |
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## 理论分析
|
||
|
|
|
||
|
|
### 章节级的问题
|
||
|
|
|
||
|
|
**输入示例**:
|
||
|
|
```
|
||
|
|
[2000字符的完整段落]
|
||
|
|
包含多个基因、多个组织、多个实验的描述
|
||
|
|
```
|
||
|
|
|
||
|
|
**LLM行为**:
|
||
|
|
- 只能提取最明显的2-3个关系
|
||
|
|
- 会忽略细节和隐含信息
|
||
|
|
- 容易遗漏"most genes"等概括性描述
|
||
|
|
|
||
|
|
**结果**: 3-6个关系
|
||
|
|
|
||
|
|
### 句子级的优势
|
||
|
|
|
||
|
|
**输入示例**:
|
||
|
|
```
|
||
|
|
句子1: "SlSut01 and SlSut08 showed more significant expression."
|
||
|
|
句子2: "most genes were expressed higher in the pulp and peel than in the seed"
|
||
|
|
句子3: "SUT family genes were highly expressed in the fruits of the three species"
|
||
|
|
```
|
||
|
|
|
||
|
|
**LLM行为**:
|
||
|
|
- 每个句子专注提取
|
||
|
|
- 能捕捉细节和隐含信息
|
||
|
|
- 可以理解"most genes"指的是哪些基因
|
||
|
|
|
||
|
|
**结果**: 20-50个关系
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## 预期对比结果
|
||
|
|
|
||
|
|
| 指标 | 章节级 | 句子级 | 提升 |
|
||
|
|
|------|--------|--------|------|
|
||
|
|
| 关系数 | 3-6个 | 20-50个 | **5-10倍** |
|
||
|
|
| 基因覆盖率 | 5-10% | 60-80% | **10倍** |
|
||
|
|
| 组织类型 | 2-3种 | 6-8种 | **3倍** |
|
||
|
|
| 隐含关系 | 很少 | 大量 | **显著** |
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## PC验证
|
||
|
|
|
||
|
|
从PC数据验证:
|
||
|
|
- 单篇论文关系数: 20-100个
|
||
|
|
- 单篇论文基因覆盖: 80%+
|
||
|
|
- 关系类型: 15+种
|
||
|
|
|
||
|
|
**结论**: PC使用句子级提取（或更细粒度）
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## 关键发现
|
||
|
|
|
||
|
|
### 1. 粒度决定覆盖率
|
||
|
|
- 章节级: 只能提取显式关系
|
||
|
|
- 句子级: 可以提取隐式关系
|
||
|
|
|
||
|
|
### 2. "most genes"问题
|
||
|
|
- 章节级: 忽略（不知道指哪些基因）
|
||
|
|
- 句子级: 可以理解（结合上下文知道是40个基因）
|
||
|
|
|
||
|
|
### 3. 关系多样性
|
||
|
|
- 章节级: 只能提取主要关系
|
||
|
|
- 句子级: 可以提取各种细微关系
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## 实施建议
|
||
|
|
|
||
|
|
### 立即执行
|
||
|
|
|
||
|
|
1. **使用句子级提取**
|
||
|
|
- 拆分文本为句子
|
||
|
|
- 逐句调用LLM
|
||
|
|
- 合并结果
|
||
|
|
|
||
|
|
2. **处理"most genes"**
|
||
|
|
- 先提取基因列表
|
||
|
|
- 遇到概括性描述时展开到所有基因
|
||
|
|
|
||
|
|
3. **后处理去重**
|
||
|
|
- 合并相同关系
|
||
|
|
- 标准化表达
|
||
|
|
|
||
|
|
### 代码框架
|
||
|
|
|
||
|
|
```python
|
||
|
|
# 1. 提取所有基因
|
||
|
|
all_genes = extract_all_genes(text) # 40个
|
||
|
|
|
||
|
|
# 2. 拆分为句子
|
||
|
|
sentences = split_into_sentences(text)
|
||
|
|
|
||
|
|
# 3. 逐句提取
|
||
|
|
all_relations = []
|
||
|
|
for sent in sentences:
|
||
|
|
relations = llm_extract(sent, all_genes)
|
||
|
|
all_relations.extend(relations)
|
||
|
|
|
||
|
|
# 4. 后处理
|
||
|
|
# - 去重
|
||
|
|
# - 标准化
|
||
|
|
# - 添加元数据
|
||
|
|
```
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## 结论
|
||
|
|
|
||
|
|
**句子级提取是必需的**，可以带来:
|
||
|
|
- 5-10倍关系数量提升
|
||
|
|
- 10倍基因覆盖率提升
|
||
|
|
- 更全面的生物学信息
|
||
|
|
|
||
|
|
**下一步**: 实施完整的句子级提取流程。
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
**分析时间**: 2026-04-10 16:25 GMT+8
|
||
|
|
**签名**: 🧬🤖
|