docs: revise ML method summaries

Author: beautifulrem

PPT_Overleaf/ml_methods_by_module.md

@@ -134,8 +134,6 @@ Training scripts produce the shipped `models/*.joblib` artifacts.
 | `training.train_ranker` | Standardized handcrafted ranking features + linear logistic SGD | `models/ranker.joblib` |
 | `training.train_typo_linker` | Dict features + linear logistic SGD | `models/typo_linker.joblib` |
 
-## 8. Short Presentation Summary
+## 8. Summary
 
-If asked "What ML methods does the project use?", the short answer is:
-
-> The core system uses classical, explainable ML: sparse n-gram features with linear logistic SGD classifiers for text classification, CRF for entity boundary detection, handcrafted feature-based linear classifiers for clarification, typo linking, source planning, and ranking, plus TF-IDF/cosine methods for retrieval and deduplication. Transformer models are only downstream exceptions: FinBERT for document sentiment and LLMs for final answer and next-question JSON generation over already retrieved evidence.
+The core system uses classical, explainable ML: sparse n-gram features with linear logistic SGD classifiers for text classification, CRF for entity boundary detection, handcrafted feature-based linear classifiers for clarification, typo linking, source planning, and ranking, plus TF-IDF/cosine methods for retrieval and deduplication. Transformer models are downstream exceptions: FinBERT handles document sentiment, and LLMs generate final answer and next-question JSON over already retrieved evidence.

PPT_Overleaf/ml_methods_by_module_cn.md

@@ -134,8 +134,6 @@ Retrieval 中的辅助非 ML 方法：
 | `training.train_ranker` | 标准化人工排序特征 + 线性 logistic SGD | `models/ranker.joblib` |
 | `training.train_typo_linker` | 字典特征 + 线性 logistic SGD | `models/typo_linker.joblib` |
 
-## 8. 答辩简短总结
+## 8. 总结
 
-如果被问到“这个项目用了哪些 ML 方法”，可以这样回答：
-
-> 项目的核心系统使用传统、可解释机器学习：文本分类主要是稀疏 n-gram 特征加线性 logistic SGD 分类器，实体边界识别使用 CRF，澄清判断、错别字链接、source planning 和排序使用人工特征加线性分类器，文档检索和去重使用 TF-IDF / cosine similarity。Transformer 模型只作为下游例外：FinBERT 用于文档情感分析，LLM 用于在已有证据基础上生成最终回答和追问 JSON。
+项目的核心系统使用传统、可解释机器学习：文本分类主要是稀疏 n-gram 特征加线性 logistic SGD 分类器，实体边界识别使用 CRF，澄清判断、错别字链接、source planning 和排序使用人工特征加线性分类器，文档检索和去重使用 TF-IDF / cosine similarity。Transformer 模型是下游例外：FinBERT 用于文档情感分析，LLM 用于在已有证据基础上生成最终回答和追问 JSON。