Skip to content

Commit eace364

Browse files
committed
revise duplication
1 parent c140e8d commit eace364

48 files changed

Lines changed: 516 additions & 487 deletions

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

content/post/9/1.jpg

2.24 MB
Loading

content/post/9/1.png

-449 KB
Binary file not shown.

content/post/9/2.jpg

2.24 MB
Loading

content/post/9/2.png

-388 KB
Binary file not shown.

content/post/9/3.jpg

3.57 MB
Loading

content/post/9/3.png

-1.01 MB
Binary file not shown.

content/post/9/featured.jpg

3.57 MB
Loading

content/post/9/featured.png

-1.01 MB
Binary file not shown.

content/post/9/index.md

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -12,17 +12,17 @@ date: 2025-08-04
1212
- 论文简介:
1313
  检索增强型生成(RAG)系统在处理知识密集型任务中展现出巨大潜力,然而,当检索到的上下文与模型的参数化知识发生冲突时,生成结果的不一致性问题成为了一个亟待解决的重大挑战。如图1所示,知识冲突导致模型的回答无法忠实于检索上下文从而导致准确率急剧下降。
1414
<figure>
15-
<img src="1.png" alt="" style="width: 50%;">
15+
<img src="1.jpg" alt="" style="width: 50%;">
1616
<figcaption style="text-align: center;">图1. 知识冲突导致模型回答不忠实于检索上下文</figcaption>
1717
</figure>
1818
&emsp;&emsp;目前的忠实性提升方法主要通过严格限制模型对上下文的依赖来解决这一问题,但这些方法往往会抑制模型的参数化知识,导致模型内部知识结构受损,增加了对上下文的误读风险。如图2所示,目前导致模型无法忠实于上下文产生的错误类型有两类:过度自信和错误匹配。过度自信指的是模型过度相信自身的参数化知识,错误匹配则指的是模型误读上下文学习到了错误的上下文知识。
1919
<figure>
20-
<img src="2.png" alt="" style="width: 50%;">
20+
<img src="2.jpg" alt="" style="width: 50%;">
2121
<figcaption style="text-align: center;">图2. 两种知识冲突场景下模型的错误输出范式</figcaption>
2222
</figure>
2323
&emsp;&emsp;针对这一难题,本文提出了 FaithfulRAG 框架,实现了在知识冲突场景下对上下文忠实性的显著提升。其框架如图3所示,首先,FaithfulRAG 将“事实级冲突建模”引入检索增强生成(RAG)流程,解决了传统方法在面对模型参数知识与检索内容不一致时产生的幻觉问题。具体来说,FaithfulRAG 通过 Self-Fact Mining模块将模型内部的参数知识显式化为可解释的自我事实(self-facts),从而精确识别与上下文之间的冲突点。其次,通过Contextual Knowledge Alignment将检索上下文按事实粒度对齐,过滤噪声并保留关键冲突片段。最后,FaithfulRAG设计了Self-Think 模块,引导模型在生成前主动推理并融合冲突信息,避免盲目依赖参数记忆或上下文。在 FaithEval、MuSiQue、SQuAD 和 RealtimeQA等多个数据集基准实验上表明,FaithfulRAG在保持高准确率的同时显著降低了“过度自信”与“错误匹配”两类错误,实现了鲁棒且可信的检索增强生成。
2424
<figure>
25-
<img src="3.png" alt="" style="width: 66.6%;">
25+
<img src="3.jpg" alt="" style="width: 66.6%;">
2626
<figcaption style="text-align: center;">图3. FaithfulRAG流程图</figcaption>
2727
</figure>
2828
&emsp;&emsp;表1展示了FaithfulRAG在知识冲突场景下的实验结果,在实体级冲突数据集(MuSiQue、SQuAD)和逻辑级冲突数据集(FaithEval)以及实时问答(RealtimeQA)四种基准数据集上,FaithfulRAG均取得了显著优于现有最佳基线的忠实且准确的生成表现。

content/publication/126/index.md

Lines changed: 10 additions & 12 deletions
Original file line numberDiff line numberDiff line change
@@ -1,23 +1,21 @@
11
---
2-
title: "Signer Diversity-driven Data Augmentation for Signer-Independent Sign Language Translation"
2+
title: "GENET: Unleashing the Power of Side Information for Recommendation via Hypergraph Pre-training"
33
authors:
4-
- Honghao Fu
5-
- Liang Zhang
6-
- Biao Fu
7-
- Rui Zhao
4+
- Yang Li
5+
- Qiao Zhao
6+
- Chen Lin
7+
- Zhenjie Zhang
8+
- Xiaomin Zhu
89
- Jinsong Su
9-
- Xiaodong Shi
10-
- Yidong Chen
1110
author_notes:
1211
-
1312
-
1413
-
1514
-
1615
-
1716
-
18-
-
19-
date: "2024-05-29T08:08:29Z"
20-
publishDate: "2025-05-29T08:08:29Z"
21-
publication_types: [direction2]
22-
publication: "**In Proc. of NAACL 2024.** (CCF-B类)"
17+
date: "2024-05-29T08:08:37Z"
18+
publishDate: "2025-05-29T08:08:37Z"
19+
publication_types: []
20+
publication: "**In Proc. of DASFAA 2024, Short Paper.**"
2321
---

0 commit comments

Comments
 (0)