<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[Daisymanman的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/Daisymanman</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; Daisymanman]]></copyright><item><title><![CDATA[论文阅读：Augmenting Reasoning Capabilities of LLMs with Graph Structures in KBQA]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/146096992</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/146096992</guid><author>Daisymanman</author><pubDate>Sat, 08 Mar 2025 11:22:45 +0800</pubDate><description><![CDATA[论文来源：EMNLP 2024 finding（暂未发布）]]></description><category></category></item><item><title><![CDATA[论文阅读：KAM-CoT: Knowledge Augmented Multimodal Chain-of-Thoughts Reasoning]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/146053482</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/146053482</guid><author>Daisymanman</author><pubDate>Thu, 06 Mar 2025 18:35:56 +0800</pubDate><description><![CDATA[LLM通过利用能够逐步思考的思维链在NLP任务中取得了很好的性能，但是为LLM扩展多模态能力时计算成本高，且需要大量的硬件资源。为了解决上述问题，本文提出了KAM-CoT框架，集成了CoT推理、知识图谱和多种模态，以全面理解多模态任务。KAM-CoT采用两阶段式训练过程，连接知识图谱以生成有效的推理和答案，通过在推理过程中整合来自KG的外部知识，使模型获得更很层次的上下文理解，以缓解幻觉，并提高答案的质量。这种知识增强的CoT推理能够使模型处理需要外部上下文的问题，提供更高质量的答案。]]></description><category></category></item><item><title><![CDATA[论文阅读：Chain-of-Question: A Progressive Question Decomposition Approach for Complex KBQA]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/145678715</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/145678715</guid><author>Daisymanman</author><pubDate>Mon, 17 Feb 2025 17:53:14 +0800</pubDate><description><![CDATA[论文来源：ACL 2024 Finding论文地址：KBQA任务通常包含两种范式用于解决复杂的KBQA：基于信息检索的方法和基于语义解析的方法。基于信息检索的方法：构建一个特定于问题的子图来捕获相关信息，并根据问题的相关性对实体进行排序；基于语义解析的方法将问题转换为符号逻辑形式，并查询知识库以获得答案。基于语义解析的KBQA方法通常是根据原始问题推导出所有的子问题，并生成逻辑形式推理出答案，但是这类方法并不适用于需要根据前一问题的答案生成下一问题的情况，如多跳推理。]]></description><category></category></item><item><title><![CDATA[论文笔记：RelationPrompt :Zero-Shot Relation Triplet Extraction]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/127061593</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/127061593</guid><author>Daisymanman</author><pubDate>Sat, 12 Oct 2024 16:48:03 +0800</pubDate><description><![CDATA[提示学习关系三元组抽取]]></description><category></category></item><item><title><![CDATA[论文笔记：Pre-training to Match for Unified Low-shot Relation Extraction]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/127259176</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/127259176</guid><author>Daisymanman</author><pubDate>Sat, 12 Oct 2024 16:47:32 +0800</pubDate><description><![CDATA[MCMN]]></description><category></category></item><item><title><![CDATA[论文笔记：Prototypical Verbalizer for Prompt-based Few-shot Tuning]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/127659534</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/127659534</guid><author>Daisymanman</author><pubDate>Sat, 12 Oct 2024 16:47:19 +0800</pubDate><description><![CDATA[提示学习、原型向量、对比学习]]></description><category></category></item><item><title><![CDATA[Knowledgeable Prompt-tuning: Incorporating Knowledge into Prompt Verbalizer for Text Classification]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/127754816</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/127754816</guid><author>Daisymanman</author><pubDate>Sat, 12 Oct 2024 16:47:00 +0800</pubDate><description><![CDATA[提示学习，语言表达器]]></description><category></category></item><item><title><![CDATA[论文笔记：KnowPrompt: Knowledge-aware Prompt-tuning withSynergistic Optimization for Relation Extraction]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/127813414</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/127813414</guid><author>Daisymanman</author><pubDate>Sat, 12 Oct 2024 16:46:38 +0800</pubDate><description><![CDATA[提示学习 RE]]></description><category></category></item><item><title><![CDATA[论文笔记：Ontology-enhanced Prompt-tuning for Few-shot Learning]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/127929425</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/127929425</guid><author>Daisymanman</author><pubDate>Sat, 12 Oct 2024 16:46:24 +0800</pubDate><description><![CDATA[提示学习]]></description><category></category></item><item><title><![CDATA[论文笔记：PTR: Prompt Tuning with Rules for Text Classification]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/128002136</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/128002136</guid><author>Daisymanman</author><pubDate>Sat, 12 Oct 2024 16:46:06 +0800</pubDate><description><![CDATA[提示学习]]></description><category></category></item><item><title><![CDATA[论文笔记：Template-Based Named Entity Recognition Using BART]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/128267261</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/128267261</guid><author>Daisymanman</author><pubDate>Sat, 12 Oct 2024 16:45:49 +0800</pubDate><description><![CDATA[提示学习]]></description><category></category></item><item><title><![CDATA[论文笔记：Label Verbalization and Entailment for Effective Zero and Few-Shot Relation Extraction]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/128472147</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/128472147</guid><author>Daisymanman</author><pubDate>Sat, 12 Oct 2024 16:45:33 +0800</pubDate><description><![CDATA[提示学习，关系抽取]]></description><category></category></item><item><title><![CDATA[论文笔记：Prompt-Based Meta-Learning For Few-shot Text Classification]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/130017103</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/130017103</guid><author>Daisymanman</author><pubDate>Sat, 12 Oct 2024 16:45:12 +0800</pubDate><description><![CDATA[Prompt-Based Meta-Learning]]></description><category></category></item><item><title><![CDATA[peft安装]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/137569828</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/137569828</guid><author>Daisymanman</author><pubDate>Tue, 09 Apr 2024 22:11:16 +0800</pubDate><description><![CDATA[【代码】peft安装。]]></description><category></category></item><item><title><![CDATA[论文笔记：IMF: Interactive Multimodal Fusion Model for Link Prediction]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/136952093</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/136952093</guid><author>Daisymanman</author><pubDate>Mon, 25 Mar 2024 21:17:42 +0800</pubDate><description><![CDATA[最近一些研究将多模态信息引入到链接预测中，但是这些方法是分别利用多模态信息的，忽略了不同模态之间的复杂交互。为了更好地建模模态间的信息，引入一种交互式多模态融合模型，整合不同模态间的知识。通过两阶段式多模态融合框架来保存特定模态的知识，并利用不同模态之间的交互。多模态融合模块没有将不同模态投影到一个统一的向量空间，而是限制了不同模态的独立表示，同时利用双线性池化进行融合，并将对比学习作为额外的约束。此外，决策融合模块在所有模态预测的上提供学习到的加权平均，以更好地整合不同模态的互补性。]]></description><category></category></item><item><title><![CDATA[论文笔记：Think-on-graph: Deep and responsible reasoning of large language model with knowledge graph]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/136525822</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/136525822</guid><author>Daisymanman</author><pubDate>Thu, 07 Mar 2024 19:21:04 +0800</pubDate><description><![CDATA[尽管大语言模型在各种任务中取得了显著的效果，但是经常存在幻觉问题。通过在LLM推理中引入外部知识图谱可以解决部分问题。本文提出一种新的LLM-KG集成范式，将LLM视为一个代理，交互探索KG上的相关实体和关系，并基于检索到的知识进行推理。作者引入一种名为在图上思考（Think-on-Graph, ToG）的方法来进一步实现这一范式，LLM代理在KG上迭代执行波束搜索，发现最优路径，并返回最有可能的推理结果。ToG方法的优点：① 与LLM相比，ToG具有更好的深度推理能力；]]></description><category></category></item><item><title><![CDATA[论文笔记：Knowledge-Augmented Language Model Promptingfor Zero-Shot Knowledge Graph Question Answering]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/136445095</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/136445095</guid><author>Daisymanman</author><pubDate>Mon, 04 Mar 2024 15:51:18 +0800</pubDate><description><![CDATA[LLM能够根据其在预训练过程中存储在参数中的内部知识进行零样本闭卷问答任务，但是这种内部知识可能是不正确的和过时的，导致LLM生成事实上错误的答案，此外，对LLM进行微调来更新其中的知识是昂贵的。为此，本文直接在LLM的输入中增加知识。具体来说，首先根据问题及相关事实之间的语义相似性从知识图谱中检索与输入问题相关的事实；然后，将检索到的事实以提示的形式与输入问题融合，并输入到LLM中，生成答案。KAPING框架不需要重新进行模型训练，因此可以适用于零样本场景。]]></description><category></category></item><item><title><![CDATA[论文笔记：Do Prompt-Based Models Really Understandthe the Meaning of Their Prompts?]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/129776413</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/129776413</guid><author>Daisymanman</author><pubDate>Sun, 26 Mar 2023 17:33:03 +0800</pubDate><description><![CDATA[提示真的有意义吗]]></description><category></category></item><item><title><![CDATA[论文笔记：Event Detection as Graph Parsing]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/121898614</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/121898614</guid><author>Daisymanman</author><pubDate>Mon, 13 Dec 2021 11:22:48 +0800</pubDate><description><![CDATA[论文地址：https://aclanthology.org/2021.findings-acl.142.pdf

论文来源：ACL 2021

Abstract

事件检测是信息抽取的一项基本任务。之前大多数方法通常将事件检测视为基于触发词的分类问题，聚焦于使用语法依赖结构或外部知识来提高分类性能。为了克服现有基于触发词分类 模型存在的问题，本文提出一种新的事件检测方法，将其描述为一个图解析问题，可以显式地对多个事件关联进行建模，并自然地利用事件类型和子类型传递的丰富信息。此外，为了解决数据稀疏问题，在不.]]></description><category></category></item><item><title><![CDATA[论文笔记-MLBiNet: A Cross-Sentence Collective Event Detection Network]]></title><link>https://blog.youkuaiyun.com/Daisymanman/article/details/120469060</link><guid>https://blog.youkuaiyun.com/Daisymanman/article/details/120469060</guid><author>Daisymanman</author><pubDate>Tue, 28 Sep 2021 09:44:32 +0800</pubDate><description><![CDATA[论文地址：https://arxiv.org/pdf/2105.09458.pdf

代码地址：

来源：ACL 2021


]]></description><category></category></item></channel></rss>