Skip to content

Commit

Permalink
Site updated: 2023-10-18 14:30:33
Browse files Browse the repository at this point in the history
  • Loading branch information
@hedarko20
hedarko20 committed Oct 18, 2023
1 parent 38b85f0 commit 64a1076
Show file tree
Hide file tree
Showing 27 changed files with 303 additions and 302 deletions.
4 changes: 2 additions & 2 deletions 2023/10/09/paper1/index.html
View file
Original file line number Diff line number Diff line change
Expand Up @@ -174,12 +174,12 @@ <h1>MUSER: A MUlti-Step Evidence Retrieval Enhancement Framework for Fake News D

<span class="tag">

<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #00bcd4">论文</a>
<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #03a9f4">论文</a>
</span>

<span class="tag">

<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #ffa2c4">阅读</a>
<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #ff7d73">阅读</a>
</span>

<span class="tag">
Expand Down
6 changes: 3 additions & 3 deletions 2023/10/10/paper2/index.html
View file
Original file line number Diff line number Diff line change
Expand Up @@ -174,17 +174,17 @@ <h1>Interpretable Multimodal Misinformation Detection with Logic Reasoning 阅

<span class="tag">

<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #ffa2c4">论文</a>
<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #ff7d73">论文</a>
</span>

<span class="tag">

<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #03a9f4">阅读</a>
<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #00bcd4">阅读</a>
</span>

<span class="tag">

<a href="/tags/%E5%81%87%E6%96%B0%E9%97%BB%E6%A3%80%E6%B5%8B/" style="color: #ffa2c4">假新闻检测</a>
<a href="/tags/%E5%81%87%E6%96%B0%E9%97%BB%E6%A3%80%E6%B5%8B/" style="color: #ff7d73">假新闻检测</a>
</span>

</span>
Expand Down
4 changes: 2 additions & 2 deletions 2023/10/10/paper3/index.html
View file
Original file line number Diff line number Diff line change
Expand Up @@ -174,12 +174,12 @@ <h1>Improving Evidence Retrieval for Automated Explainable Fact-Checking 阅读

<span class="tag">

<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #ff7d73">论文</a>
<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #03a9f4">论文</a>
</span>

<span class="tag">

<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #00bcd4">阅读</a>
<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #00a596">阅读</a>
</span>

</span>
Expand Down
4 changes: 2 additions & 2 deletions 2023/10/11/paper4/index.html
View file
Original file line number Diff line number Diff line change
Expand Up @@ -174,12 +174,12 @@ <h1>Reasoning Over Semantic-Level Graph for Fact Checking 阅读笔记</h1>

<span class="tag">

<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #ff7d73">论文</a>
<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #ffa2c4">论文</a>
</span>

<span class="tag">

<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #00bcd4">阅读</a>
<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #00a596">阅读</a>
</span>

</span>
Expand Down
4 changes: 2 additions & 2 deletions 2023/10/11/paper5/index.html
View file
Original file line number Diff line number Diff line change
Expand Up @@ -174,12 +174,12 @@ <h1>Learn over Past, Evolve for Future: Forecasting Temporal Trends for Fake New

<span class="tag">

<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #00a596">论文</a>
<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #ffa2c4">论文</a>
</span>

<span class="tag">

<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #ffa2c4">阅读</a>
<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #00a596">阅读</a>
</span>

</span>
Expand Down
2 changes: 1 addition & 1 deletion 2023/10/15/tutorial1/index.html
View file
Original file line number Diff line number Diff line change
Expand Up @@ -179,7 +179,7 @@ <h1>embedding pytorch 实现</h1>

<span class="tag">

<a href="/tags/deeping-learning/" style="color: #03a9f4">deeping learning</a>
<a href="/tags/deeping-learning/" style="color: #00bcd4">deeping learning</a>
</span>

</span>
Expand Down
6 changes: 3 additions & 3 deletions 2023/10/16/paper7/index.html
View file
Original file line number Diff line number Diff line change
Expand Up @@ -174,17 +174,17 @@ <h1>Automated Fact Checking: Task formulations, methods and future directions

<span class="tag">

<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #03a9f4">论文</a>
<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #ff7d73">论文</a>
</span>

<span class="tag">

<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #00bcd4">阅读</a>
<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #00a596">阅读</a>
</span>

<span class="tag">

<a href="/tags/%E4%BA%8B%E5%AE%9E%E9%AA%8C%E8%AF%81/" style="color: #00a596">事实验证</a>
<a href="/tags/%E4%BA%8B%E5%AE%9E%E9%AA%8C%E8%AF%81/" style="color: #03a9f4">事实验证</a>
</span>

</span>
Expand Down
11 changes: 6 additions & 5 deletions 2023/10/17/paper8/index.html
View file
Original file line number Diff line number Diff line change
Expand Up @@ -174,17 +174,17 @@ <h1>Unsupervised Dense Information Retrieval with Contrastive Learning 阅读笔

<span class="tag">

<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #03a9f4">论文</a>
<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #00bcd4">论文</a>
</span>

<span class="tag">

<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #00bcd4">阅读</a>
<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #00a596">阅读</a>
</span>

<span class="tag">

<a href="/tags/%E4%BF%A1%E6%81%AF%E6%A3%80%E7%B4%A2/" style="color: #00a596">信息检索</a>
<a href="/tags/%E4%BF%A1%E6%81%AF%E6%A3%80%E7%B4%A2/" style="color: #ffa2c4">信息检索</a>
</span>

</span>
Expand All @@ -197,8 +197,9 @@ <h1 id="作者和出处"><a href="#作者和出处" class="headerlink" title="
<span id="more"></span>

<h1 id="摘要"><a href="#摘要" class="headerlink" title="摘要"></a>摘要</h1><p>近年来，信息检索领域出现了基于神经网络的密集检索器，作为基于词频的经典稀疏方法的替代。这些模型在有大训练数据的数据集和任务上取得了最先进的结果。然而，它们在新应用中没有训练数据时迁移效果较差，并且在诸如 BM25 等无监督词频方法的性能面前落败。在这项工作中，我们探讨了对比学习作为一种训练无监督密集检索器的方法，并表明它在各种检索设置中导致了强大的性能。在我们的 BEIR 基准测试中，我们的无监督模型在 15 个数据集中有 11 个在 Recall@100 上优于 BM25。当使用我们的对比模型作为微调前的预训练，无论是使用领域内几千个示例，还是使用大型 MS MARCO 数据集，都可以在 BEIR 基准测试中取得改进。最后，我们评估了我们的方法在多语言检索中的应用，在这种情况下，训练数据甚至比英语更少，并表明我们的方法导致了强大的无监督性能。我们的模型在仅使用监督英语数据进行微调并在斯瓦希里语等低资源语言上进行评估时，也表现出强大的跨语言转移。</p>
<h1 id="引言和结论"><a href="#引言和结论" class="headerlink" title="引言和结论"></a>引言和结论</h1><p>在本文中，我们做出了以下贡献。首先，我们证明了对比学习可以导致强大的无监督检索器：我们的模型在 BEIR 基准测试的 most 数据集上取得了与 BM25 竞争的 Recall@100 结果。其次，在少量示例设置中，我们证明了我们的模型可以从少量训练示例中受益，并获得了比从诸如 MS MARCO 之类的大型数据集中转移的模型更好的结果。第三，当我们使用 MS MARCO 进行微调之前，我们的技术在 BEIR 基准测试中导致了强大的性能。我们进行了一些 ablation 以激发我们的设计选择，并表明裁剪比反向 Cloze 任务效果更好。最后，我们使用对比学习训练了一个多语言密集检索器，并表明它取得了最先进的性能。</p>
<h1 id="方法"><a href="#方法" class="headerlink" title="方法"></a>方法</h1><h1 id="个人感悟"><a href="#个人感悟" class="headerlink" title="个人感悟"></a>个人感悟</h1>
<h1 id="引言和结论"><a href="#引言和结论" class="headerlink" title="引言和结论"></a>引言和结论</h1><p>文档检索是一个在大量集合中寻找相关文档来回答特定查询的任务。<br>传统的检索系统依赖查询和文档之间的近乎精确的匹配，导致其泛化效果不佳。<br>神经网络具有强大的检索能力，但是创建数据集需要手动匹配。在大型数据集上几乎不可能做到。所以，基于神经网络的稠密检索器的效果往往不如基于词频的做法，因为基于词频的做法是无监督的。此外，大型数据集通常不包含除英语以外的其他版本，因此，使用大量的监督数据集不适合训练多语言的检索系统。<br>在本文中，我们做出了以下贡献。首先，我们证明了对比学习可以导致强大的无监督检索器：我们的模型在 BEIR 基准测试的 most 数据集上取得了与 BM25 竞争的 Recall@100 结果。其次，在少量示例（few-shot）设置中，我们证明了我们的模型可以从少量训练示例中受益，并获得了比从诸如 MS MARCO 之类的大型数据集中转移的模型更好的结果。第三，当我们使用 MS MARCO 进行微调之前，我们的技术在 BEIR 基准测试中导致了强大的性能。我们进行了一些 ablation 以激发我们的设计选择，并表明裁剪比反向 Cloze 任务效果更好。最后，我们使用对比学习训练了一个多语言密集检索器，并表明它取得了最先进的性能。</p>
<h1 id="方法"><a href="#方法" class="headerlink" title="方法"></a>方法</h1><p>用一个Transformer架构的编码器同时对query和doc分别进行编码，最后一层取平均，然后点积求相似性。<br>$$s(q,d)&#x3D;\langle f_\theta(q),f_\theta(d)\rangle $$<br>如果对query和doc分别设计不同的编码器进行编码，可能会降低零样本和少样本情况下的鲁棒性。</p>
<h2 id="在未标注的文档上进行无监督学习"><a href="#在未标注的文档上进行无监督学习" class="headerlink" title="在未标注的文档上进行无监督学习"></a>在未标注的文档上进行无监督学习</h2><h1 id="个人感悟"><a href="#个人感悟" class="headerlink" title="个人感悟"></a>个人感悟</h1>
</div>


Expand Down
4 changes: 2 additions & 2 deletions 2023/10/17/tutorial2/index.html
View file
Original file line number Diff line number Diff line change
Expand Up @@ -174,12 +174,12 @@ <h1>记录下Python中有关path的bug及解决方案</h1>

<span class="tag">

<a href="/tags/debug/" style="color: #ff7d73">debug</a>
<a href="/tags/debug/" style="color: #00bcd4">debug</a>
</span>

<span class="tag">

<a href="/tags/python/" style="color: #00bcd4">python</a>
<a href="/tags/python/" style="color: #ff7d73">python</a>
</span>

</span>
Expand Down
38 changes: 19 additions & 19 deletions archives/2023/10/index.html
View file
Original file line number Diff line number Diff line change
Expand Up @@ -190,7 +190,7 @@ <h3>Hexo教程</h3>

<span class="tag">

<a href="/tags/Hexo/" style="color: #00bcd4">Hexo</a>
<a href="/tags/Hexo/" style="color: #ff7d73">Hexo</a>
</span>

</span>
Expand All @@ -217,17 +217,17 @@ <h3>MUSER: A MUlti-Step Evidence Retrieval Enhancement Framework for Fake News D

<span class="tag">

<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #ff7d73">论文</a>
<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #00bcd4">论文</a>
</span>

<span class="tag">

<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #ffa2c4">阅读</a>
<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #03a9f4">阅读</a>
</span>

<span class="tag">

<a href="/tags/%E5%81%87%E6%96%B0%E9%97%BB%E6%A3%80%E6%B5%8B/" style="color: #00a596">假新闻检测</a>
<a href="/tags/%E5%81%87%E6%96%B0%E9%97%BB%E6%A3%80%E6%B5%8B/" style="color: #ffa2c4">假新闻检测</a>
</span>

</span>
Expand All @@ -254,17 +254,17 @@ <h3>Interpretable Multimodal Misinformation Detection with Logic Reasoning 阅

<span class="tag">

<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #ffa2c4">论文</a>
<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #ff7d73">论文</a>
</span>

<span class="tag">

<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #00bcd4">阅读</a>
<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #03a9f4">阅读</a>
</span>

<span class="tag">

<a href="/tags/%E5%81%87%E6%96%B0%E9%97%BB%E6%A3%80%E6%B5%8B/" style="color: #03a9f4">假新闻检测</a>
<a href="/tags/%E5%81%87%E6%96%B0%E9%97%BB%E6%A3%80%E6%B5%8B/" style="color: #00a596">假新闻检测</a>
</span>

</span>
Expand Down Expand Up @@ -296,7 +296,7 @@ <h3>Improving Evidence Retrieval for Automated Explainable Fact-Checking 阅读

<span class="tag">

<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #00bcd4">阅读</a>
<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #03a9f4">阅读</a>
</span>

</span>
Expand All @@ -323,7 +323,7 @@ <h3>Reasoning Over Semantic-Level Graph for Fact Checking 阅读笔记</h3>

<span class="tag">

<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #ffa2c4">论文</a>
<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #00bcd4">论文</a>
</span>

<span class="tag">
Expand Down Expand Up @@ -360,7 +360,7 @@ <h3>Learn over Past, Evolve for Future: Forecasting Temporal Trends for Fake New

<span class="tag">

<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #ffa2c4">阅读</a>
<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #00a596">阅读</a>
</span>

</span>
Expand All @@ -387,7 +387,7 @@ <h3>A Survey on Automated Fact-Checking 阅读报告</h3>

<span class="tag">

<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #03a9f4">论文</a>
<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #ff7d73">论文</a>
</span>

<span class="tag">
Expand Down Expand Up @@ -419,12 +419,12 @@ <h3>embedding pytorch 实现</h3>

<span class="tag">

<a href="/tags/code/" style="color: #03a9f4">code</a>
<a href="/tags/code/" style="color: #00a596">code</a>
</span>

<span class="tag">

<a href="/tags/deeping-learning/" style="color: #00bcd4">deeping learning</a>
<a href="/tags/deeping-learning/" style="color: #ff7d73">deeping learning</a>
</span>

</span>
Expand All @@ -451,12 +451,12 @@ <h3>Automated Fact Checking: Task formulations, methods and future directions

<span class="tag">

<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #ffa2c4">论文</a>
<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #00bcd4">论文</a>
</span>

<span class="tag">

<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #ff7d73">阅读</a>
<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #03a9f4">阅读</a>
</span>

<span class="tag">
Expand Down Expand Up @@ -488,17 +488,17 @@ <h3>Unsupervised Dense Information Retrieval with Contrastive Learning 阅读笔

<span class="tag">

<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #ffa2c4">论文</a>
<a href="/tags/%E8%AE%BA%E6%96%87/" style="color: #03a9f4">论文</a>
</span>

<span class="tag">

<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #00bcd4">阅读</a>
<a href="/tags/%E9%98%85%E8%AF%BB/" style="color: #00a596">阅读</a>
</span>

<span class="tag">

<a href="/tags/%E4%BF%A1%E6%81%AF%E6%A3%80%E7%B4%A2/" style="color: #03a9f4">信息检索</a>
<a href="/tags/%E4%BF%A1%E6%81%AF%E6%A3%80%E7%B4%A2/" style="color: #ff7d73">信息检索</a>
</span>

</span>
Expand All @@ -525,7 +525,7 @@ <h3>记录下Python中有关path的bug及解决方案</h3>

<span class="tag">

<a href="/tags/debug/" style="color: #ffa2c4">debug</a>
<a href="/tags/debug/" style="color: #ff7d73">debug</a>
</span>

<span class="tag">
Expand Down
Loading

0 comments on commit 64a1076

Please sign in to comment.