copenlu/dynamicqa
收藏资源简介:
--- license: mit configs: - config_name: static data_files: - split: test path: static.csv - config_name: temporal data_files: - split: test path: temporal.csv - config_name: disputable data_files: - split: test path: disputable.csv task_categories: - question-answering language: - en pretty_name: DynamicQA size_categories: - 10K<n<100K --- # DYNAMICQA This is a repository for the paper [DYNAMICQA: Tracing Internal Knowledge Conflicts in Language Models](https://arxiv.org/abs/2407.17023) accepted at Findings of EMNLP 2024. <p align="center"> <img src="main_figure.png" width="800" alt="main_figure"> </p> Our paper investigates the Language Model's behaviour when the conflicting knowledge exist within the LM's parameters. We present a novel dataset containing inherently conflicting data, DYNAMICQA. Our dataset consists of three partitions, **Static**, **Disputable** 🤷♀️, and **Temporal** 🕰️. We also evaluate several measures on their ability to reflect the presence of intra-memory conflict: **Semantic Entropy** and a novel **Coherent Persuasion Score**. You can find our findings from our paper! The implementation of the measures is available on our github [repo](https://github.com/copenlu/dynamicqa)! ## Dataset Our dataset consists of three different partitions. | Partition | Number of Questions | | --------- | ------------------- | | Static | 2500 | | Temporal | 2495 | | Disputable | 694 | ### Details 1. Question : "question" column 2. Answers : Two different answers are available: one in the "obj" column and the other in the "replace_name" column. 3. Context : Context ("context" column) is masked with \[ENTITY\]. Before providing the context to the LM, you should replace \[ENTITY\] with either "obj" or "replace_name". 4. Number of edits : "num_edits" column. This denotes Temporality for temporal partition, and Disputability for disputable partition. ## Citation If you find our dataset helpful, kindly refer to us in your work using the following citation: ``` @inproceedings{marjanović2024dynamicqatracinginternalknowledge, title={DYNAMICQA: Tracing Internal Knowledge Conflicts in Language Models}, author={Sara Vera Marjanović and Haeun Yu and Pepa Atanasova and Maria Maistro and Christina Lioma and Isabelle Augenstein}, year={2024}, booktitle = {Findings of EMNLP}, publisher = {Association for Computational Linguistics} } ```
许可证:MIT协议 配置项: - 配置名称:静态(static) 数据文件: - 拆分集:测试集 文件路径:static.csv - 配置名称:时序(temporal) 数据文件: - 拆分集:测试集 文件路径:temporal.csv - 配置名称:有争议(disputable) 数据文件: - 拆分集:测试集 文件路径:disputable.csv 任务类别:问答(question-answering) 语言:英语 规范名称:DynamicQA 样本规模区间:10000 < 样本量 < 100000 # DYNAMICQA 本仓库对应发表于2024年EMNLP Findings的论文《DYNAMICQA:追踪大语言模型(Large Language Model, LLM)内部知识冲突》(https://arxiv.org/abs/2407.17023)。 <p align="center"> <img src="main_figure.png" width="800" alt="主示意图"> </p> 本论文研究了当大语言模型参数内部存在冲突知识时的模型行为。我们构建了一个包含固有冲突数据的全新数据集DYNAMICQA,该数据集包含三个分区:**静态(Static)**、**有争议(Disputable)** 🤷♀️以及**时序(Temporal)** 🕰️。 我们还评估了多种指标在反映模型内部记忆冲突存在性方面的能力,包括**语义熵(Semantic Entropy)**以及全新提出的**一致性说服得分(Coherent Persuasion Score)**。相关研究结论可参见论文。 上述指标的实现代码可在我们的GitHub仓库(https://github.com/copenlu/dynamicqa)中获取! ## 数据集详情 本数据集包含三个不同分区: | 分区名称 | 问题数量 | | --------- | ------------------- | | 静态分区(Static) | 2500 | | 时序分区(Temporal) | 2495 | | 有争议分区(Disputable) | 694 | ### 数据集说明 1. 问题:对应数据集中的`question`列 2. 答案:共包含两种不同答案,分别位于`obj`列与`replace_name`列中 3. 上下文:上下文内容(对应`context`列)已使用`[ENTITY]`进行掩码。在将上下文输入大语言模型之前,需将`[ENTITY]`替换为`obj`或`replace_name`中的对应值 4. 编辑次数:对应`num_edits`列。对于时序分区,该字段表示时序性程度;对于有争议分区,该字段表示争议性程度 ## 引用声明 若您的工作中使用了本数据集,请通过以下引用格式致谢: @inproceedings{marjanović2024dynamicqatracinginternalknowledge, title={DYNAMICQA: Tracing Internal Knowledge Conflicts in Language Models}, author={Sara Vera Marjanović and Haeun Yu and Pepa Atanasova and Maria Maistro and Christina Lioma and Isabelle Augenstein}, year={2024}, booktitle = {Findings of EMNLP}, publisher = {Association for Computational Linguistics} }




