Datasets:
Update README.md
Browse files
README.md
CHANGED
|
@@ -14,3 +14,30 @@ configs:
|
|
| 14 |
- split: test
|
| 15 |
path: "disputable.csv"
|
| 16 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
- split: test
|
| 15 |
path: "disputable.csv"
|
| 16 |
---
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
# DYMANICQA
|
| 20 |
+
|
| 21 |
+
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.
|
| 22 |
+
|
| 23 |
+
<p align="center">
|
| 24 |
+
<img src="main_figure.png" width="650" alt="main_figure">
|
| 25 |
+
</p>
|
| 26 |
+
|
| 27 |
+
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** 🕰️.
|
| 28 |
+
|
| 29 |
+
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!
|
| 30 |
+
|
| 31 |
+
The implementation of the measures is available on our github [repo](https://github.com/copenlu/dynamicqa)!
|
| 32 |
+
|
| 33 |
+
## Dataset
|
| 34 |
+
|
| 35 |
+
Our dataset consists of three different partitions.
|
| 36 |
+
|
| 37 |
+
| Partition | Number of Questions |
|
| 38 |
+
| --------- | ------------------- |
|
| 39 |
+
| Static | 2500 |
|
| 40 |
+
| Temporal | 2495 |
|
| 41 |
+
| Disputable | 694 |
|
| 42 |
+
|
| 43 |
+
|