Datasets:

Modalities:
Tabular
Text
Formats:
csv
Languages:
English
ArXiv:
Libraries:
Datasets
pandas
License:
File size: 2,462 Bytes
e44e0ac
 
 
 
0351a07
 
1fbc30c
e44e0ac
0351a07
 
1fbc30c
e44e0ac
0351a07
 
1fbc30c
 
 
 
 
 
 
 
e44e0ac
49badc8
 
1fbc30c
49badc8
 
 
 
9353f66
49badc8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d0a663b
 
 
 
 
 
 
 
 
 
7c51a08
 
 
 
 
 
82cb85d
7c51a08
 
 
82cb85d
 
7c51a08
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
---
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}
}
```