# The CoreSearch Dataset
A cleaner version of the CoreSearch dataset</br>

## Cleaning done in this version
Below examples of fixes done in this dataset version

### Fix event mention spans 
| Before Span Fix   | After Span Fix |
| :------ | :----- |
| Academy Awards 65th ceremony    |  Academy Awards   |
| 65th Annual Academy Awards    |  Annual Academy Awards   |
| 2015 Southland Conference Women's Basketball Tournament    |  Southland Conference Women's Basketball Tournament   |
|   2010 Austin suicide attack  |  attack   |
|  an earthquake in 1948   |   earthquake  |
|   tragic plane crash  |   crash  |
|   Lokomotiv Yaroslavl plane tragedy  |  tragedy   |
|   just been shot  |  shot   |
|   wounded by gunfire  |  wounded   |
|  Kennedy's assassination in Los Angeles   |  assassination   |
|  Arab League summit in Cairo   |  summit   |
|  plane crash on 4 July 1943   | crash |

### Resolve cases of two events in a single span
| Before Span Fix   | After Span Fix |
| :------ | :----- |
| killing or abduction of several hundred villagers |  killing |
| kidnapping and murder of Aldo Moro | kidnapping |


### Remove erroneous event mentions (Entities)
| Mention Removed   | Reason (not event) |
| :------ | :----- |
| I Want Your Love | Song name  |
| 2006 Screen Actors Guild Award for Best Ensemble | Award name |
| in 1896 | Year |


- **Paper:** [Cross-document Event Coreference Search: Task, Dataset and Modeling](https://arxiv.org/abs/2210.12654)

### Languages

English

## Load Dataset
You can read/download the dataset files following Huggingface Hub instructions.</br>
For example, below code will load CoreSearch DPR folder:

```python
from huggingface_hub import hf_hub_url, cached_download
import json
REPO_ID = "datasets/Intel/CoreSearchV2"
DPR_FILES = "/dpr/"

dpr_files = ["dpr/Dev.json", "dpr/Train.json", "dpr/Test.json"]

dpr_jsons = list()
for _file in dpr_files:
    dpr_jsons.append(json.load(open(cached_download(
        hf_hub_url(REPO_ID, _file)), "r")))
```

### Data Splits
- **Final version of the CD event coreference search dataset**<br>
|                                                    | Train   | Valid | Test | Total |
| -----                                              | ------ | ----- | ----  | ----  |
| WEC-Eng Validated Data                             |        |       |       |       |
| &nbsp;&nbsp;&nbsp;&nbsp;# Clusters                 | 229    |  48   | 226   | 503   | 
| &nbsp;&nbsp;&nbsp;&nbsp;# Passages (with Mentions) | 1,429  |  335  | 1,206 | 2,970 |
| # Added Destructor Passages                        | 922,736  |  923,376  | 923,746   | 2,769,858 |
| # Total Passages                                   | 924,239 |  923,717 | 925,012 | 2,772,968 |

## Citation
```
@inproceedings{eirew-etal-2022-cross,
    title = "Cross-document Event Coreference Search: Task, Dataset and Modeling",
    author = "Eirew, Alon  and
      Caciularu, Avi  and
      Dagan, Ido",
    booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.emnlp-main.58",
    pages = "900--913",
    abstract = "The task of Cross-document Coreference Resolution has been traditionally formulated as requiring to identify all coreference links across a given set of documents. We propose an appealing, and often more applicable, complementary set up for the task {--} Cross-document Coreference Search, focusing in this paper on event coreference. Concretely, given a mention in context of an event of interest, considered as a query, the task is to find all coreferring mentions for the query event in a large document collection. To support research on this task, we create a corresponding dataset, which is derived from Wikipedia while leveraging annotations in the available Wikipedia Event Coreferecene dataset (WEC-Eng). Observing that the coreference search setup is largely analogous to the setting of Open Domain Question Answering, we adapt the prominent Deep Passage Retrieval (DPR) model to our setting, as an appealing baseline. Finally, we present a novel model that integrates a powerful coreference scoring scheme into the DPR architecture, yielding improved performance.",
}
```


## License
We provide the following data sets under a <a href="https://creativecommons.org/licenses/by-sa/3.0/deed.en_US">Creative Commons Attribution-ShareAlike 3.0 Unported License</a>. It is based on content extracted from Wikipedia that is licensed under the Creative Commons Attribution-ShareAlike 3.0 Unported License

## Contact
If you have any questions please create a Github issue at <a href="https://github.com/AlonEirew/CoreSearch">https://github.com/AlonEirew/CoreSearch</a>.