Datasets:

mfromm
/

AMSR

We release a new dataset of peer-reviews from different computer science conferences with annotated arguments, called AMSR (**A**rgument **M**ining in **S**cientific **R**eviews).

1. Raw Data
conferences_raw/ contains directories for each conference we scraped (e.g., [iclr20](./data/iclr20)).
The respective directory of each conference comprises multiple `*.json` files, where every file contains the information belonging to a single paper, such as the title, the abstract, the submission date and the reviews.
The reviews are stored in a list called `"review_content"`.

2. Cleaned Data
conferences_cleaned/ contains reviews and papers where we removed all unwated character sequences from the reviews.
For details on the details of the preprocessing steps, please refer to our paper "Argument Mining Driven Analysis of Peer-Reviews".

3. Annotated Data
conferences_annotated/ contains sentence_level and token_level data of 77 reviews, annotated each by 3 annotators.
We have three labels:
PRO - Arguments supporting the acceptance of the paper.
CON - Arguments opposing the acceptance of the paper.
NON - Non-argumentative sentences/tokens which have no influence on the acceptance of the paper.

And following we have three tasks:

Argumentation Detection:
A binary classification of whether a text span is an argument.
The classes are denoted by ARG and NON, where ARG is the union of PRO and CON classes.

Stance Detection:
A binary classification whether an argumentative text span is supporting or opposing the paper acceptance.
he model is trained and evaluated only on argumentative PRO and CON text spans.

Joint Detection:
A multi-class classification between the classes PRO, CON and NON, i.e. the combination of argumentation and stance detection.

4. Generalization across Conferences
conferences_annotated_generalization/ contains token_level data of 77 reviews split diffrently than in 3.

We studied the model’s generalization to peer-reviews for papers from other (sub)domains.
To this end, wereduce the test set to only contain reviews from the GI’20conference.
The focus of the GI’20 conference is ComputerGraphics and Human-Computer Interaction, while the otherconferences are focused on Representation Learning, AI andMedical Imaging.
We consider the GI’20 as a subdomain since all conferences are from the domain of computer science.
NO-GI:
The original training dataset with all sentences from reviews of GI’20 removed.
ALL
A resampling of the original training dataset of the same size as NO-GI, with sentences from all conferences.

5. jupyter-Notebook
ReviewStat is a jupyter notebook, which shows interesting statistics of the raw dataset.