Table of Contents
This is the source code for experiments of DQNC2S: DQN-based Cross-stream Crisis event Summarizer paper accepted at ECIR 2024.
The repository is in the creation phase; some files could be missing
The employed dataset is CrisisFACTS 2022.
To download the dataset, refer to https://github.com/crisisfacts/utilities/tree/main and the competition website.
The module environment contains a Gymnasium Env. It could be used with different frameworks.
The suggested environment is SimilarityCrisisEnv, which includes the BERT embedding, max similarity, and remaining space. The following are its attributes:
Action Space: Discrete(2)
Observation Space: Box(770, N)
The requests_file ipc should have the following schema:
-
('eventID', String)
-
('requestID', String)
-
('dateString', String)
Each file with annotations contained in data_folder should have the following schema:
-
('text', String), the passage
-
('source_type', String)
-
('unix_timestamp', Int64)
-
('answer', String), the answer to the query given text
-
('score', Float64),, confidence for the answer
-
('query', String), the query
To better understand the meaning, refer to CrisisFacts 2022 data.
The original setup of this repository is by Daniele Rege Cambrin.
For a full list of all authors and contributors, see the contributors page.
This project is licensed under the Apache 2.0 license. See LICENSE for more information.
The project makes use of CleanRL under the MIT license. See CLEANRL_LICENSE for more information.