Multimodal Entity Linking (MEL) consists in combining information from several modalities (textual, visual...) to map an ambiguous mention to an entity in a knowledge base (KB). We propose a MEL dataset based on Twitter posts, and elaborate a process for collecting and constructing a fully annotated MEL dataset, where entities are defined in a Twitter KB. This repository contains:
- the data and their annotation that correspond to the corpus used in our LREC 2020 article [1].
- the programs that allows to build your own MEL dataset from Twitter, with the process proposed at ECIR 2020 [2]
conda create -n mael python=2.7
conda activate mael
conda install -c conda-forge tweepy python-wget
conda install -c anaconda nltk
conda install numpy # for program splitGroudTruths.py only
# with python 2.7 only
conda install configparser
The corpus is available at this address under the licence CC BY-NC-SA 3.0.
We provide the identifiers of the tweets that were used in [1,2]. Due to the Twitter policy we do not release the full (hydratated) content : see here at Redistribution of Twitter content. We also provide a program to retrieve a Tweet content from its ID and convert it to the appropriate format. All the material is in the folder corpus:
mel_train_ids35,976 ids of the training evaluation corpusmel_dev_ids16,599 ids of the dev evaluation corpusmel_test_ids36,521 ids of the test evaluation corpuskb2,657,213 ids of the knowledge database
Send us a mail if you have problem.
You can create a corpus similar to that we used by using the program we provide. For this, you need Twitter API credentials. Then use the programs and seed files in code/corpus. The steps are:
- collect data and store them into a (sqlite3) database
- convert data into files to create (i) the knowledge database (ii) the evaluation corpus
A seed file is a text file that contain all Twitter screen names (@xxxx, one per line) to query. The corresponding accounts are collected as is:
- all the tweets (up to the imit of the Twitter API) of the timeline to create the knowledge database
- the recent tweets as possible samples to be included in the evaluation corpus
To collect data and store it in the database my_knowledge.db let report your Twitter API credentials into twitterDataDB.py (line 24-27), then:
python twitterDataDB.py -db_name my_knowledge.db -seed_file seed_files/xxx.txt -query e
python twitterDataDB.py -db_name my_knowledge.db -seed_file seed_files/yyy.txt -query m
If you use -query e it collects tweet from entities to create the knowledge database. With -query m it collects tweets with potential mentions to create the evaluation corpus.
Once data are stored in my_knowledge.db you can create the corresponding file with sqlite3 my_knowledge.db then:
sqlite> .output timelineKB.txt
sqlite> .separator "\t" "\n"
sqlite> select ('@'||userScreenName),tweetId,replace(replace(replace(tweetFullText,CHAR(10),' '),CHAR(13),' '),CHAR(9),' '),mediaURL from timeLineTweets where mediaURL!='' order by userScreenName;
The knowledge dataset is then the text file timelineKB.txt. It has one line per sample and the columns (tab separator) are:
- original screen names
- tweet identifiers
- textual content
- visual content (image URL)
You can download the images using their URL (fourth column)
You need to identify ambiguous users such that the corpus is more challenging. To seek ambiguous last names use:
python get_ambiguous_users.py -db_name ambiguousUsers.db \
-lastName_seed seed_files/popLastNames.txt \
-screenName_seed seed_files/HouseRepublicans.txt \
-seed_type ln
use -seed_type sn to seek ambiguous screen names (less useful).
A version of such a database of ambiguous users is available here.
Then let generate a map between screen names and mentions, using sqlite3 ambiguousUsers.db:
sqlite> .output mapSreenNameToMention.txt
sqlite> select ('@'||userScreenName),userSearchQueryLasttName from twitterUsers;
Using mapSreenNameToMention.txt, you can now generate tweets with ambiguous mentions from table searchTweets in my_knowledge.db:
python generate_groundTruthTweets.py -db_name my_knowledge.db -o groundTruth.txt
You can finally create the corresponding {Train/Dev/Test}.txt files with:
python splitGroudTruths.py -i groundTruth.txt
Since the full process is quite long, we provide a small seed file to test the programs.
python twitterDataDB.py -db_name test0.db -seed_file quick_test/AAA.txt -query e
python twitterDataDB.py -db_name test0.db -seed_file quick_test/AAA.txt -query m
sqlite3 test0.db < quick_test/create_timeline.sql
wget --no-check-certificate 'https://docs.google.com/uc?export=download&id=1qYGTUJlkzyFbSTeGP-g1PVefkq8VgUwA' -O ambiguousUsers.db.gz
gunzip ambiguousUsers.db.gz
sqlite3 ambiguousUsers.db < quick_test/create_mapSreenNameToMention.sql
python generate_groundTruthTweets.py -db_name test0.db -o groundTruth.txt
As of May 2020, groundTruth.txt has only 4 tweets for this quick test, splitted into 2 tweets for Train.txt and one in Dev.txt and Test.txt.
TODO
If you find this material useful for your research, please cite
[1] O. Adjali, R. Besançon, O. Ferret, H. Le Borgne and B. Grau (2020) Building a Multimodal Entity Linking Dataset From Tweets, 12th International Conference on Language Resources and Evaluation (LREC)
[2] O. Adjali, R. Besançon, O. Ferret, H. Le Borgne and B. Grau (2020) Multimodal Entity Linking for Tweets, 42nd European Conference on Information Retrieval (ECIR): Advances in Information Retrieval
Bibtex entries:
@inproceedings{adjali2020ecir,
title={Multimodal Entity Linking for Tweets},
author={Adjali, Omar and Besancon, romaric and Ferret, olivier and {Le Borgne}, Herv{\'e} and Grau, Brigitte},
booktitle={European Conference on Information Retrieval (ECIR)},
year={2020},
month={april},
day={14--17},
address={Lisbon, Portugal}
}
@inproceedings{adjali2020lrec,
title={Building a Multimodal Entity Linking Dataset From Tweets},
author={Adjali, Omar and Besancon, romaric and Ferret, olivier and {Le Borgne}, Herv{\'e} and Grau, Brigitte},
booktitle={International Conference on Language Resources and Evaluation (LREC)},
year={2020},
month={may},
day={11--16},
address={Marseille, France}
}