Dataset Card for "cardiffnlp/tweet_topic_single"

Dataset Summary

This is the official repository of TweetTopic ("Twitter Topic Classification , COLING main conference 2022"), a topic classification dataset on Twitter with 6 labels. Each instance of TweetTopic comes with a timestamp which distributes from September 2019 to August 2021. See cardiffnlp/tweet_topic_multi for multi label version of TweetTopic. The tweet collection used in TweetTopic is same as what used in TweetNER7. The dataset is integrated in TweetNLP too.

Preprocessing

We pre-process tweets before the annotation to normalize some artifacts, converting URLs into a special token {{URL}} and non-verified usernames into {{USERNAME}}. For verified usernames, we replace its display name (or account name) with symbols {@}. For example, a tweet

Get the all-analog Classic Vinyl Edition
of "Takin' Off" Album from @herbiehancock
via @bluenoterecords link below: 
http://bluenote.lnk.to/AlbumOfTheWeek

is transformed into the following text.

Get the all-analog Classic Vinyl Edition
of "Takin' Off" Album from {@herbiehancock@}
via {@bluenoterecords@} link below: {{URL}}

A simple function to format tweet follows below.

import re
from urlextract import URLExtract
extractor = URLExtract()
def format_tweet(tweet):
    # mask web urls
    urls = extractor.find_urls(tweet)
    for url in urls:
        tweet = tweet.replace(url, "{{URL}}")
    # format twitter account
    tweet = re.sub(r"\b(\s*)(@[\S]+)\b", r'\1{\2@}', tweet)
    return tweet
target = """Get the all-analog Classic Vinyl Edition of "Takin' Off" Album from @herbiehancock via @bluenoterecords link below: http://bluenote.lnk.to/AlbumOfTheWeek"""
target_format = format_tweet(target)
print(target_format)
'Get the all-analog Classic Vinyl Edition of "Takin\' Off" Album from {@herbiehancock@} via {@bluenoterecords@} link below: {{URL}}'

Data Splits

split	number of texts	description
test_2020	376	test dataset from September 2019 to August 2020
test_2021	1693	test dataset from September 2020 to August 2021
train_2020	2858	training dataset from September 2019 to August 2020
train_2021	1516	training dataset from September 2020 to August 2021
train_all	4374	combined training dataset of train_2020 and train_2021
validation_2020	352	validation dataset from September 2019 to August 2020
validation_2021	189	validation dataset from September 2020 to August 2021
train_random	2830	randomly sampled training dataset with the same size as train_2020 from train_all
validation_random	354	randomly sampled training dataset with the same size as validation_2020 from validation_all
test_coling2022_random	3399	random split used in the COLING 2022 paper
train_coling2022_random	3598	random split used in the COLING 2022 paper
test_coling2022	3399	temporal split used in the COLING 2022 paper
train_coling2022	3598	temporal split used in the COLING 2022 paper

For the temporal-shift setting, model should be trained on train_2020 with validation_2020 and evaluate on test_2021. In general, model would be trained on train_all, the most representative training set with validation_2021 and evaluate on test_2021.

IMPORTANT NOTE: To get a result that is comparable with the results of the COLING 2022 Tweet Topic paper, please use train_coling2022 and test_coling2022 for temporal-shift, and train_coling2022_random and test_coling2022_random fir random split (the coling2022 split does not have validation set).

Models

model	training data	F1	F1 (macro)	Accuracy
cardiffnlp/roberta-large-tweet-topic-single-all	all (2020 + 2021)	0.896043	0.800061	0.896043
cardiffnlp/roberta-base-tweet-topic-single-all	all (2020 + 2021)	0.887773	0.79793	0.887773
cardiffnlp/twitter-roberta-base-2019-90m-tweet-topic-single-all	all (2020 + 2021)	0.892499	0.774494	0.892499
cardiffnlp/twitter-roberta-base-dec2020-tweet-topic-single-all	all (2020 + 2021)	0.890136	0.776025	0.890136
cardiffnlp/twitter-roberta-base-dec2021-tweet-topic-single-all	all (2020 + 2021)	0.894861	0.800952	0.894861
cardiffnlp/roberta-large-tweet-topic-single-2020	2020 only	0.878913	0.70565	0.878913
cardiffnlp/roberta-base-tweet-topic-single-2020	2020 only	0.868281	0.729667	0.868281
cardiffnlp/twitter-roberta-base-2019-90m-tweet-topic-single-2020	2020 only	0.882457	0.740187	0.882457
cardiffnlp/twitter-roberta-base-dec2020-tweet-topic-single-2020	2020 only	0.87596	0.746275	0.87596
cardiffnlp/twitter-roberta-base-dec2021-tweet-topic-single-2020	2020 only	0.877732	0.746119	0.877732

Model fine-tuning script can be found here.

Dataset Structure

Data Instances

An example of train looks as follows.

{
    "text": "Game day for {{USERNAME}} U18\u2019s against {{USERNAME}} U18\u2019s. Even though it\u2019s a \u2018home\u2019 game for the people that have settled in Mid Wales it\u2019s still a 4 hour round trip for us up to Colwyn Bay. Still enjoy it though!",
    "date": "2019-09-08",
    "label": 4,
    "id": "1170606779568463874",
    "label_name": "sports_&_gaming"
}

Label ID

The label2id dictionary can be found at here.

{
    "arts_&_culture": 0,
    "business_&_entrepreneurs": 1,
    "pop_culture": 2,
    "daily_life": 3,
    "sports_&_gaming": 4,
    "science_&_technology": 5
}

Citation Information

@inproceedings{dimosthenis-etal-2022-twitter,
    title = "{T}witter {T}opic {C}lassification",
    author = "Antypas, Dimosthenis  and
    Ushio, Asahi  and
    Camacho-Collados, Jose  and
    Neves, Leonardo  and
    Silva, Vitor  and
    Barbieri, Francesco",
    booktitle = "Proceedings of the 29th International Conference on Computational Linguistics",
    month = oct,
    year = "2022",
    address = "Gyeongju, Republic of Korea",
    publisher = "International Committee on Computational Linguistics"
}