Twitter_data

Textual Content for Each Tweet

• Under the folder "twitter2015"
• Under the folder "twitter2017"

Visual Content for Each Tweet

• Under the folder "twitter2015_images"
• Under the folder "twitter2017_images"

Description

Data Split

• We randomly split our annotated data into training (60%), development(20%), and test sets (20%).

Format

• We provide two kinds of format, one is ".txt" for LSTM-based models, and another is "tsv" for BERT models.

• For example, each row of "train.tsv" for twitter2015 is one sample: (1). the first column is index; (2). the second column is sentiment label (0 refers to negative, 1 refers to neutral, and 2 refers to positive); (3). the third column is the id for the corresponding image of this tweet, which can be found in the folder "twitter2015_images"; (4). the fourth and fifth columns respectively refer to the original tweet by masking the current opinion target and the opinion target (i.e., entity).

  Note that each tweet may contain multiple opinion targets (i.e., entities), it may correspond to several continuous samples. E.g., the first and second samples in "train.tsv" for twitter2015 are about the same tweet but different entities.

• The ".txt" file is similar to "train.tsv", but every four lines in the file is one sample: (1). the first line refers to the original tweet by masking the current opinion target;  (2). the second line refers to  the opinion target (i.e., entity);  (3). the third line is sentiment label (Note that here -1 refers to negative, 0 refers to neutral, and 1 refers to positive);  (4). the fourth line is the id for the corresponding image of this tweet, which can be found in the folder "twitter2015_images".