Dataset Card for TROPICAL

Dataset Summary

The TROPICAL dataset is a French-language dataset for sentiment analysis. The dataset contains comments left by French-speaking tourists' on TripAdvisor after their visit to French Polynesia, each review either concern a hotel or a guesthouse. The format is JSON. The comments spanning from January 2001 to April 2023, the dataset contain 1592 comments along with 10729 ASTE triplets (aspect, opinion, sentiment). The unsplitted dataset is available in our Github repository.

Languages

The text in the dataset is in French as it was written by French speakers.

Dataset Structure

Data Instances

Normaly the polarity of the triplets are either "POS", "NEG" or "NEU", due to using ClassLabel the polarity is represented by 0, 1 or 2.

String label	Int label
POS	0
NEG	1
NEU	2

An example from the TROPICAL original dataset looks like the following:

{
  "id_comment": "16752", 
  "words": ["Nous", "avons", "passé", "4", "nuits", "dans", "cet", "établissement", "Ce", "fut", "un", "très", "bon", "moment", "Le", "personnel", "très", "aimable", "et", "serviable", "Nous", "avons", "visité", "les", "plantations", "d'ananas", "en", "4/4", "et", "ce", "fut", "un", "agréable", "moment", "nous", "avons", "fait", "le", "tour", "de", "l'île", "et", "c't", "une", "splendeur", "Nous", "sommes", "revenus", "enchantés"], 
  "triplets": [
      {"aspect_term": ["Aspect inexistant"], "opinion_term": ["revenus", "enchantés"], "aspect_position": [-1], "opinion_position": [47, 48], "polarity": "POS"}, 
      {"aspect_term": ["tour", "de", "l'île"], "opinion_term": ["une", "splendeur"], "aspect_position": [38, 39, 40], "opinion_position": [43, 44], "polarity": "POS"}, 
      {"aspect_term": ["moment"], "opinion_term": ["agréable"], "aspect_position": [33], "opinion_position": [32], "polarity": "POS"}, 
      {"aspect_term": ["personnel"], "opinion_term": ["serviable"], "aspect_position": [15], "opinion_position": [19], "polarity": "POS"}, 
      {"aspect_term": ["personnel"], "opinion_term": ["très", "aimable"], "aspect_position": [15], "opinion_position": [16, 17], "polarity": "POS"}, 
      {"aspect_term": ["moment"], "opinion_term": ["très", "bon"], "aspect_position": [13], "opinion_position": [11, 12], "polarity": "POS"}
  ], 
  "general_polarity": "POS"
}

Data Fields

'id_comment': a string containing the review id
'words': an array of strings composing the comment
'triplets': a list of dictionnaries containing the following informations
- 'aspect_term': an array of strings composing the aspect term (can be a single word or a multi-word expression)
- 'opinion_term': an array of strings composing the opinion term (can be a single word or a multi-word expression)
- 'aspect_position': an array of integers indicating the position of the aspect term in the words array (can be a single integer list or a list of integers)
- 'opinion_position': an array of integers indicating the position of the opinion term in the review (can be a single integer list or a list of integers)
- 'polartiy': an integer, either 0, 1, or 2, indicating a positive, negative, or neutral sentiment, respectively
'general_polarity': an integer, either 0, 1, or 2, indicating a positive, negative, or neutral sentiment, respectively

Data configurations

The TROPICAL dataset has 3 configurations: original, no overlapping, and overlapping.The first one contains the 1592 comments. The overlapping dataset contains the comments that have at least one overlapping triplet. The no overlapping dataset contains the comments that have no overlapping triplet.

Dataset Configuration	Number of comments	Number of triplets	Positive triplets	Negative triplets	Neutral triplets
original	1,592	10,729	9,889	734	106
no_overlapping	467	2,235	2,032	184	19
overlapping	1,125	8,494	7,857	550	87

The following table show the splits of the dataset for all configurations:

Dataset Configuration	Train	Test	Val
original	1,114	239	239
no_overlapping	326	70	71
overlapping	787	169	169

The split values for train, test, validation are 70%, 15%, 15% respectively. The seed used is 42.

Use this dataset

from datasets import load_dataset
dataset = load_dataset("TROPICAL", "original") # or "no_overlapping" or "overlapping"

Dataset Creation

Source Data

All the comments were collected from the TripAdvisor website. The comments range from January 2001 to April 2023. The dataset contains 1592 comments along with 10729 ASTE triplets (aspect, opinion, sentiment).

Who are the source language producers?

The dataset contains tourists' comments about French Polynesia stored on the TripAdvisor website.

Known limitations

The dataset contains only comments about French Polynesia. Moreover, the dataset is not balanced, the number of positive triplets is much higher than the number of negative and neutral triplets.

Datasets:

GePaSud
/

TROPICAL