The UGSD toolkit is a representation learning framework designed to generate domain-specific sentiment dictionaries from online customer reviews (UGSD: User Generated Sentiment Dictionaries from Online Customer Reviews) published in AAAI 2019. For example, given reviews on TripAdvisor, our framework attempts to generate sentiment lexicons of different ratings from 5-star to 1-star, which contain words conveying positive or negative polarities on attractions or tours respectively.
In specific, the UGSD toolkit can generate sentiment dictionaries of different domains given the reviews and the associated ratings in a specific domain. Especially, the lexicons from the UGSD toolkit include not only sentiment words but also representations of them, which can be used in other advanced applications such as sentiment analysis or dictionary expansion.
Please refer to the slides to know the main concept of the paper.
- python2.7
- Stanford CoreNLP
- Stanford NER software
- proNet
We provide two datasets, TripAdvisor and Yelp, which are used in our paper.
.
|-- datasets/
|-- tripadvisor/
|-- yelp_round9/
The TripAdvisor dataset was collected from TripAdvisor, where the top 25 cities in 2016 were selected, and the reviews of the top 20 attractions or tours of each city were included. The Yelp dataset was the ninth round of Yelp Dataset Challenge and you can see the offical Yelp website for more information.
$ git clone https://github.com/cnclabs/UGSD
$ cd ./UGSD
The UGSD toolkit uses Stanford CoreNLP to tag part-of-speech of words. So, we need to download the Stanford CoreNLP software from Stanford CoreNLP software official website.
Then, users need to rename Stanford CoreNLP directory to be "stanford-corenlp" and move it in this repository, or just add the Stanford CoreNLP software path when users run programs.
$ mv stanford-corenlp-full-2018-XX-XX stanford-corenlp
The UGSD toolkit uses Stanford NER to recognize named entity. Therefore, we need to download the Stanford NER software from Stanford NER software official website.
Then, again, users need to rename Stanford NER directory to be "stanford-ner" and move it in this repository, or just add the Stanford NER software path when users run programs.
$ mv stanford-ner-2018-XX-XX stanford-ner
The UGSD toolkit uses proNet to model the direct proximity of nodes in a network in order to learn the relationships between sentiment words and rating symbols. So, we need to download this package and compile them. Users can refer to proNet github repository for more details about the usage of this framework.
$ git clone https://github.com/cnclabs/proNet-core.git
$ cd proNet-core
$ make
$ cd ..
$ pip install -r requirements.txt
Users need to prepare reviews in a specified input format. At first, users specified a source folder, which is called ./data/ as the default path in programs. Then users will put all reviews in the path ./data/reviews/ in the JSON format, which are named as Amsterdam_01.json or Paris_05.json for example. Review files include the category, the rated entity and associated reviews. For each of reviews, there are information about the index, the rating and the corresponding review text.
Here is an example for the review format in all review files.
{
"category": "Amsterdam",
"entity": "Room-Escape-Games",
"reviews": [
{
"index": 1,
"rating": 5,
"review": "Fun! ..."
},
{
"index": 2,
"rating": 5,
"review": "Good place! ..."
},
{
"index": 3,
"rating": 4,
"review": "Nice! ..."
}
]
}
Then, after downloading Stanford CoreNLP, Stanford NER software and proNet, the default tree structure for this repository is listed as the following. Otherwise, users need to specify the needed path for tools when they run programs.
.
|-- data/
|-- reviews/
|-- Amsterdam_01.json
|-- Amsterdam_02.json
|-- Paris_01.json
|-- Paris_02.json
|-- proNet-core/
|-- README.md
|-- requirement.txt
|-- stanford-corenlp/
|-- stanford-ner/
|-- UGSD/
Users can run the shell script for the whole procedures.
$ cd UGSD
$ sh main.sh
However, users are recommended to run each program respectively because the steps for candidate sentiment word selection and review transformation take a long time. Then, we explain the usage and the parameters of each step accordingly.
Our framework selects candidate sentiment words from all reviews by leveraging POS information.
$ python SelectCandidates.py [--src_folder <string>] [--freq_thre <int>] [--corenlp_path <string>] [--ner_path <string>] [--verbose]
--src_folder <string>
data source folder
--freq_thre <int>
word frequency threshold for candidate sentiment word selection
--corenlp_path <string>
path for Stanford CoreNLP
--ner_path <string>
path for Stanford NER tagger
--verbose
verbose logging or not
In this step, our framework substitutes all entities in reviews with associated ratings.
$ python SubstituteEntity.py --src ../data/reviews/Amsterdam_01.json [--src_folder <string>] [--corenlp_path <string>] [--verbose]
--src <string>
review path
--src_folder <string>
data source folder
--corenlp_path <string>
path for Stanford CoreNLP
--verbose
verbose logging or not
If users need to have custom ways to recognize entities, they can write their own functions about the regular expression generation in accordance with reviews in different domains. And, they can pass the function to the constructor of the class in the program, SubstituteEntity.py. In the sample code, the program aims to recognize entities in reviews from TripAdvisor.
Our framework merges all processed reviews into a file for the next step, network construction.
$ python MergeCorpus.py [--src_folder <string>] [--verbose]
--src_folder <string>
data source folder
--verbose
verbose logging or not
In this step, our framework constructs co-occurrence networks, which will be used for the representation learning.
$ python ConstructNetwork.py [--src_folder <string>] [--min_count <int>] [--window_size <int>] [--vocab_path <string>] [--cooccur_path <string>] [--auto_window] [--verbose]
--src_folder <string>
data source folder
--min_count <int>
words with less than the minimum count are excluded
--window_size <int>
window size for constructing networks
--vocab_path <string>
path for vocabularies in the network
--cooccur_path <string>
path for co-occurrence networks
--auto_window
automatically calculate the window size or not
--verbose
verbose logging or not
Users use proNet to model the direct proximity of nodes in the network in order to learn low-dimensional representations. For more details of this package, please refer to the proNet.
$ ../proNet-core/cli/line -train ../data/network/cooccur/starred.txt -save ../data/network/cooccur/starred_repr.txt -order 1 -dimensions 200 -sample_times 25 -negative_samples 5 -threads 2
With the learned vector representations of words and rating symbols, our framework constructs dictionaries of all ratings by the maximum-cosine-similarity scheme and the z-score scheme. The generated dictionaries will be in the src_folder/starred_lexicons/.
$ python ConstructLexicon.py [--src_folder <string>] [--min_star <int>] [--max_star <int>] [--star_scale <int>] [--std_threshold <float>] [--verbose]
--src_folder <string>
data source folder
--min_star <int>
minimum rating scale
--max_star <int>
maximum rating scale
--star_scale <int>
rating scale
--std_threshold <float>
the threshold for the z-score scheme
--verbose
verbose logging or not
@inproceedings{WangUGSD,
author = {Chun-Hsiang Wang and
Kang-Chun Fan and
Chuan-Ju Wang and
Ming-Feng Tsai},
title = {UGSD: User Generated Sentiment Dictionaries from Online Customer Reviews},
booktitle = {The Thirty-Third AAAI Conference on Artificial Intelligence, {AAAI}
2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019.},
pages = {313--320},
year = {2019},
crossref = {DBLP:conf/aaai/2019},
url = {https://aaai.org/ojs/index.php/AAAI/article/view/3800},
biburl = {https://dblp.org/rec/bib/conf/aaai/WangFWT19},
bibsource = {dblp computer science bibliography, https://dblp.org}
}