项目作者: tianyiwangnova

项目描述 :
a PyPI package for analyzing customer reviews using aspect-based opinions mining and sentiment analysis
高级语言: Python
项目地址: git://github.com/tianyiwangnova/2021_project__ReviewMiner.git
创建时间: 2021-01-21T16:36:04Z
项目社区:https://github.com/tianyiwangnova/2021_project__ReviewMiner

开源协议:MIT License

下载


ReviewMiner

PyPI version
Build Status
codecov

reviewminer is built for analyzing customer reviews, or any text datasets that are similar to review data (short opinions
collected from multiple individuals)
. It is built on top of nltk and TextBlob.
reviewminer takes the pain out of building NLP pipelines (for analyzing customer reviews) and provides handy tools
for quickly organizing review data into digestible insights.

Features:

  • Aspects and opinions extraction The key methodology in this package is aspect-based opinoins mining. The package has
    its own algorithm to extract aspects and the related opinion words from the review data.
  • Sentiment on comment and aspect level The package can offer sentiment scores on both comment level and aspect level
  • Negative reviews investigation The users can quickly display the negative sentences in the comments. They can also
    investigate negative comments by aspects.

Installation

  1. $ pip install reviewminer

Quickstart

One-stop text analysis

We use the Women’s Clothing E-Commerce dataset on Kaggle to run the examples.

  1. import reviewminer as rm
  2. import pandas as pd
  3. # read our sample data
  4. reviews_df = pd.read_csv("https://raw.githubusercontent.com/tianyiwangnova/2021_project__ReviewMiner/main/sample_data/Womens%20Clothing%20E-Commerce%20Reviews.csv")
  5. # create a reviewminer object
  6. sample_rm = rm.ReviewMiner(reviews_df, id_column="Id", review_column='Text')
  7. # run the one time analysis and you will see
  8. sample_rm.one_time_analysis()

The function will print out 4 visualizations:

  • Popular aspects and opinions
    popular

This chart displays 9 most common aspects found in the reviews and the most popular opinions words people used to
describe them. In each bar chart, the heights represent the percentages of the people using the opinion words.

  • Distribution of sentiment scores of all comments
    sentiment

  • Radar chart of the most common aspects and their average sentiment scores
    radar

From this chart you can quickly compare customers’ average sentiment on each of the common aspects. Here “size” seems to be an aspect that customers are not quite satisfied with.

  • Aspects with the most negative comments
    negative

Exclude certain aspects

You might want to exclude some aspects. For example, if you don’t want the aspect “colors”, you can do the following:

  1. print("Before:", sample_rm.top_aspects)
  2. sample_rm.aspect_mute_list = ['colors']
  3. print("After:", sample_rm.top_aspects)

exclude

When aspect_mute_list has changed, the visualizations will change as well when the related methods are calling, but the
base intermediate output tables (e.g. aspect_opinion_df) won’t change.

Check out negative comments of an aspect

From the radar chart above we saw that customers might not be very satisfied with “sizes” of the clothes. Let’s check out the negative comments around “size”

  1. sample_rm.negative_comments_by_aspects_dict['size']

size

Check out the most common opinion words of an aspect

  1. sample_rm.single_aspect_view("material")

material

This dataset is not very large so the numbers are not quite prominent.

Radar chart of average sentiments for a list of aspects

  1. sample_rm.aspects_radar_plot(['shirt','skirt','sweater','blouse','jacket','dress'])

radar_customized

Tips

  • It’s better to feed in review data on a specific product or service. If you run it on the review data for a specific
    ramen restaurant, it’s easier to find meaningful aspects. If you feed in Amazon reviews for 5 totally different
    products, the insights might not be very clear.

  • Sometimes a sample of the data can tell the whole story. If you have a million reviews, the result will be very similar
    to the result you get from a random sample of 10k reviews. Don’t rush to feed all your data in, try with a sample first ;)