Training data for sentiment analysis [closed]

问题:

Where can I get a corpus of documents that have already been classified as positive/negative for sentiment in the corporate domain? I want a large corpus of documents that provide reviews for companies, like reviews of companies provided by analysts and media.

I find corpora that have reviews of products and movies. Is there a corpus for the business domain including reviews of companies, that match the language of business?

回答1:

http://www.cs.cornell.edu/home/llee/data/

http://mpqa.cs.pitt.edu/corpora/mpqa_corpus

You can use twitter, with its smileys, like this: http://web.archive.org/web/20111119181304/http://deepthoughtinc.com/wp-content/uploads/2011/01/Twitter-as-a-Corpus-for-Sentiment-Analysis-and-Opinion-Mining.pdf

Hope that gets you started. There's more in the literature, if you're interested in specific subtasks like negation, sentiment scope, etc.

To get a focus on companies, you might pair a method with topic detection, or cheaply just a lot of mentions of a given company. Or you could get your data annotated by Mechanical Turkers.

回答2:

This is a list I wrote a few weeks ago, from my blog. Some of these datasets have been recently included in the NLTK Python platform.

Lexicons

• Opinion Lexicon by Bing Liu

  • URL: http://www.cs.uic.edu/~liub/FBS/sentiment-analysis.html#lexicon
  • PAPERS: Mining and summarizing customer reviews
  • NOTES: Included in the NLTK Python platform

• MPQA Subjectivity Lexicon

  • URL: http://mpqa.cs.pitt.edu/#subj_lexicon
  • PAPERS: Recognizing Contextual Polarity in Phrase-Level Sentiment Analysis (Theresa Wilson, Janyce Wiebe, and Paul Hoffmann, 2005).

• SentiWordNet

  • URL: http://sentiwordnet.isti.cnr.it
  • NOTES: Included in the NLTK Python platform

• Harvard General Inquirer

  • URL: http://www.wjh.harvard.edu/~inquirer
  • PAPERS: The General Inquirer: A Computer Approach to Content Analysis (Stone, Philip J; Dexter C. Dunphry; Marshall S. Smith; and Daniel M. Ogilvie. 1966)

• Linguistic Inquiry and Word Counts (LIWC)

  • URL: http://www.liwc.net

• Vader Lexicon

  • URLs: https://github.com/cjhutto/vaderSentiment, http://comp.social.gatech.edu/papers
  • PAPERS: Vader: A parsimonious rule-based model for sentiment analysis of social media text (Hutto, Gilbert. 2014)

---

Datasets

• MPQA Datasets

  • URL: http://mpqa.cs.pitt.edu

  • NOTES: GNU Public License.

    • Political Debate data
    • Product Debate data
    • Subjectivity Sense Annotations

• Sentiment140 (Tweets)

  • URL: http://help.sentiment140.com/for-students
  • PAPERS: Twitter Sent classification using Distant Supervision (Go, Alec, Richa Bhayani, and Lei Huang)
  • URLs: http://help.sentiment140.com, https://groups.google.com/forum/#!forum/sentiment140

• STS-Gold (Tweets)

  • URL: http://www.tweenator.com/index.php?page_id=13
  • PAPERS: Evaluation datasets for twitter sentiment analysis (Saif, Fernandez, He, Alani)
  • NOTES: As Sentiment140, but the dataset is smaller and with human annotators. It comes with 3 files: tweets, entities (with their sentiment) and an aggregate set.

• Customer Review Dataset (Product reviews)

  • URL: http://www.cs.uic.edu/~liub/FBS/sentiment-analysis.html#datasets
  • PAPERS: Mining and summarizing customer reviews
  • NOTES: Title of review, product feature, positive/negative label with opinion strength, other info (comparisons, pronoun resolution, etc.)

  Included in the NLTK Python platform

• Pros and Cons Dataset (Pros and cons sentences)

  • URL: http://www.cs.uic.edu/~liub/FBS/sentiment-analysis.html#datasets
  • PAPERS: Mining Opinions in Comparative Sentences (Ganapathibhotla, Liu 2008)
  • NOTES: A list of sentences tagged <pros> or <cons>

  Included in the NLTK Python platform

• Comparative Sentences (Reviews)

  • URL: http://www.cs.uic.edu/~liub/FBS/sentiment-analysis.html#datasets
  • PAPERS: Identifying Comparative Sentences in Text Documents (Nitin Jindal and Bing Liu), Mining Opinion Features in Customer Reviews (Minqing Hu and Bing Liu)
  • NOTES: Sentence, POS-tagged sentence, entities, comparison type (non-equal, equative, superlative, non-gradable)

  Included in the NLTK Python platform

• Sanders Analytics Twitter Sentiment Corpus (Tweets)

  • URL: http://www.sananalytics.com/lab/twitter-sentiment

  5513 hand-classified tweets wrt 4 different topics. Because of Twitter’s ToS, a small Python script is included to download all of the tweets. The sentiment classifications themselves are provided free of charge and without restrictions. They may be used for commercial products. They may be redistributed. They may be modified.

• Spanish tweets (Tweets)

  • URL: http://www.daedalus.es/TASS2013/corpus.php

• SemEval 2014 (Tweets)

  • URL: http://alt.qcri.org/semeval2014/task9

  You MUST NOT re-distribute the tweets, the annotations or the corpus obtained (from the readme file)

• Various Datasets (Reviews)

  • URL: https://personalwebs.coloradocollege.edu/~mwhitehead/html/opinion_mining.html
  • PAPERS: Building a General Purpose Cross-Domain Sentiment Mining Model (Whitehead and Yaeger), Sentiment Mining Using Ensemble Classification Models (Whitehead and Yaeger)

• Various Datasets #2 (Reviews)

  • URL: http://www.text-analytics101.com/2011/07/user-review-datasets_20.html

---

References:

• Keenformatics - Sentiment Analysis lexicons and datasets (my blog)
• Personal experience

回答3:

Here are a few more;

http://inclass.kaggle.com/c/si650winter11

http://alias-i.com/lingpipe/demos/tutorial/sentiment/read-me.html

回答4:

If you have some resources (media channels, blogs, etc) about the domain you want to explore, you can create your own corpus. I do this in python:

• using Beautiful Soup http://www.crummy.com/software/BeautifulSoup/ for parsing the content that I want to classify.
• separate those sentences meaning positive/negative opinions about companies.
• Use NLTK to process this sentences, tokenize words, POS tagging, etc.
• Use NLTK PMI to calculate bigrams or trigrams mos frequent in only one class

Creating corpus is a hard work of pre-processing, checking, tagging, etc, but has the benefits of preparing a model for a specific domain many times increasing the accuracy. If you can get already prepared corpus, just go ahead with the sentiment analysis ;)

回答5:

I'm not aware of any such corpus being freely available, but you could try an unsupervised method on an unlabeled dataset.

回答6:

You can get a large select of online reviews from Datafiniti. Most of the reviews come with rating data, which would provide more granularity on sentiment than positive / negative. Here's a list of businesses with reviews, and here's a list of products with reviews.