I've been working in a system, which evaluates products based on the comments of the consumers, but I've had some issues in detecting the patterns of responses. I was considering applying some statistical test for it. Which test I should use?
You might want to try some simple measures of lexical association, like the t-test and the log-likelihood ratio. This will help you find bigrams (and perhaps longer ngrams) that occur more often than chance, which can perhaps be useful. Various implementations already exist, including those in the Ngram Statistics Package http://ngram.sourceforge.net
That would depend on what the responses look like. Are they open, constructed responses where the consumers freely write in answers? Are they answers to a Likert-scale survey (i.e. 1 = strongly disagree, 6 = strongly agree)? Are they dichotomous, yes/no questions like "would you recommend this product to other people"? The type of test you'd use would depend on the type of data you have.
I'm sorry Ryan I didn't say about how are the questions. It are open questions and the clients freely write everything they want about the product ou service. The database which I'm working is a site for "happy costumers" they write down good things about the products and all the answers with bad thngs about are deleted. So, they usually recommends about the store or product. I was constructing a system which parsing all the database with some keywords as a presets (like "good product" "I recommend", etc) after that I was applying a factor analysis for categorize the data (like good product is more related of fast delivery than low prize), but I think factor analysis is not good teste for it.
Well, CFA (Confirmatory Factor Analysis) is commonly used for similar purposes, so I don't think it's necessarily bad, but you can follow that up with more specific tests as well. For example, it seems that you might be looking at frequency of co-occurrences like how often "good product" or similar terms were mentioned together with price, delivery, ease of use, durability, or some other thing. The simplest, most direct way to see if there is a statistically significant difference in the frequency of those associations would be to do a chi-squared test. That would tell you whether the observed distribution was different than a hypothetical distribution where there was no relationship between the pairings.
You might want to try some simple measures of lexical association, like the t-test and the log-likelihood ratio. This will help you find bigrams (and perhaps longer ngrams) that occur more often than chance, which can perhaps be useful. Various implementations already exist, including those in the Ngram Statistics Package http://ngram.sourceforge.net
Check out what you find typing katerina veselovska subjectivity into google. It seems she does pretty much what you are trying to do
A good way is to use RQA that does not impose you any type of constraints. You can use it in a global way measuring the global recurrence of a text (with no choice of a specific pattern to be tested) or in a focused way selecting a giving pattern of interest and measuring the departure from chance of its dynamic of recurrence along a series.
See paper attached.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics