CODEWITHSUNDEEP
azureazure-cognitive-services
Ferdinand Fejskid
Ferdinand Fejskid

Reputation: 537

What can I do with inconsistent sentiment detection provided by cognitive services?

Using Text Analytics for sentiment detection I receive sometimes results I consider being inconsistent.

They can be demonstrated by one simple example:

I'm sad was marked as 1% (0% means very negative)

Hello I'm sad was marked as 85% (100% means very positive)

Is there a way to improve/contribute to Text Analytics service for sentiment detection? Or to use own model similar to LUIS to detect sentiment?

Alternatively is there some recommended service/library to use to change input text prior to sentiment detection attempt to get better results?

Please see the way I used to test given example at https://azure.microsoft.com/en-us/services/cognitive-services/text-analytics/

screenshot with result

and the same result for via API using https://westeurope.api.cognitive.microsoft.com/text/analytics/v2.0/sentiment

Input: 

{"documents": [{"id": "101","text": "I'm sad","language":"en"},
{"id": "111","text": "Hello I'm sad.","language":"en"}]}

Result:

{"documents":[{"score":0.0038561224937438965,"id":"101"},
{"score":0.84333503246307373,"id":"111"}],"errors":[]}

Upvotes: 2

Views: 239

Answers (2)

Ferdinand Fejskid
Ferdinand Fejskid

Reputation: 537

I didn't find exactly what I was looking for but I consider last of my steps satisfactory.

1) No response when I tried :-) and :-( inside Azure Portal

2) No response whet I tried 1) and checked "Microsoft can email you about your feedback"

3) I received advice to post to Stackoverflow When I tried Twitter and have put @Azure #CognitiveServices into my tweet. The response came in less than 20 minutes.

4) Here I received an advice to ask the same question at [email protected]

5) After short email exchange I received following advice that answers my question in the most exhaustive manner so far:

One possible solution would be to use the MS Spell check API to correct the text grammar before sending it to sentiment analysis, and also considering breaking the text to smaller sentence segments if possible.

For example, if this is part of a chat conversation or a letter, and there’s a new line or a comma after the greeting.

We would also be adding a few enhancements in the near future that will enable you to influence the model by providing dictionary hints. So you’ll be able to tell the model to ignore words like: “hello” or “hi” if you see they affect your specific scenario quite often.

Please let us know if you have any additional feedback, and feel free to suggest features in our UserVoice.

Upvotes: 1

Luis Cabrera
Luis Cabrera

Reputation: 546

To answer your original question: The best way to provide feedback for incorrect scores is to contact the text analytics team ([email protected])

For the specific example provided, we'll look into what may be causing this issue. Somehow the machine learned that that was positive while of course we know it is not.

Luis Cabrera | Text Analytics Program Manager | Cloud AI Platform, Microsoft

Upvotes: 1

Related Questions