Hacker Newsnew | past | comments | ask | show | jobs | submit | owen_pearson's commentslogin

This is great, but the sentiment analysis is way off. Take these 'negative' reviews for the Staples Hyken office chair:

> For months i was waiting for the Staples Hyken Mesh chair to drop to a good price. I pulled the trigger at 150ish, received it really fast, and set it up right away. I actually really love the chair, totally worth it. Perfect size for me and i can sit in in for 12+ hours just fine.

> I personally haven't found the thinks to be more comfortable than other chairs, I'd rather go with a staples Hyken. Also, I see these go for much less. This looks like a V1. I sold one like that for $200 in Virginia.

> Hello all, I went on a deep plung into chairs over the weekend to find a good new chair (instead my junky racer/gaming chair that hurt my back and neck. I ended up finding the Staples Hyken and Staples Dexley. I went to store and tried them both out. I really liked the Hyken way more than the dexley....

In the small number of negative reviews I've looked at it seems like around half of them are categorised incorrectly.


That's indeed an issue. We're currently working on a new version of that, stay tuned!


GPT-3 does a much better job of sentiment analysis. Parts in italics are output from GPT-3

------------

I am a sentiment measurement bot. I take arbitrary text input at the INPUT: prompt then reply with a SENTIMENT: output that includes a numeric sentiment indicator ranging from -5, negative to 5, positive, a PRODUCT: output that extracts the name of the product being described and a CONFIDENCE: output that estimates the confidence in in the rating, from 0% where there is no confidence, to 100% where there is complete confidence.

INPUT: For months i was waiting for the Staples Hyken Mesh chair to drop to a good price. I pulled the trigger at 150ish, received it really fast, and set it up right away. I actually really love the chair, totally worth it. Perfect size for me and i can sit in in for 12+ hours just fine.

SENTIMENT: 4 PRODUCT: Staples Hyken Mesh chair CONFIDENCE: 100%

INPUT: I personally haven't found the thinks to be more comfortable than other chairs, I'd rather go with a staples Hyken. Also, I see these go for much less. This looks like a V1. I sold one like that for $200 in Virginia.

SENTIMENT: 3 PRODUCT: Thinks CONFIDENCE: 60%

INPUT: Hello all, I went on a deep plung into chairs over the weekend to find a good new chair (instead my junky racer/gaming chair that hurt my back and neck. I ended up finding the Staples Hyken and Staples Dexley. I went to store and tried them both out. I really liked the Hyken way more than the dexley.

SENTIMENT: 4 PRODUCT: Staples Hyken CONFIDENCE: 100%

---------------------

You can even use it to extract comparisons. Again output from GPT-3 in italics:

---------------------

I am a comparison evaluation bot. I take arbitrary text input at the INPUT: prompt then reply with a COMPARISON: boolean output that indicates if the text is a comparison between products and a BETTER: output with the name of the better product and WORSE: product with the name of the worse product

INPUT: For months i was waiting for the Staples Hyken Mesh chair to drop to a good price. I pulled the trigger at 150ish, received it really fast, and set it up right away. I actually really love the chair, totally worth it. Perfect size for me and i can sit in in for 12+ hours just fine.

COMPARISON: FALSE

INPUT: I personally haven't found the thinks to be more comfortable than other chairs, I'd rather go with a staples Hyken. Also, I see these go for much less. This looks like a V1. I sold one like that for $200 in Virginia.

COMPARISON: TRUE

BETTER: Staples Hyken

WORSE: Think

INPUT: Hello all, I went on a deep plung into chairs over the weekend to find a good new chair (instead my junky racer/gaming chair that hurt my back and neck. I ended up finding the Staples Hyken and Staples Dexley. I went to store and tried them both out. I really liked the Hyken way more than the dexley.

COMPARISON: TRUE

BETTER: Staples Hyken

WORSE: Staples Dexley

-----------------------


This is pretty amazing; were these single attempts?


Yep! I’m guessing one of the openly available models would do similarly as well. Obviously more expensive compute-wise, results are pretty good though.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: