Hacker News Reader

Comment by byoung2 | Hacker News Reader

You're looking for sentiment analysis and you won't get good results unless the text is in a specific domain. I did this work for a previous company where we analyzed restaurant reviews to extract actionable feedback. The trouble is that the written word is missing most of the nonverbal cues that give you the tone for spoken word (think air quotes or rolled eyes). So when you see a review that says "My steak came out well done...great job guys!" It looks positive but could be sarcastic. Or when someone says "We came to Taco Bell because we were looking for the most authentic Mexican cuisine" of course this is sarcastic but there is no way you could teach a computer to spot that.

View on HN

Hey now, taco bell is the sh*t! But seriously, thanks for the insight. Do u think this is solvable? It would be so helpful for online discourse.

Like I said, if the domain is super specific, it is solvable. That Taco Bell example is real. A private equity company was considering purchasing them from Yum Brands and they wanted to know what the customers liked or disliked. If all of your text is Taco Bell reviews, then you can train a model to spot sarcasm by having a human tag a portion and letting the model extrapolate from there. The model will determine which words (e.g. authentic, gourmet, and homemade) correlate with sarcasm. This breaks down if you try to apply it to all restaurants because what is clearly sarcasm for Taco Bell may not be for Lupe's Mexican Cantina. In the end, we went with a more objective approach. I broke the reviews into sentences and analyzed independent and dependent clauses, identified nouns, verbs, and adjectives, and used those to determine the main idea of the clause. I noticed that 1st person nouns and verbs are what the reviewers said about themselves (e.g. I was hungry or I rolled into the drive-thru) and 2nd and 3rd person was about the restaurant (you need to train your employees, the taco was cold or the cashier gave me the wrong order). Looking at the sentiment of the clause and its subject/verb phrase gave us exactly what we needed. In the end we had a list of ideas like "cold taco - negative", "waited forever - negative", "super cheap - positive" which we used a logistic regression classifier to put into categories like price, service, selection, ambience, etc.

Taco Bell is food for chihuahuas, not humans.

To correctly analyze the sentiment of that, your software would have to have the "Yo quiero Taco Bell" ad series from, what, the 90s maybe, in its database. It's snarky, but it's not the outright slam that the words themselves would seem to indicate.

That's the problem. There's sarcasm. There's parody, which if done well can be impossible to tell from the real thing. (That's somebody's law - too lazy to look it up right now.) Those are bad enough. But when a sentence drags in cultural knowledge, you have to understand the cultural knowledge to accurately assess the sentence. It's tough to ask a program to "get" not just what the reference is, but how that affects what a sentence is actually saying.

Reply on news.ycombinator.com