The comments sections are WILD | YouTube sentiment analysis - Data science project for beginners | Tina Huang

0:00Of course, Ericor's spent 15 bloody minutes trying to find the cost of certification. But he sits at his golf course and I mean literally think about it. I think you probably played more than I do, Jim. America sent a fool on an area and he has cut her legs off and displayed the stops. So this is how it all started. Bye, Diana! Goodbye! For those of you that have reached out to me asking about beginner data science projects, I highly encourage you to also stay until the very end of this video, where I'll give it a run down of a notebook and how to pull YouTube comments yourself. So you can start your own analysis on your favorite YouTube videos or channels.

1:00So, I started off doing some sentiment analysis. I hear the problem. You guys are so nice to me. You guys are too nice to me. And okay, I mean the worst comment I've gotten with doing the sentiment analysis is I've got mad yellow fever. Not saying that you guys should be mean to me though. But for the sake of this analysis, let's also explore another YouTube and I know just the guy. Don't make me laugh. Hey, you're literally making me laugh right now. Okay, you have to be normal. Hey Ken, how's it going? Great, how about you?

2:01Good, good. How are you papayas? They're pretty good. I only had three today. All right, I called you for a reason. It's something you're supposed to say Ken. I actually did some sentiment analysis on your channel using two different types of sentiment analysis modules. One is called text fault and the other one is called baited. Are you ready to hear some results? So basically this is going to be like, are you ready to react some wholesome and mean comments? Yeah, absolutely. Let's see what I mean, I have read every single one of my comments. So hopefully there are no surprises. But I've forgotten some by design. We'll see what happens when you unearth them. All right, so okay, what do you want to do? Hold some ones first or mean ones first? To rip the bandaid off, let's go for the mean ones. Let us see. First order from Vader. Are you ready? A abandoned video at 0.38. Want to know why?

3:02Read. First video to get, visit videos to get a quick subject and topic list of math required for data science. Videos suggest articles read on topic. Stop video video article offers link to Corsair of course on topic. Visit Corsair of course. Spend 15 bloody minutes trying to find the cost of certification. Do not find it. Give up search frustrated. Return here to leave comment. Abandon watching and rest of the video. Conjure of images of wanting to punch the wall in frustration on time wasting results still on the crap. Comment up. Now 30 minutes later, I still need to list the subjects and topics. Well, I think if that person had just watched the whole video, I would have actually given all of the subjects and topics. I feel bad for that person, but I also wish they had taken just maybe a little more time to to watch past like 40 seconds into the video. Here's a whole text ball. Very annoying background music also in podcasts. Otherwise, perfect.

4:02That one's not bad. That's like constructive criticism. My biggest problem is communicating. I'm very bad at communicating. But what do you think? I feel like this person is pretty good at communicating. But they're bad at communicating. And maybe they're a very effective communicator via the written language, but not the spoken language. So I actually did notice I gave that example because both beta and text block tend to classify things as negative, just because there's negative words in it. Right? So I feel like that's mean that it's not very good at doing is where is the sentiment directed towards? And I wonder if this is something that's kind of ubiquitous in classification. I think it probably is. I mean, that's clearly a negative sentiment. Like I'm not good at something. It's really hard, especially in this to get direction. I think of sentiment as kind of a blunt force tool, where it's great at, let's say, evaluating 100,000 tweets. You can tell on a topic. Most people are positive or negative about this, but on a comment by comment basis,

5:02I think we can both agree after looking at this data that it's a little shaky. Let's move on to some wholesome points. It's beautiful how you're thankful for us, but actually we're happier and more thankful for you truly. You educated and inspired me and led me to great resources. Also, thank you because your work is not just useful, but also beautiful and well-made. I can't wait to reach my fifth not bad project and tenth nice project and 15 amazing project and look back at the journey, a journey full of learning and growing and amazing experience. Well, I'm like, I'm legitimately turning red. That's that's something that if you're putting stuff out there into the world, I mean, that's some of the nicest things that someone could say about your work. You just have so many wholesome comments. Some advice, always the best tips. Thanks, man. And there's another one that's like, these are great tips for me since I myself work from home.

6:04That's also a great game. So, what do you think are the ten words that most really differentiates you based on your comments? So, let me give you some context first on what it is that I did. So, I use scatter text with species English, Core Web, large, healthy model, and I created a corpus with a collection of words. Then I found the ten words that are the most defining of your channel. Well, I do everything you just said sounded like a foreign language to me, but probably I would expect data and science to be in there. Probably project. Maybe Kaggle tens a lot of words. Papaya. I know that's not in there. You wish. You wish. I'm not sure. Not bad. Not bad. So, the ones that came out, Kaggle is in there for top 10. Tweets YouTube. Gee, your last name. Coursera.

7:05Jupiter, you did me Twitter and LinkedIn. Interesting. But when I expand that to 20 though, you do get data science coming up, you get like data camp coming up, Numpey, Twitter scraper, bootcamp, COVID, Colock as well. I wonder what G is what. Maybe it's it's because people, some people call me like Mr. G or something like that. Or I get a lot of serves. You can just call me camp guys. Honestly. But they call me sir too. Serena. There's actually so many ways of improving this. Like using themes as opposed to words. I have a lot of ideas to improve this in the future. Let me know in the comments below too, what you think I can do to improve this. I also send out some of the stuff that I found. And seeing as you know, the scoreboard that your building can, I think we can really expand that out. And just having like a whole slew of things that could be useful.

8:05Heck yeah. I think that there's tons of opportunity here. I'm pretty excited about the kind of. Potential to do more projects around you too. Let's take this to the next level. Okay, so Ken sentiments are generally very positive. Although he's had some variations over time, it makes sense though because he's a really nice guy and has really useful content that's not really controversial and hasn't been involved in any drama that we know of at least. Okay, I don't know he's going to do anything heinous soon. Which unfortunately for this sentiment analysis means that he doesn't have any main comments or polarizing opinions. So I thought to myself, what is the single most polarizing thing in the past few months or so? The pandemic. Yeah, probably the pandemic, but that's been gone for like a year now and COVID is already taking over our lives. So I was like, let's do something else. The US collection. So I grabbed the comments from this video which covers the first residential debate.

9:07They have a plan. He won't even meet with them. The Republicans won't meet with us in it. But he sits at his golf course and, I mean, literally think about it. You probably won't be more than I do, Joe. There's some of the comments that got the most likes. Gordon Ramsay will be a much better moderator in the next debate. This isn't a presidential debate. This is an emergency meeting discussion in among us. The debate is hilarious as hell until you realize one of them will be president. Completely agree with that one. And many more. I like the last one too. They need an Italian grammar as the moderator. Yes, much better. As you can see, this is much more polarizing. And both Vader and Tex-Bop do a much better job at classifying sentiment. So here's some of the most positive comments. Biden, 2020, a lot of emojis, hearts, really long,

10:09civil-acquise. That's about salvation of souls. Best pizza night. Make Russia great again. Okay, that makes sense. America sent a fool on an area and he has cut her legs off and displayed the stops. Chris Wallace, a terrible moderator. Yeah, that makes sense. Pathetic. Yeah, I can see why. That would be definitely classified as negative. I also attempted to do some cameine's clustering to see if I can find clusters of comments. I first did more preprocessing. As you can see, this data set is really kind of unclean. In the sense that there's like a lot of spelling errors, a lot of issues in terms of like, the structure sentences, comments that literally don't make any sense at all. So I tried to do as much preprocessing as I could. I played around with stemming versus tokenizing, also like removing emojis, trying to check spelling. After all of that, I then used an elbow graph

11:10to show the most awful number of clusters. Which honestly didn't show a super clear elbow. But I guess two scenes promising and I was like, oh, maybe it's like one is Trump and one is Biden. But these are actually clusters that we ended up with. You get Trump, Biden, I did debate because I used stemming in this case. So the words are not four words. But debate, Joe, like president, this, and all these things. So and then for the other one, you get Trump, that word, go vote Biden, you know, savage. Biden, notice that towards the end. Here you get beautiful. I don't know clearly. Peace, bless, hard heard, love. So I actually feel like what the clusters ended up being is the first one is much more negative. And the second one was, I mean, it was kind of mixed, but seems to have more positivity involved in it.

12:10So I thought that was interesting, because I was really like, expecting to see the Biden Trump split. But yeah, I don't think these are really good though. So I try to do some more processing, but unfortunately to no avail. I think the biggest challenge for this data set though is just sufficiently cleaning the comments, since you get some pretty wild stuff. What do you guys think? I hope this video was insightful on how I went about going through my first ever project as a total NLP. It's nowhere near perfect. And I would also argue that projects are never finished. I do want to say lots of thanks to Kenji for interrupting how to script you to comments and the club. To everyone that wants to do a super beginner-friendly project, where I also provide you guys with most of the code already, you can find links to the notebook in the description's below. All you have to do is go to the YouTube ATAPI, get your developer key and replace the developer key with yours, and then you can run that code. You can then switch off the channel ID and YouTube IDs

13:12with those other channels in the videos that you want. I uploaded two collabs since that was the easiest for me, but you can also play around them with the data in any way that you like. And there you go. Try out the analysis I've done and take it away. I would love to see any analysis that you guys do, so please let me know what you do. And let me know if you improve my analysis too, possibly by trying out some of the suggestions that I mentioned earlier. Next up, I'm really enjoying learning about NLP, and I was laying in bed. I had some other ideas. So do let me know what you guys think about this type of video, and if you guys like watching, and if you do enjoy watching this type of video,