del.icio.us
blinklist
digg
Facebook
Furl
ma.gnolia
Newsvine
Pownce
reddit
StumbleUpon
Technorati
Twitter
Audio:
Audio clip: Adobe Flash Player (version 9 or above) is required to play this audio clip. Download the latest version here. You also need to have JavaScript enabled in your browser.
Slides: Sentiment Analysis
Transcript:
Okay, so off we go. Hold up your hand if you said wait please. Half of you sure, I have got a job today here. Right so how does it work guys? [some one enters] Okay, so who am I, I have got a few personal tags; I made that after studies today. I am the Founder and Chief Executive of the company called Brand watch Repulse Monitoring Company. We have been doing, we have been around for about 4 years for the product launched 2 years ago, 2 years of R&D before that. We monitor the web and we analyze what we were saying in the social media and in the news. …. and I am getting attached to November, and I am really regretting it in a big way. But if you want to sponsor me, please go to November and search for me. Okay, so over there, over there , sentiments so why…I would talk about sentiments ……sentiment analysis …part of it we do at Brand watch so I reasonably qualified to talk about it. Why is it important? It’s important if you imagine internet is the world’s biggest focus group. People have said this before. Lots of people communicating, talking, throwing comments out there, if you could figure out what they were talking about and understand it, it would probably be a very useful resource and also you know asking them questions just kind of listening which you say is biasing, biasing it….in particular if two people ..I just picked up a couple of …this one is going to be interesting …This is Gordon Brown speech last year, ..Before and after in social media, before and after in professional media. Basically professional media news there was an increase in positive sentiment…and with social media there was an increase in negative sentiment which is going to be interesting. (01:42) And I think that reflects the way we feel about Gordon, we are not really that kind of warmth towards him already, but in view seem to be more so…also companies like it because its predictive this is a chart we do for Edge online games magazines, biggest jump this was a few years ago now [xxx].2010. and sure enough [xxxx] 2010 was the biggest seller, couple would say so, so I am not saying it, it’s absolutely predictive but is definitely a measure. As Robin was saying first thing in the morning that, historically people talked about attitudes driving behavior actually does, his argument was the other way around behavior drives attitude which I think is probably true to an extent but I think attitude do drive behavior. You only will have to look at word of mouth marketing, review sites and more people saying about various things you ask. You have to get online and take a look at what other people’s experiences are and their attitudes tends to drive your behavior. So I think, I think sentiment and attitudinal data is pretty influential. So what I am going to do is I’m going to take a deep dive, which is the phrase of one of our clients … (02:55) I am really ….I am going to talk about simple analysis and how it works … take a deep dive which is one of our American customers use\age phrases I am going to use it much more in conversation. I think I will add something to that…to my…..inter sentimental analysis…first of all I am going to hopefully …with that is a little bit technical try and make it more….but…I’m going to say some stats we have got might be interesting here is the…. where is my picture ? …there is an image; there is a beautiful picture which shows … the web is 58% neutral…. Am so that’s we marked up and manually looked over 4 hundred thousand web pages in the last few years… cross section of different things …1000 different things people…places, politics, brands, products companies, and 58% of whole content is neutral. (04:00) 24% is positive and, sorry 28% is positive and 14% is negative if I remember correctly, which shows that we are as a kind of a race…as a population….this is in English language. Twice as I say something positive is negative….which am, which isn’t the case of news, right you don’t see that on the news [xxx]so interestingly when we talking in social spaces, we are more positive than we are than negative…. Which I can’t say that’s interesting. But when it comes to sentiment analysis there is a little bit subjectivity involved…same school….john on the left there and somebody else from…on the right, so same result same data, different take on the sentiment.,. and we did a so there is an element of subjectivity we all do…and we are going to test where we give a thousand pieces of information to two people and two native English speakers and variety of different topics ..So wasn’t just about one thing. The one person may be more positive about and the other…..it was a kind of broad reasonably, I think reasonably scientific controlled test and we found that people agree with each other 85% of the time when it comes through three. Three bucket sentiment things positive, negative and neutral so there is some, that is interesting…here is some stats that we pulled out of the system yesterday ….over here we have got the volume of conversations…October, English speaking web, the volume of unique mentions ranked by the top to bottom, see who comes on to top [xxx] that’s surprising [5.44], European Union edging ahead Barrack Obama he would probably be too pleased about that… Windows.7 jumping right up there…and if you add in volume…sentiments. So this middle column takes a count of the sentiment of the posts…Google still coming on top, Microsoft amazingly jumping from fourth to second. Microsoft would be pleased about that. Sony jumping too from eight to third, eBay doing pretty well Barrack and EU drops to the bottom probably because of a lot of; Beatle is unbelievable they are still coming in the top 15… And if we look at just center … this is the average sentiments score and it goes up to ten. For everything on the web (Bridgton) comes on top. which is kind of interesting hurry up probably because we say fabulous…most right … [xxx] comes out for fifth, which seems completely wrong and I will come back to that in a minute. I so it little….you think ….that’s my favorite add [xxx] pointing up in the air. .I just love that…so… I found that last night… how you measure as I said I am going to get boring now. How do you measure the sentiments of millions of web pages, you could get crowds of it, you could go to ….Mechanical Turk.if you go to there is some other…..doesn’t want to be part of the crowd…and you can get people….you can give them bits of text and you can make them say negative, neutral positive….in our experience you have to pay reasonable amount money to do that (07:18) We tried it we offered 5P for people to mark up a paragraph as positive negative neutral with respect to a particular brand of keyword and we found that we we got…pretty shabby results back because 5 P wasn’t enough so people just kind of clicking everything. So you got to pay reasonable amount it’s slow, it’s inconsistent you got to you are using different people around the world …train them potentially and is expensive. So by calculations 10,000 articles will cost about about 2 grant, which is part of money so can we use machines well…..yes we can we do… and reliably, kind of and how to do. Okay you use artificial intelligence you use software called machine learning software basically you teach the software, you teach the program to do what you do…to do what I should be doing with you….so you decide what text to feed in. and of course if he has got a web page which is talking about different things, which I’ll show you in a second…and you have got to make decisions which part of that page to feed in to the machine to ask it to do some sentimental analysis you got to train it you got to basically tell the machine, show how to do its job before it does itself and you got to keep doing it again and again it is not quick, it takes a long time. So here’s a page, this is just a random page that I pulled off last night, talked about the iphone loss .so if I was doing sentimental analysis on the iphone.,.,,. I probably put that whole page in. There is some bits here which don’t necessarily talk about it but I think that is probably all about the iphone, whereas this page is a forum and there is a few little mentions down here and this page goes on and on and on. so making a decision about which part of that page is talking about the iphone is a lot harder. (9:01) How do you then train machines, one minute, half minute, 20 secs, Jesus Christ! So you basically have to get somebody to do it first…also take this machine to millions of [xxx] and then it gets reasonably good at it and you can’t do a foreign language this is all about statistics. We have done it we have got 6, 50,000 mentions that we marked off across 60 different industries because the language which is in different industries is very different.. online games is different…it is very different, insurance for example and you cant get roughly 75% accuracy figure…that is bloody hard. So finally two things, the small volumes[xxx]because you can afford to pay for it…the larger volumes has got a big data set you have to do sentimental .analysis on it….you can use machines but is you can get the top 5% and you can look at it even better…so it’s a bit of a recap… and in the end we have got new product coming up…At the end of this month we are giving Away 50 [xxx] accounts so if you want one go to there are sign off[xxx]/4 . thank you very much.