SF Text: Stefan Gower, Q&A with Alexy Khrabrov @Lithium
Recording: SF Text: Stefan Gower, Q&A with Alexy Khrabrov @Lithium
hello everybody I'm Alexi krau the organizer of SF text a new Meetup devoted to text mining NLP search Ai and basically uncovering human intent behind the text and here we are with li at lithium and we have Stefan go with us who who is the creator of topic Scout uh a new uh company in new technology which is very interesting because uh it says it can do high quality topics at enormous scale which we didn't previously see so it's uh it's very interesting at first I want to ask Stan can you tell us a little bit about your background uh your technology uh tool set and what you know what made you what interested you in this space well my my background in includes work in large scale expert systems including some unusual Ones based on S of uh sort of almost functional flows that were very large uh and and those did forms of classification as well very differently uh I got into idea of classification because I got exposed to some tools whose names I won't mention but they just didn't work very well you could get them to work for maybe a 100 topics with a lot of effort maybe 200 and it was clear it just got worse from from there mhm so it seemed to be an interesting problem and that's how I got that's how I got into it uh really how I got into this one day it's the only time it really happened in my life I had an epiphany walking around a lake and it was very quick it's the only time this has ever happened to me and it sounded like well that might really work took me four months to run the experiment and it was very promising and that's how I got into it and it turns out to be a lot of work much more than I ever ever anticipated uh but there there is a way to actually do the topics that scale interesting looking forward to your talk about that but why topics you know there are so many different applications and algorithms why topics specifically well Ian I think I think a lot of people are engineering thing see something exposed to it may come away sometime and you think know that thing didn't work very well could I do better I mean it doesn't really matter what subject area it is but a lot of people in technology that's how we think then sometimes you know we not we we have the observation and maybe something comes to us or we figure the problem's worth attacking and it kind of leads from there so if we do find a way where we think well we could do better maybe maybe we're right maybe we're wrong then maybe we go attack the problem that was kind of the case here though it was interesting when I started and it started a long time ago like 2001 2002 people then when they talked about topics experts in the field would say well who would need more than 100 or 200 topics I'm really not joking now the world has changed a lot in the years that followed and almost a decade and a half ago people don't question the importance of topics anymore I think people have a very different attitude to them today yes because of all the social media and the web and all the text content right and I think ad targeting especially there is a lot of interest in in very specialized topics uh so you mentioned you were working on Expert systems before and you know what's fascinating to me in the uh uh computer science it's very cyclical so some things you know rise and fall so there was a time I I remember you know I might dating myself but you know expert systems was a thing it was it was hot right and then you know there were a lot of books and expert systems and how you build expert systems how you manage the knowledge and then it all gone away and I'm wondering you know where all where did all this expertise go right like what do you think happening like are we Reinventing the wheel you know are we discovering this all over again or did this knowledge get transfered somehow to to the new uh incarnations well I mean we don't have a long time to talk so I can just say that what I've done in topic Scout has really nothing to do with rule based experis because they don't scale and topic Scout is all about scalability yes so if something doesn't scale it doesn't matter from day one my goal was was any algorithm when scale throw it out the window I thought from day one from going to 100,000 topics it was my goal from very early on so expert systems actually if you want to study them the rule-based ones teach you a lot about scalability problems because generally they don't they they're very hard to maintain so in this case they're almost a counter example and what I was want to work on with things that would you know would achieve scale and maintainability very importantly okay uh can you talk a little bit about your tool set what programming languages are using what technology are using because you know if you need to scale to that to that size is it a distributed system is it on single node can talk a bit about that yeah I I can though I think in this particular case it's more about the algorithms which are very proprietary so I don't talk about those really uh which is really where the scale comes in the language I happen to do it what is in Java because the language I'm comfortable with and I like but it could be done in some other other language as well okay and in terms of scaling uh you know the system as it stands now is designed a certain way to allow it to be paralyzed but that was very much by Design but as we all know don't paralyze something until it works someone who taught me a guy who was very smart said to me uh you know build things so you can optimize them but don't optimize too early in the same way you might want to design something you paralyze it but make sure it works first before you go you know do the final step in terms of Stack there besides Java does use map ruce though map ruce is really a means to an end uh in this case it's but it works fine actually it actually really quite well actually uh so Java M ruce uh it it uses file systems a lot not really no SQL it's it's usually very text text has a lot to do with files right yes so there's just a lot of files I mean I can't remember for 10,000 topics it's something like I don't know is it 20 million documents or something there's some very large number of documents involved in it uh I don't quote me on that number that I'm speaking to a camera now I'm pulling out of my head but largely it's about about files and processing them well okay I mean this is what you know we as programmers know very well right text and files so this you know um very comfortable with that uh and you know Java is uh an interesting good choice you know I'm also happen to help organize s of scholar and I'm always interested in kind of uh doing scientific Computing on gvm so I wonder uh do you think that gvm is kind of good choice going forward for this kind of work do you find I mean is just a convenience for you do do you find gaps in this it's really just a convenience I mean it's a kind of system which could be done in different language and I really think it would be you know programmers really go to different language for different purposes some for style or choice or performance so I think with certain algorithms you have a lot of choices on how you might want to implement them and that would really be up to the implementer and what they want to do yes but I yes uh another angle of this is I use any open source libraries right because Java has a lot of them right and that probably would make things easier if you rely on some some tools yeah it it helps I mean I assume other language like say well I'm not a python programmer there are a lot of Lang there a lot of libraries in python as well so I I would suspect there are other language choices that would give you similar kinds of variety and so it really really come down to one's personal choice okay uh so I wonder how does this connect to another topics we disc because at the different up is is functional programming and you mentioned that you doing functional programming before basically most people doing functional programming and fair large scale it was very very large scale by today's definition some people might call functional program because it was almost like a spreadsheet where the cells could change a massive propagation so by today's definition pseudo functional I suppose uh but it was very very large so it was one of those systems where you had to learn how to come up with algorithms to allow the evaluations to scale so it was it was a good training grounds for thinking about you know what happens at scale MH and I and I I I fortunately was exposed to some people who who taught me early on to think in those terms I'm I'm very grateful for them there was a a a guy uh professor at swen called Tory re and uh he he he he taught me everything I knew about scale uh how how do you use these techniques in your topics project is it kind of inspiration do you use some of these approaches it's really a way of thinking I I think that you know most people and it it always changes from what people are working on but when they're dealing with scale you know they often are thinking of you know what would happen to something as a size changes where would the edges be I mean you know and and scenes go larger and increase by 10 times sometimes they don't need to go 100 times thinking but think it out what would occur as SCA as scale changes sometimes not much sometimes it can the scale of things can have enormous effects and I I think just asking those questions you know is important and it's important to you know data size performance maintainability uh so so I don't think there's anything I don't think it's any special in the way I think about it I think a lot of people think about in a very similar fashion but I do think that for all you know it's it's it's good to think about those things and if it's if it's a place where excuse me scale has been difficult then you kind of have extra reason to be kind of really paranoid because in in topic classification scale has been an enormous problem I mean it's just you know people have a lot of trouble getting things to scale and maintain yes so because there is web scale you know set of documents often you need to to classify yeah and in in scale even getting to 10,000 is like uh Notions of scale for topics is don't think in the billions first start thinking can you reach a 100,000 to a million and but people have a lot trouble even reaching you know a couple of thousand is really really difficult yes indeed uh so if this is a new meet up right and uh you're Veterans of software industry like you've seen different aspects so I wonder do you have an advice uh for us what you know how can we kind of connect academics with engineers and what would you personally like to to learn ATF Tex you know what kind of Spees are you looking for what would you you know what are you looking for in the community well I I think I fall I have long-term interests and short-term interests right now trying to get certain things off the ground I I'm more in in short-term mode uh topic Scout is is really not focused it's not on NLP that's not its basis but I'm interested in applying more NLP to do certain things M so I would be interested in some NLP trying to figure out what works what doesn't because like I use some of the gate stuff I ran some simple tests on it and the results were very they were not that great so for doing things like entity recognition or not it'd be interesting you know have people come in talk about things and see really what their experiences are the good side and the bad side you know so get you know if you're a practitioner you know what can you respect uh you know especially people who have looked at data and tried things for themselves not just to be you know really optimistic but really try different data sets because some of these techniques work well on some data sets but then work very poorly on others so I someone who could help share that with me that would be that would be helpful great I mean I think that's all about you know this committee is about sharing sharing experiences sharing open source which works right and kind of uh cross pollination so it's you know I think that we hope to achieve this goal and it's great to with us and we're looking forward to your talk thank you thanks