SF Text: Adithya Rao, Q&A with Alexy Khrabrov @Lithium
Recording: SF Text: Adithya Rao, Q&A with Alexy Khrabrov @Lithium
hello everybody I'm Alexi kov the organizer of SF text a new metup devoted to text mining AI search NLP and basically discovering human anten behind you know the strings of text and this is our second meet up we're here at lithium and have ad uh lead research engineer at lithium with us and uh we formerly were colleagues at Cloud so I know quite a bit about cloud and uh the topic today is topics U which is something that cloud does really well and Adia works on is going to talk about tonight so we'll have a the talk separately so I'm just going to ask a few questions to kind of get more perspective on on this whole Space yeah so it's great to have your DET with us thanks foring us uh so uh uh how did you get interested in this whole space I know that you know what you were doing at Cloud you know and you're really interested in the science of these things how they work so can you tell me a little bit about your background you know and what brought you to to these problems definitely um I think it started mainly when I started looking at uh I I come from originally from an electrical engineering background and I was looking at wireless networks as my initial field of study okay uh from wireless network Works communication was a uh kind of an easy walk through and there uh information Theory became like a very interesting subject as I started going deeper into that whole life I realized that it's probably uh easier to work with when you look at it from the higher abstraction uh and that's when I kind of moved into the domain of like machine learning in particular so uh at I did my masters from Stanford uh and there I mainly specialized in machine learning and data mining and that really got me interested into the various domains where you can apply machine learning and uh text was one of the big ones and uh that was around maybe four years ago when uh you know the social media boom was just like taking off there A lot of lot of people everybody uh getting overwhelmed with so much data around and I thought it was a per perfect opportunity to kind of look at uh uh information in social media and uh how machine learning can be applied to that aspect and uh yeah so I joined Cloud at that point and with basically there was this huge data set to work on and uh it's been pretty exciting yes yeah I know you know some of the story and that's you know it always amazes me that you know with a relatively small team Cloud was able to achieve so much right because so now as I understand it it ingests essentially most of the Social Web yeah a lot of it yes I mean uh I mean I remember you know it was quite a few itations to get there but how I mean how do you find balance basically with you know kind of a startup setting you know be able to to regularly daily you know ingest all of this data and do some science on this like how how do you balance the engineering work versus the research work right right so I think uh I think the important part is the engineering work at the end of the day because unless you're able to consume uh this kind of data and build up Frameworks that allow you to kind of iterate quickly or try to do things which where you are able to push things out fast to production and so on no matter how complex your algorithms are it's not going to amount to anything at the end of the day so I think uh that is where like a lot of effort went in in the beginning years of cloud at least and uh finally when we did have a framework where we could ingest this data and run like pipelines every day then it became easy to kind of let's say let's try try out this algorithm let's try out this model let's try out these features let's maybe vary it up a bit and then maybe the Cycles became faster and iteration became faster so I guess uh engine ing played plays a very big role when it comes to data science although uh when we talk about data science in particular we often ignore that aspect of it yes uh so I remember uh you used a lot of different tools you know we at Cloud can you talk a little bit about what kind of uh tools are using uh in your daily work sure yeah so uh I think most of the stand for machine learning at least we use most of the standard ones like uh python libraries and uh WCA and so on uh but in terms of the data pipeline we use uh main the map produ stack with the with our hdfs And Hive on top of that for our quering language Hive is actually a pretty powerful abstraction because it allows us to write like complex queries which trigger like multiple jobs without actually worrying about how how you would go about optimizing those things and the really interesting part is also that it allows us to have udfs which uh actually encapsulate a lot of the complex logic that you may not be able to write in Hive but you end up uh like let's say you wanted to apply a predictive model uh to a a row in Hive you would want to have that as a UDF that can maybe look up some data from different sources do some extraction do some NLP and then apply a model and all of that can exist in uh a UDF so I think that combination of the Hadoop stack along with the ability to write udfs and then uh training offline models using uh machine learning tools has really kind of helped us in uh Skilling up cool so you know so when you're doing it's very interesting because you're doing you know using both Python and weekend uh you know one of my huge drives in the scholar Community is to kind of bring scientific comping to the jvm right right and so there is a lot of libraries which exists already but there are some gaps compared to Python and there is just a habit of a lot of data scientists to use Python because that's usually the first language uh they pick up in in school it's also probably the most developed in terms of machine learning in uh because there there are so many tools that are present over there but yeah right so and so that you know that's my question to you is kind of practicing with both uh to tool sets what does the jvm community need to do to kind of bring the python functionality right where do you see the most gaps and where the community can contribute the most and you know what should happen so for you to stay within the jvm when you you know do your experiments right uh I think the main thing is the comprehensive nature of the tools that are present in Python I think a lot of the algorithms that are present in like the python toolkits are not always in the jvm toolkits and maybe if they are present they're not implemented in a fashion that is very uh efficient when you want to scale up so uh I think a lot of uh uh there are a lot of small things which the smaller pieces once they are there and they kind of build up like Lego bricks that's when maybe the jvm uh Community would be at that same scale uh I'm looking forward to using spark MLB I haven't used it personally yet but that's seems to be a promising area which could combine like the advantages of streaming data along with uh some of the machine learning libraries so uh I'm sure like there are a lot of avenues for uh implementing a a lot of the things that are existing in the python toolkits in into the jvm space yeah and Spark also will run Hive right because you know are committed to run everything on spark so you know there is spark cql but also Hive itself will will have spark back end right so which is which is really exciting yes yeah um interesting interesting so uh so you mentioned NLP can you talk a little bit about what kind of NLP uh are you guys doing what kind of tools are you using and what works in practice Yeah so uh I think one of the biggest uh things that we found in our uh experiments is that if you try to use like s very sophisticated NLP techniques like let's say you wanted to use Stanford named entity recognition for passing out uh entities from your text uh the what happens is that sure you could run that on a small scale but you when you want to run something on 700 million messages per day it's not going to scale very well uh so uh one of the approaches that we took was that we actually used uh open data sets such as uh Freebase Wikipedia and all of these knowledge grafts and using those knowledge crafts we try to create dictionaries that are uh and that are pre-trained and pre pre-fabricated so that they are they already encapsulate all the entities that are relevant and the advantage of doing that is also that when new entities start coming up let's say there's a completely new phenomenon on social media today that we want to incorporate in our dictionaries we are able to update our dictionaries very easily so rather than trying to extract uh semantic meaning in uh of course sometimes that is very useful but in some cases what happens is just using a simple dictionary and trying to have those dictionary terms mapping to certain entities in your topical ontology are uh is pretty effective in uh going in going maybe 90% of the way and sure at the end of the day if you really want to be like very highly accurate you have to uh get more sophisticated in terms of the NLP techniques but even doing Simple techniques at scale uh can sometimes get you a long way if you're using the correct kind of data sets and the data sources interesting so so basically you do your most of yourp offline and and you when you do training so uh when you do that can you describe what kind of Stack are you using sure uh so basically we use uh Freebase as one of our primary uh data sources so again we use Hive to kind of extract all these uh entities uh we have a bunch of tools that we have that in the udfs that do things like tokenization stemming and maybe a lot of other Transformations that so it's on jvm as as udfs in in yeah it's all jvm yes it's all in jvm and uh some of those udfs are also open- sourced in uh the project called Brick House okay so uh it's actually free to use for a lot of people so uh using so they have UDF such as uh where you can pass in a dictionary and you can pass in a piece of text and then the text would efficiently extract the certain entities out of the dictionary and so on so uh uh we have a bunch of these udfs which we then use to kind of uh uh extract entities from Freebase we use certain ranking techniques and relevance techniques to say that uh these are the most important entities because there's a lot of of noise in free Bay as well so you would probably pick the cream do some sort of ranking pick the cream maybe top half a million keywords and then map those half a million keywords into 10,000 topics because multiple keywords would map to the same topic and so on so uh mainly the stack that we use is still hi in udfs for building these models uh with a little bit of uh python to just kind of do the uh little bit whatever a little bit of machine learning training is required interesting so I mean that sounds really kind of a scalable production infrastructure and I must say that you know B's I think lead is Jerome Banks who is our common former colleague now at F we and actually we are planning to present some of the open sourcing uh stages of that project and some of scups uh coming coming forward yeah uh interesting so so given that you you you kind of set up this infrastructure you toate it and you have it in kind of nice you know ready working production order and you can iterate an experiment what are the next steps where do you see challenges for for for this project in the future yeah I think uh one of the most exciting things for me is uh actually in the domain of deep learning MH uh so deep learning is just I mean everybody's talking about it it's it's kind of the hot area of machine learning right now uh the challenge there is that if you wanted to run any algorithm on deep learning you it's I mean it's debatable whether you can do it on the cloud and on the jvm uh right now the best uh approaches to that have been to run it on uh gpus uh now if you wanted to scale gpus but now you're using something like the Hado stack then how you going to really go about it so I think uh there are certain challenges there that uh we'll probably be I mean I'll be keeping my eye out on uh in the next couple of years to see how that plans out but uh uh that is a very exciting area of research because uh if you are able to do unsupervised learning using uh deep learning on like large corpuses then uh you pretty much have like uh a huge step ahead than what everybody has is that currently on uh in in terms of production systems you know it's interesting you mention this because you know we have this uh uh conference coming up text by the bay and one of the talks they actually Dearing with deing for J on spark and that tool kit actually uses gpus right so because you you can there is a linear algebra implementation nd4j which can use Cuda and and gpus so hopefully uh we make some advances there right and I've looked into de learning for it looks very interesting Park too right so hopefully all this stuff will come together and I think actually it a great application and hopefully you know uh we'll see what's next and I mean I think the committee will be happy to to work on this stuff yeah well thanks for sharing us we're looking forward to your talk thank you thanks so much yeah all right thanks