English Subtitles for FASTTEXT - Project Presentation at LvCS 2016



Subtitles / Closed Captions - English

the topic is called past txt use the picture right so we came across the

facebook paper was really popular in four weeks ago ago and it's about using back of tricks and makeup work tricks to better sex education so we go to speed up to the training of models projectors for sentiment analysis and they achieve by bashing and crushing my ribs so how it looks so we are creating entrance onto and and we're using so we're using prime numbers so we have a attitude to grab a prime number

and for the freedom we have two prime numbers and they didn't we describe how we get this prime number suggested some some random number and so we have to find the right number to the world connects everything the work that we have found in other texts that is being located and index which is unique and we're just multiplying it and afterwards we are taking the modulo operation and we are changing and an index which we can use in some kind of

the area so we are fixing and we're fixing on the right size to the size of the capillary and afterwards we are creating actors so we have been training and model which is consistent which is consisting of of two lands it's a simple forward propagation model and we're using back propagation to train the weights but it's not keep learning so in the project description you may have seen that we described as

sentiments and it is without the planning because we go to keep any techniques so we basically just create vectors which which has it that you want vectors where the one indicates where the hash force was activated in this field and that we tried models so we implemented different types of reckless so there's one option what we use basic one hot breakfast and you can see it here so we congratulate on the input of the

one step and content related to a huge vectors so considering that we have something like forty thousand words and our dictionary then every single worst would be 40,000 i mentioned input vector and we just concatenate them together so as you can see it's really stuff that's around five hundred thirty seconds and the accuracy awesome the parade so then we try something else what we just average the vectors so we just end

up with one huge vector consisting of of something like forty thousand dimensions and it's obviously a pasta maker serious as email and then we tried something else purchase the approach that i just presented to you it's what we call it in context tensions because we create the the two grams and grams and we get that right with us about it it's really fast with snow and the accuracy as 17 purples and secrecy is as

much higher that we tried was it if you different approaches when we combine those two techniques and here are the results are quite similar the waiting time is completely different so we can compute the actually contact sessions approach presented in the paper is is quite good and you can use it to Train faster models and if you don't trust us you can see the go completely

on nine and four questions we don't have any time but if you have any of this company of words



Video Description

This project was completed during the Lviv Data Science Summer School 2016 (http://cs.ucu.edu.ua/en/summerschool). The project goal was to speed up training models for Sentiment Analysis using hashing of n-grams.

Project presentation - http://www.slideshare.net/lvcs_ucu/fasttext