# Using Semantic Trees In Place of Sentences | Munashe Shumba | OpenAI Scholars Demo Day 2018

## Метаданные

- **Канал:** OpenAI
- **YouTube:** https://www.youtube.com/watch?v=dFpQ830Y5qo
- **Дата:** 06.10.2018
- **Длительность:** 10:49
- **Просмотры:** 6,754
- **Источник:** https://ekstraktznaniy.ru/video/11612

## Описание

Munashe Shumba talks about Using Semantic Trees In Place of Sentences on OpenAI Scholars Demo Day on September 20, 2018. 

Learn more: https://openai.com/blog/openai-scholars-2018-final-projects#munashe

## Транскрипт

### Segment 1 (00:00 - 05:00) []

hi everyone thank you for taking time to come and see what we've been working on last a couple of months my name is Manoj Jaya and my project was on investigating the effectiveness of using somatic trees in place of regular sentences when it comes to natural language processing our problems so to get started let's take a look at some quiz questions okay I have pairs of sentences that either mean the same thing or doughnut hole or somewhere in the middle I want you to take a look at them think how closely related in their meaning they are and come up with a score from 1 to 5 if that's too hard for you could come up with a score from low medium to high alright I'm gonna give you couple seconds to think all right now I'm gonna tell you this course the first two sentences they are pretty closely related in the meaning talking about dogs fighting wrestling sort of the same thing the second one not really related except that it's somebody doing something the last one somewhere in the middle in one sentence there are kids with a football in the other there kids with that if the ball so like I say my goal is to investigate the effectiveness of using semantic trees in natural language processing instead of using a regular sentences alright so for the tests that I just gave you the quiz if one is to build a model for it you're probably lsdm s because we're talking about sentences it's sequences and it looks something like this data would flow through the Ella stem cells from the first element of the input to the end all right and then at the end of the Elysium layers you have some combination that gives you a prediction so now let's take a look at the inputs it's a sentence and represented as a sequence but is that really how we really think about sentences portray the meaning of a sentence if I were to summarize this sentence let's say maybe it's about I mean it's really about our fighting ok I'll put that down fighting but who's fighting some dogs and when is it happening it's happening now or which dogs two dots okay you can see from the top you have the most important aspect of the sentence basically the essence are the meaning and then as you go down you get more details about how this is happening the further down you get even more details about the children are far the top-level aspect same thing goes with the other sentence which was in the quiz two dogs are wrestling and hugging it's about wrestling and they's hugging happening with wrestling okay let's move on the task that I chose for this investigation was semantic relatedness so you get two sentences just like in the quiz and you have to come up with a school that saves how similar again in meaning the sentence is armed and I used it is it cold sick no idea come up with that name this dataset has 10,000 pairs of our sentences 10,000 scores to go into the the sentences and I also used a glove which is an embedding for words okay in order to convert my sentences from this data from this sick database this is sick dealer said I used synthetic net which is very well known as parsing on phosphates and the output looks something like this make some mistakes sometimes so you have to manually correct it so from there I built a model again based on SDO I'm sorry LS CMS and it looks like that same as from before just with additional details so the question now is how do I feed my trees into an LST M because remember elysium sequences they work with sentences but trees are you know not quite sequences

### Segment 2 (05:00 - 10:00) [5:00]

so what I did is was to express my trees in sequential form pretty easy I did depth-first search so the sentence on the left ribs end about that tree which is truth of the finding Simpson is from before that becomes fighting the main idea forward by the children are fighting children are dogs and are ok but dogs is a child of its own so the child appears in its onset Bart gets set of brackets right in front of dogs so I trained my models in fact I've trained two models one I used depends the trees oh by the way this fries are called dependency trees so one model I use that this depends the trees and in the other I just use regular sentences and again when I say dependency trees here are I'm talking about the sequential presentation alpha the tip it is dependent the trees from the precursor slide over here what I observed was that the model trained on dependency trees had significantly lower min squid error a surprise there 0. 35 and the one that was trying to radiances had 1. 3 so that's about three point seven times the amount of the air if you look at the scale on the horizontal axis you can see that the trees model only trained for very few number a very small number of cycles so it got to the perfect level at about 150 steps but the other model it took about one point eight so 106 million steps so significant significantly more they had about the same amount of for training loss it's just one arrived at the same loss after significantly more steps the next step from here is to instead of using a LCM cells I'm planning to use three LST ohms which is this STM that is a specifically for our trees so instead of for working on sipping seeds it works on trees I'm also going to try to use that this same idea on question answering are using the squad database sorry the squad dataset what I've noticed is that with that depends the trees it's actually really easy to manipulate them so if for instance you wanted to up meant the dealer could maybe change the order of five children in the trees and then come up with a new set of trees which you can just pile onto your data set you have more data it works in most cases there's some cases where you just can't do that I'd like to thank my mentor I was not able to come to this event my name is that yes mean she's been a great help and also I'd like to thank the open e aí team for supporting us throughout this program all right I'm gonna open out for questions so feel free yeah good that's a good question this question was how were the scores for the sentence pears are calculated wasn't actually an empirical score it was decided by people there was significant consensus but it's really not an empirical so you can really quite measure it I'm sure you could try to come up with ways of creating a criteria for you know what's a 1 and what's a 5 and wasn't middle but uh yeah it was a lot of people who curated the data said this cause and there was a significant consensus on now on the schools and you can go and take a look at this cause and I'll see how you feel about it when I was looking at the date of myself just you know taking browsing through it I sort of agreed with this course it makes sense but yeah it's not empirical yes great question the question was how did I fit in the words for the tree and what I do with the parentheses I used our glove for embedding the words and then for the parentheses because you know parentheses actually do up here in language you could have sentences that do have parentheses so I couldn't

### Segment 3 (10:00 - 10:00) [10:00]

really use parentheses so what I did is I created my own symbols and I specifically made him so that they were far away from the rest of the words I give them you know really high values for their vectors and the two symbols for the opening parentheses in the closing parentheses closer together but again they were really far away from the rest of the day just to make it create the model that this is these are not you know this is not the same kind of data all right I think that's all I have all right sorry I ran out of time but thank you so much