Giters

MLJejuCamp2017 / REGAN_Kyuye_Rachel

Relational GAN (REGAN) : Generating affective sentence using Relation Network

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Kyuye Song_(MENTOR : Chanwoo Lee) _Machine Learning Camp in Jeju, 2017

Relational GAN (REGAN) : Generating affective sentence using Relation Networks

Base theory : Relation Network published by DeepMind, 2017

Relational Reasoning is learning to understand relations between different objects(ideas). This is considered an essential characteristic of intelligence.To this end, Relation Networks (RNs) are proposed to solve problems hinging on inherently relational concepts. To be more specific, RN is a composite function:

The Relational Network for O (O is the set of objects you want to learn relations of) is a function fɸ. gθ is another function that takes two objects :oi , and oj. The output of gθ is the ‘relation’ that we are concerned about. Σ i,j means , calculate gθ for all possible pairs of objects, and then sum them up. where o represents inidividual object while f and g are functions dealing with relational reasoning which are implemented as MLPs.

Example : bAbi task using Relation Network

1 Mary moved to the bathroom

2 John went to the hallway.

3 Where is Mary? bathroom 1

Model : Relational GAN (REGAN) to generate affective sentnece

  • Motivation

    I think Relation Networks could analayze all kinds of things in NLP (Grammar,emotion,tone, etc.) I'd like to make dialog-bot which reflects human's emotion,tone,anything using RN&GAN

* How (REGAN MOdel)

* Expected output,goal

* Contributions

Prerequisites

import pandads
import tensorflow as tf

Git Struture

  1. DataSet : all types of dataset
  2. REGAN : code of main model
  3. cloudML : code to train in cloud ML Engine enviroment
  4. local : code to train local Engine enviroment
  5. 230 seminar ppt: Cognitive science behind ML_Kyuye 07.07.pdf
  6. README.md

Installing

  1. Dataset

    1. raw data

    • twitter dataset with emotion tag : twitter_emotion_v2(p,n,N).csv
    • twitter dataset with emotion tag : twitter_emotion_v2(p,n,N).txt

    1. targeted emotion all sentence among dataset

    • Positive emotion : Positive.tsv
    • Negative emotion : Negative.tsv
    • Neutral emotion : Neutral.tsv

    1. tageted emotion one sentence among dataset

    • Positive sentence :

      men always remember love because of romance only The best love is the kind that awaken the soul that makes us reach for more that plants the fire in our hearts and brings peace to our minds That's what I hope to give you forever The greatest happiness of life is the declaration that we are loved loved for myself or rather loved in hurt of myself The best and most beautiful things in this world cannot be seen or even heard but must be felt with the heart

    • Negative sentence :

      My sadness has become an addiction when i am not sad i feel lost I start to panic trying to find my way back which leads me back to my original state You were rarely wishing for the end of pain the monster said your own pain end to how it you It is the most human wish of all everyone in life is gonna hurt you you just have to figure out which people are worth the pain The World is mad and the people are sad The saddest thing is when you are feeling real down you look around and realize that there is no shoulder for you I guess that is what saying goodbye is always like jumping off an edge The
      worst art is making the choice to do it Once you are in the air there is nothing you can do but let go

    • Neutral sentence :

      You cannot visit the past but thanks to modern photography you can try to create it Just ask I was a student at a school and picture her travel across returned to the site exactly 30 years later The picture decided to create some of her favorite picture from back in the day I thought it would be a fun picture project for my YouTube channel tells I was amazed at how little these places had changed Before she left he finish out her old photo albums and scan favorite images Once in she successful track down the exact locations and follow her pose from 30 years previous creating new versions of her favorite she has showed the then and now picture on her YouTube

    4. train dataset - training output : train_set.txt - embeddings(key:value) : word2vec_map.json

  2. Model code

    1. model.py
    2. data_loader.py
    3. preprocessing.py
    4. ustils.py
    5. word2vec.py
[raw data]
twitter_emotion_v2(p,n,N).csv

tweet_id,Sentiment,author,content,,,,,,,
1956967341,Neg,xoshayzers,@tiffanylue i know  i was listenin to bad habit earlier and i started freakin at his part =[,,,,,,,
1956967666,Neg,wannamama,Layin n bed with a headache  ughhhh...waitin on your call...,,,,,,,
1956967696,Neg,coolfunky,Funeral ceremony...gloomy friday...,,,,,,,
1956967789,Pos,czareaquino,wants to hang out with friends SOON!,,,,,,,
1956968416,neutral,xkilljoyx,"@dannycastillo We want to trade with someone who has Houston tickets, but no one will.",,,,,,,
....

[targeted emotion all sentence among dataset]
Negative.tsv

i know i was listenin to bad habit earlier and i started freakin at his part
Layin n bed with a headache ughhhh...waitin on your call...
Funeral ceremony...gloomy friday...
Re-pinging why didn't you go to prom? BC my bf didn't like my friends
I should be sleep, but im not! thinking about an old friend who I want. but he's married now. damn, he wants me 2! scandalous!
Hmmm. is down
Charlene my love. I miss you
I'm sorry at least it's Friday?
..

[train dataset]
train_set.csv

7,7
7,51
7,22
7,4642
22,51
..

word2vec_map.json

{"clutches": [-0.40713953971862793, -0.7928721904754639, 0.5400254726409912, -0.7113916873931885, -0.2588541507720947, 0.31255054473876953, 0.3822500705718994, -0.8085153102874756, -0.25881052017211914, -0.8527586460113525, 0.8659896850585938, 0.3053250312805176, 0.06038784980773926, -0.9497530460357666, 0.4123997688293457, 0.5235145092010498, 0.7868549823760986, 0.37369585037231445, -0.8102536201477051, -0.5631704330444336, 0.7328002452850342,.....]}

Training

  1. Experiment 1 : generating sentence based on 3 emotion label(positive,negative,netural)

    • 3emo_raw.sh

  2. Experiment 2-1 : generating sentence based on 1 emotion all dataset

    1. Positive : 1emo_pos.sh
    2. Negative : 1emo_neg.sh
    3. Neutral : 1emo_neu.sh

  3. Experiment 2-2 : genetrating sentence based on 1 emotion only one sentence

    1. Positive : 1sent_pos.sh
    2. Negative : 1sent_neg.sh
    3. Neutral : 1sent_neu.sh
./1emo_pos.sh

Result

  • Example of result : Training of negative emotion dataset as a real data

  1. Experiment 1 : generating sentence based on 3 emotion label(positive,negative,netural)

  2. Experiment 2-1 : generating sentence based on 1 emotion all dataset

    1. Positive
    2. Negative
    3. Neutral

  3. Experiment 2-2 : genetrating sentence based on 1 emotion only one sentence

    1. Positive
    2. Negative
    3. Neutral
Give an example

Built With

  • Contributing

Please read READEME.mg for details on our code of conduct, and the process for submitting pull requests to us.

Authors

Acknowledgments

About

Relational GAN (REGAN) : Generating affective sentence using Relation Network

Languages

Language:Python 88.8%Language:Shell 11.2%