Master Thesis Slide

Dipartimento di Informatica

Università degli studi di Salerno, Fisciano

Capturing Emotions with Deep Learning

Jacek Filipczuk

25 February 2016

1

Content

IntroductionProblem Definition

The goalEmotionsEmotion ModelsDeep Learning Neural NetworksMLPCNNLSTM

PreprocessingDatasetsFeature CreationNLTKFeature ExtractionLabeling

Training and TestingConclusions

Jacek Filipczuk | Emotions with Deep Learning

2

Index






3

Emotion Extraction from Text

I Why is it important?I Is it useful?I How can it be done?


4

Problem Definition

Figure: A positive review


5

Problem Definition

Figure: A negative review


6

Index






7

Sentiment Analysis

Sentiment Analysis is the process of determining the emotional tonebehind a text written by some user.It helps to understand the attitudes, opinions and emotions expressedwith an online method.


8

Why emotions?

Social media sentiment analysis can be an excellent source ofinformation and can provide insights that can:

I determine market strategy;I improve campaign success;I improve product messaging;I improve customer service;I test business performance.


9

Different Models

Figure: Different Emotion Models.


10

Plutchik

Figure: Plutchik Emotion Model.Jacek Filipczuk | Emotions with Deep Learning

11

The hourglass

Figure: The Hourglass of Emotions ModelJacek Filipczuk | Emotions with Deep Learning

12

The emotions

The target emotions where inspired from the previous models.


13

Deep Neural Networks

The process of hierarchical feature learning is divided in two steps.First multiple layers of non-linear features are extracted. Second thelayers are passed to a classifier that combines all the features tomake predictions.Deep Neural Networks were inspired by the human brain.


14

A Neural Network

Figure: The architecture of the first known deep network trained by AlexeyGrigorevich Ivakhnenko in 1965.


15

MLP architecture

Figure: The architecture of a generic MLP neural Network.


16

Convolutional Neural Network

The CNN take into account the spatial structure of the dataset.


17

Convolutional Neural Network

Figure: An example of a CNN.


18

LSTM Neural Network

Long short-term memory networks are a special kind of Recurrentneural networks capable of learning long-term dependencies.LSTMs are explicitly designed to remember information for longperiods of time.


19

The architecture

Figure: An example of the repeating module in an LSTM neural network.


20

The cell

Figure: An example of the cell.


21

Index






22

The datasets

In this work were used two different datasets:I the IMDB dataset;I the ISEAR dataset.


23

IMDB

I Who created the dataset? —> UsersI What are the data inside? —> ReviewsI How big it is? —> 25 000


24

ISEAR

I Who created the dataset? —> PsychologystsI What are the data inside? –>Short sentencesI How big it is? –> 7500


25

Natural Language problems

Natural language processing problems are directly caused bylanguage problems. In fact there are some major language problemslike:

I Languages are changing everyday, new words, new rules, etc.I Words can have different meanings depending on context, and

they can acquire new meanings over time (apple [a fruit], Apple[a company], they can even change their parts of speech(Google –> to google).

I Computers do not understand words. They do not understandtheir true meanings as humans do.


26

Natural Language Toolkit

A general NLP pipeline:I Tokenization;I Tagging;I Chunking.


27

Chunking grammar

grammar = "NP:<DT>?<JJ>*<NN>"

Figure: Segmentation and Labeling at both the Token and Chunk Levels.


28

Building the dataset

Adverbs + Adjectives = Emotion Score

Feature[i] += 2+coeff

where coeff=+-1


29

Labeling the Dataset

The target emotion is the one with the greatest score.I ExpandedI Compressed


30

Index






31

The goal

Train and test with different Deep Neural Networks!


32

MLP Testing

Table: 16x7586 ISEAR

Accuracy Loss Function Learning Rate Epochs40,93% 1,594191 0,01 30043,07% 1,577638 0,01 30043,60% 1,531735 0,01 30043,27% 1,519117 0,001 30044,53% 1,483311 0,001 30042,87% 1,535861 0,001 30033,93% 1,821047 0,0001 30035,33% 1,826459 0,0001 30036,36% 1,803543 0,0001 300


33

MLP Testing


Accuracy Loss Function Learning Rate Epochs Accuracy Loss Function Epochs40,13% 1,592692 0,01 300 43,51% 1,531616 50042,67% 1,576612 0,01 300 43,93% 1,571319 50042,91% 1,531140 0,01 300 41,79% 1,583024 50042,71% 1,529107 0,001 300 44,69% 1,521561 50044,31% 1,482301 0,001 300 44,74% 1,433542 50043,97% 1,515163 0,001 300 42,82% 1,492315 50034,33% 1,721917 0,0001 300 37,96% 1,798342 50035,53% 1,821449 0,0001 300 37,12% 1,802671 50034,76% 1,803183 0,0001 300 35,61% 1,828191 500


34

MLP Testing




35

ISEAR results


36

MLP Testing

Table: 16x2156 IMDB



37

MLP Testing

Table: 116x2156 IMDB

Accuracy Loss Function Learning Rate Epochs Accuracy Loss Function Epochs89.00% 0.394762 0,01 300 88.25% 0.409069 50089.75% 0.391763 0,01 300 88.50% 0.432282 50089.50% 0.404379 0,01 300 86.00% 0.430368 50090.00% 0.389020 0,001 300 91.50% 0.364005 50087.00% 0.510808 0,001 300 90.50% 0.416307 50087.50% 0.492533 0,001 300 89.75% 0.407521 50082.00% 1.174111 0,0001 300 83.12% 1.162348 50077.75% 1.247049 0,0001 300 79.32% 1.152141 50075.50% 1.235449 0,0001 300 80.43% 1.167145 500


38

MLP Testing


Accuracy Loss Function Learning Rate Epochs91.75% 0.388250 0,01 30092.50% 0.366930 0,01 30093.00% 0.346140 0,01 30092,25% 0,443516 0,001 30090,00% 0.475815 0,001 30087.25% 0.598331 0,001 30077.75% 1.151705 0,0001 30081.75% 1.058237 0,0001 30078.50% 1.200452 0,0001 300


39

MLP Testing


Accuracy Loss Function Learning Rate Epochs96.25% 0.277372 0,01 30094.45% 0.335562 0,01 30095.84% 0.289589 0,01 30095,55% 0,316808 0,001 30095,30% 0,323706 0,001 30096,25% 0,297388 0,001 30087.95% 0.594173 0,0001 30087.60% 0.604254 0,0001 30088.30% 0.567574 0,0001 300


40

IMDB results


41

CNN Testing

Table: Results of the test with Convolutional Neural Network.

Dataset Accuracy Loss Function Learning Rate EpochsIMDB 96,65% 0,113728 0,001 300ISEAR 36,21% 1,689331 0,001 300


42

LSTM Testing

Table: Results of the test with Long Short-Term Neural Network.

Dataset Accuracy Loss Function Learning Rate EpochsIMDB 93,20% 0,28546 0,001 300ISEAR 41,00% 1,586147 0,001 300


43

Comparing results


44

The big difference


45

Index






46

Final thoughts

In conclusion deep neural networks are a really powerful tool that canachieve incredibly good results even for difficult problems likecapturing emotion.


47

Future improvements

In future it will be interesting to improve the emotion labeling steppresented in this work and experiment more with deep neuralnetworks, adding more layers or trying different type of layers.


Master Thesis Slide

Documents