Download presentation
Presentation is loading. Please wait.
Published byKathleen Gardner Modified over 9 years ago
1
Learn to Comment Lance Lebanoff Mentor: Mahdi
2
Emotion classification of text In our neural network, one feature is the emotion detected in the image Generated comment should show similar emotion Studied 2 papers
3
Detecting Emotion in Text Kaitlyn Mulcrone
4
Vector Space Model Represent corpus as a term-by-document matrix Row -> word Column -> document Entry in the matrix -> td-idf score of the word for the document
5
VSM Example Document 1: You got a new phone? Document 2: Your new dog is a cute dog! Document 1Document 2... You.060 Got.060 A00 New00 Dog0.08 Your0.04...
6
Matrix Reduction Term-by-document matrix is too large and sparse Latent Semantic Analysis (LSA) using Singular Vector Decomposition (SVD) Returns an approximated matrix with smaller dimensions
7
Emotional synset vectors Synset – set of all synonyms of a given word For each emotion, retrieve the synset using WordNet lexicon Get vectors for each synonym from the term-by-document matrix Save each synset as a list of vectors Emotion (happiness) WordNet Synset joy glee delight Term-by- document matrix Vector Vector Vector Emotional synset vectors
8
Testing Given input sentence, compare the sentence to each emotional synset Choose the emotion whose synset is most similar to the sentence Sentence dog has joy Term-by- document matrix Vector Vector Vector Sentence word vectors happiness sadness fear Emotions cosine
9
Iterative method For each word in the sentence, get the word’s vector from the term-by- document matrix Compare each word in the input sentence to each word in the first emotional synset Find cosine similarity (cosine of the angle between the vectors) Take the average of the cosines
10
Iterative method joy glee delight Sentence word vectors dog has joy cosine Emotional synset vectors (happiness)
11
Iterative method joy glee delight Sentence word vectors dog has joy Emotional synset vectors (happiness) cosine
12
Iterative method joy glee delight Emotional synset vectors (happiness) Sentence word vectors dog has joy Similarity = average of all 9 of these cosines cosine
13
Combinational method Treat all of the words in the input sentence as if they were the same word for each document d in matrix for each word w in sentence vector d = ( Σ tf w,d ) * ( Σ idf w,d ) Similarly, treat all of the words in the emotional synset as the same word Find cosine similarity between the sentence vector and the emotional synset vector
14
Combinational method Sentence words dog has joy Term-by- document matrix tf, idf Sentence vector ( Σ tf ) * ( Σ idf ) Emotional synset words (for happiness) joy delight glee Term-by- document matrix tf, idf Emotional synset vector ( Σ tf ) * ( Σ idf ) cosine Similarity
15
Classification Find cosine similarity between input sentence and all emotion synsets Choose the emotion with the max value For our method, probably save all cosine similarities as a descriptor
16
Prior and contextual emotion of words in sentential context Diman Ghazi, Diana Inkpen, Stan Szpakowicz
17
Lexicons NRC-Emotion lexicon dictionary of words and their corresponding emotions 8 emotions Prior polarity lexicon polarity (positive / negative / neutral / both) type (weak / strong) 00010100 abandon angerantic.disgust fear joysadness trust surprise
18
Features For the input sentence, find words that are in the emotional lexicon Save features for each of these emotional words He was not sorrowful but instead excited. Features
19
Emotion-word features Retrieved from lexicons Emotion Polarity
20
Part of speech features Use Stanford tagger Find the part of speech of the emotional word Part of speech of 4 neighboring words
21
Dependency-tree features Use Stanford parser
22
Dependency-tree features neg – word is modified by negation word amod – word is a noun modified by an adjective, or is an adjective modifying a noun advmod – word is an adverb or head of adverbial phrase, or is modified by an adverb modifies positive – word modifies a positive word form the prior-polarity lexicon modifies-negative modified-by-positive modified-by-negative
23
Learning Input these features into SVM or Logistic Regression Logistic Regression had better results
24
Learning For sentences with multiple emotional words: For each emotional word, get probability distribution of emotional classes For each emotional class, add up the probabilities for all of the emotional words Select the class with the max probability
25
Future Work Test Stanford CoreNLP code to extract features from input sentence Organize features to be input into our neural network Study other NLP methods
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.