Automated Classification of Book
Blurbs According to the Emotional Tags
of the Social Network Zazie
V. FRANZONI, V. POGGIONI AND F. ZOLLO
DIPARTIMENTO DI MATEMATICA E INFORMATICA
UNIVERSITÀ DEGLI STUDI DI PERUGIA
Zazie
Zazie is an Italian social network for book readers that introduces a new
dimension on book characterization, the emotional icon tagging.
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Zazie was created by
Digit-Pub with Marco
Ghezzi and Barbara Sgarzi
on the model of Anobii,
with the introduction of
emotional icon tags.
lightningread again
Zazie’s Mood Icons
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Zazie’s Mood Icons
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
smile
sad
love
angry
think
cry
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
smile
sad
love
angry
think
cry
Automated classification of books
?
Always present:• Title• Author• Editor• Pages• Blurb
Which information can we use?
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
for a supervised learning approach
?
The Idea
Emotional automated classification according toZazie.
The necessity arises from the presence of a lot ofbooks that have not been tagged yet by the userswith the goal of an emotion-driven search.
Lexical analysis of the book blurbs.
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
The Approach
Correlation between the characteristics of a bookblurb and the emotional icons associated to thebook by the users.
Book blurbs can contain relevant emotionalinformation.
Blurbs are written to attract the reader, emphasizingsome book aspects with the use of emotional terms.
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
System Architecture
Zazie DB
Filtering
• Book filtering• Grouping• Tag filtering• Mood filtering
Filtered DB
Blurb analysis
• Preprocessing:- Stop words- Tokenization- Lemmatization
• Emotion extraction
DatasetClassifiers
• J48• BFTree• …
Classification Model
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Architecture of the model
Zazie DB
Filtering
• Book filtering• Grouping• Tag filtering• Mood filtering
Filtered DB
Blurb analysis
• Preprocessing:- Stop words- Tokenization- Lemmatization
• Emotion extraction
DatasetClassifiers
• J48• BFTree• …
Classification Model
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Zazie Database
38374 records:
associations of tags in the MOOD set to books.
8 fields:
(user_id, book_isbn, mood)
(book_isbn, title,pages, publisher, author,blurb)
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Architecture of the model
Zazie DB
Filtering
• Book filtering• Grouping• Tag filtering• Mood filtering
Filtered DB
Blurb analysis
• Preprocessing:- Stop words- Tokenization- Lemmatization
• Emotion extraction
DatasetClassifiers
• J48• BFTree• …
Classification Model
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Architecture of the model
Zazie DB
Filtering
• Book filtering• Grouping• Tag filtering• Mood filtering
Filtered DB
Blurb analysis
• Preprocessing:- Stop words- Tokenization- Lemmatization
• Emotion extraction
DatasetClassifiers
• J48• BFTree• …
Classification Model
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
DB filtering
Book filtering: books, most tagged by the community
Grouping: tag, count for each book (book_isbn, mood)
Tag filtering: books, predominant moods (standard
deviation)
Mood filtering: emotional moodsangry,cry,love,sad,smile,think
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Book filtering
Distribution of the records with respect to the MOODS, after the book filtering step.
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Tag filtering
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Architecture of the model
Zazie DB
Filtering
• Book filtering• Grouping• Tag filtering• Mood filtering
Filtered DB
Blurb analysis
• Preprocessing:- Stop words- Tokenization- Lemmatization
• Emotion extraction
DatasetClassifiers
• J48• BFTree• …
Classification Model
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Architecture of the model
Zazie DB
Filtering
• Book filtering• Grouping• Tag filtering• Mood filtering
Filtered DB
Blurb analysis
• Preprocessing:- Stop words- Tokenization- Lemmatization
• Emotion extraction
DatasetClassifiers
• J48• BFTree• …
Classification Model
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Filtered Database
Distribution of records, with respect to selected MOODs at the end of the filtering steps.
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
40%
37%
9%
6%
300 records
High variance
Unbalanced distribution
Preliminary dataset
A new dataset is under testing
Architecture of the model
Zazie DB
Filtering
• Book filtering• Grouping• Tag filtering• Mood filtering
Filtered DB
Blurb analysis
• Preprocessing:- Stop words- Tokenization- Lemmatization
• Emotion extraction
DatasetClassifiers
• J48• BFTree• …
Classification Model
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Architecture of the model
Zazie DB
Filtering
• Book filtering• Grouping• Tag filtering• Mood filtering
Filtered DB
Blurb analysis
• Preprocessing:- Stop words- Tokenization- Lemmatization
• Emotion extraction
DatasetClassifiers
• J48• BFTree• …
Classification Model
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Preprocessing
Normalization of the DB:
1. Stop words deletion e.g, articles and preposition
2. Tokenization ignoring punctuation marks and digits
3. Lemmatization using Morph-it! Reducing noise due to
variabilities such as singolar or plural, male or female etc.
All lemmata are kept in case of ambiguity.
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Architecture of the model
Zazie DB
Filtering
• Book filtering• Grouping• Tag filtering• Mood filtering
Filtered DB
Blurb analysis
• Preprocessing:- Stop words- Tokenization- Lemmatization
• Emotion extraction
DatasetClassifiers
• J48• BFTree• …
Classification Model
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Architecture of the model
Zazie DB
Filtering
• Book filtering• Grouping• Tag filtering• Mood filtering
Filtered DB
Blurb analysis
• Preprocessing:- Stop words- Tokenization- Lemmatization
• Emotion extraction
DatasetClassifiers
• J48• BFTree• …
Classification Model
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Emotion extraction
Synset retrieval (WordNet/MultiWordNet) for each lemma.
Exploitation of the affective domain WordNet-Affect to
associate an emotion to each synset.
Terms which don’t convey emotional information are
filtered out.
Multiple occurrences of the same emotion are counted.
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
WordNet-Affect
Emotional hierarchyof WordNet-Affect(296 nodes)
…too finely pronged!
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Emotion reduction
Two techniques were implemented:
To the third level in WordNet-Affect hierarchy
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
WordNet-Affect
Third level ofemotional hierarchyof WordNet-Affect(32 nodes)
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Emotion reduction
Two techniques were implemented:
To the third level in WordNet-Affect hierarchy
To an extended set of Ekman model of emotions:
anger, disgust, fear, happiness, sadness, surprise
+
neutral, ambiguous
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Emotion extraction
Example:
Emotions extracted from the blurb of the book «The Count of
Montecristo» by Alexandre Dumas (emotion[#occ]).
anxiety[3], enthusiasm[1], love[1], affection[1],
joy[1], negative-fear[2], general-dislike[1]
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Architecture of the model
Zazie DB
Filtering
• Book filtering• Grouping• Tag filtering• Mood filtering
Filtered DB
Blurb analysis
• Preprocessing:- Stop words- Tokenization- Lemmatization
• Emotion extraction
DatasetClassifiers
• J48• BFTree• …
Classification Model
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Architecture of the model
Zazie DB
Filtering
• Book filtering• Grouping• Tag filtering• Mood filtering
Filtered DB
Blurb analysis
• Preprocessing:- Stop words- Tokenization- Lemmatization
• Emotion extraction
DatasetClassifiers
• J48• BFTree• …
Classification Model
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Dataset
Selected features for book representation:
Author (nominal attribute)
Emotions extracted from the blurb (32 or 8 numerical
attributes)
Mood (nominal attribute): class attribute
Publisher, pages attributes were discarded as
representative of a specific edition.
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Classification model building
Multiclass classification model: each book is associated to
one of the 6 selected moods angry,cry,love,sad,smile,think
Classification models were built by means of Weka software,
using different machine learning algorithms:
Decision tree
Decision rules
Bayesian classifiers
Random forest
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Classification model evaluation
Cross validation technique with ten folds, in particular algorithms based
on decision trees, without pruning showed the best results for accuracy,
precision and recall.
Decision trees also give a more readable model.
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
TP: totally right classifications
N: #instances
NC: #classes
For each class i:TPi: true positive
FPi: false positiveFNi: false negative
Experiments
Best accuracy levels obtained with J48 and BFTree.
Classification results with respect to selectedemotional MOODs
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Conclusions
Experiments are encouraging, considering ongoing
improvements.
The blurb is confirmed to be a good source of
emotional information about a book, to be analyzed
with the aim of sentiment analysis and emotion
recognition.
Zazie provides directly a emotional model of classes:
we don’t need a manually annotation.
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Further developments
Dataset improvement in both preprocessing/filtering
(use of web-based proximity measures) and emotion
extraction with ontology-driven approach that uses
the ArsEmotica ontology
Binary classification
Feedback process from Zazie’s side
Extension to a multilabel classification
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.
Questions and comments
V. Franzoni, V. Poggioni and F. Zollo, Dipartimento di Matematica e Informatica, Università degli Studi di PerugiaAutomated Classification of Book Blurbs According to the Emotional Tags of the Social Network Zazie.