a fuzzy approach for multi-domain sentiment analysis

24
A Fuzzy Approach For Multi- Domain Sentiment Analysis Mauro Dragoni Fondazione Bruno Kessler (FBK), Shape and Evolve Living Knowledge Unit (SHELL) https://shell.fbk.eu/index.php/Mauro_Dragoni - [email protected] work done in collaboration with Prof. Andrea G.B. Tettamanzi and Prof. Celia da Costa Pereira INRIA Sophia Antipolis June, 19 th 2014

Upload: mauro-dragoni

Post on 14-Dec-2014

149 views

Category:

Technology


4 download

DESCRIPTION

An emerging field within Sentiment Analysis concerns the investigation about how sentiment polarities towards concepts have to be adapted with respect to the different domains in which they are used. In this paper, we explore the use of fuzzy logic for modeling concept polarities, and the uncertainty associated with them, with respect to different domains. The approach is based on the use of a knowledge graph built by combining two linguistic resources, namely WordNet and SenticNet. Such a knowledge graph is then exploited by a graph-propagation algorithm that propagates sentiment information learned from labeled datasets. The system implementing the proposed approach has been evaluated on the Blitzer dataset by demonstrating its viability in real-world cases.

TRANSCRIPT

Page 1: A Fuzzy Approach For Multi-Domain Sentiment Analysis

A Fuzzy Approach For Multi-Domain Sentiment Analysis

Mauro Dragoni

Fondazione Bruno Kessler (FBK), Shape and Evolve Living Knowledge Unit (SHELL)

https://shell.fbk.eu/index.php/Mauro_Dragoni - [email protected]

work done in collaboration with Prof. Andrea G.B. Tettamanzi and Prof. Celia da Costa Pereira

INRIA Sophia Antipolis June, 19th 2014

Page 2: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Outline

1. Background on Sentiment Analysis and Fuzzy Logic

2. Motivations

3. The Approach

4. Evaluation of the Implemented System

Page 3: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Sentiment Analysis - 1

Natural Language Processing task for identifying the opinion given by someone with respect to something.

Opinions may be positive, negative, or neutral.

The value associated with the opinion is called “polarity”.

Page 4: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Sentiment Analysis - 2

Basic challenges: Identification of the polarities for each term in the text.

Deciding how to aggregate the different polarities.

Advanced challenges: Identification of the entities in each sentence (subjects).

Identification of the features describing each entity.

Adaptation of the sentiment model to different domain.

Manage the uncertainty of each learned information within the single domain

Page 5: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Fuzzy Logic

Allows to represent imprecise information.

With respect to classical logic, truth-values of assertions may assume all values in the interval [0, 1]

The main element of the fuzzy logic are Fuzzy Sets

Hot temperature.

x

y

Page 6: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Motivations - 1

The same concept may have different polarities in different domains.

The polarity associating a concept to a domain may be uncertain due to the different contexts in which it is used.

Page 7: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Motivations - 2

The assignment of a unique polarity value to the entire text leads to imprecise information.

In the same text, different aspects have to be analyzed.

A significant concept extraction capability is required.

“I bought a new smartphone: the screen is awesome, even if some

colors are not very brilliant, but the battery is too short”

Page 8: A Fuzzy Approach For Multi-Domain Sentiment Analysis

The Approach

Creation of the knowledge base.

Concept extraction.

Learning of the preliminary sentiment information.

Propagation of the learned information through the knowledge graph.

Modeling of the fuzzy shapes.

Page 9: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Creation of the Knowledge Base

Based on the integration of WordNet with SenticNet

WordNet has been enriched with terms extracted from the Roget’s Thesaurus

The links between WordNet and SenticNet have been built by taking into account the synonyms of each WordNet synset and the synonyms of each SenticNet concept.

In order to avoid ambiguities not all associations have been created.

Example: concept “base”

WordNet: 20 senses (for the noun)

SenticNet: base (beneath, below, understructure) WordNet sense 2

Page 10: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Concept Extraction - 1

Two samples:

1. Today I went to the mall and bought some desserts and a lot of very nice Christmas gifts.

2. The touchscreen is awesome but the battery is too short.

Page 11: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Concept Extraction - 2

Page 12: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Multi-Domain Fuzzy Propagation - 1

Polarity information is propagated through the knowledge base by using an algorithm implementing the simulated annealing strategy.

The propagation of the values is driven by three parameters: annealing rate, propagation rate, and convergence limit.

The intermediate polarity values measured on each concept at the end of each iteration are stored in order to build the final fuzzy shape associated with each combination concept-domain.

A different model is learned for each domain.

Page 13: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Multi-Domain Fuzzy Propagation - 2

Domain Initialization:

Information Propagation:

Stop Condition:

Page 14: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Multi-Domain Fuzzy Propagation - 3

Trade-offs:

high propagation rate risk of polarity convergence

low convergence limit risk of polarity convergence

lower propagation rate incomplete graph

high annealing rate premature stop of the algorithm

high convergence limit premature stop of the algorithm

Page 15: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Multi-Domain Fuzzy Propagation - 4

Page 16: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Modeling of Fuzzy Shapes - 1

Value computed from the training set.

Value obtained after the propagation phase.

Support computed based on the variance value.

Page 17: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Modeling of Fuzzy Shapes - 2

Type 1 level of uncertainty: the core of the fuzzy trapezoid crosses the neutral polarity

Type 2 level of uncertainty: only the support of the fuzzy trapezoid crosses the neutral polarity

Page 18: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Evaluation of the System - 1

Evaluation on the Blitzer dataset: 25 domains

~3000 reviews for each domain in the balanced dataset

75% of instances for the training, 25% for the validation

Three baselines: SVM, Max-Entropy, and Naïve-Bayes

Compared the performance by discarding the different levels of uncertainty

Evaluation on: Elementary Polarity Computation

Concept Extraction + Polarity Computation

Page 19: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Evaluation of the System - 2

How fuzzy polarities are aggregated?

x

Page 20: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Evaluation of the System - 3

Page 21: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Evaluation of the System - 4

Approach Avg. Precision Avg. Recall

SVM 0.8068 1.0

Naïve-Bayes 0.8227 1.0

Max-Entropy 0.8225 1.0

MDFSA 0.8617 ~ 1.0

MDFSA (Type 1 Uncertainty excluded)

0.8735 ~ 0.7

MDFSA(Type 1 & 2 Uncertainty excluded)

0.8991 ~ 0.5

Elementary Polarity Computation:

Page 22: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Evaluation of the System - 5

Approach Precision Recall F-Measure

MDFSA 0.25 0.26 0.25

IBM 0.24 0.14 0.18

UNI-NEGEV 0.12 0.05 0.07

Concept Extraction + Polarity Computation (ESWC 2014 Challenge):

Page 23: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Future Work

Integration of more knowledge bases into the system.

Improve how ambiguities are addressed.

Improve the concept extraction module.

Extending the approach for addressing multilingualism.

Apply the approach to the social network environment.

Page 24: A Fuzzy Approach For Multi-Domain Sentiment Analysis

Mauro Dragonihttps://shell.fbk.eu/index.php/Mauro_Dragoni [email protected]