Identifying features in opinion mining via intrinsic and extrinsic domain relevance

Guided By:
Ms. Vandana Jha
Ph. D. Scholar
UVCE Bangalore
Presented By:
Gajanand Sharma
M. E. Scholar
UVCE Bangalore

 Introduction
 Related Work
 Methodology
 Algorithm
 Experiments
 Performance
 Conclusion
 Bibliography

 At present, the approaches to opinion feature extraction use pattern mining only from a single review
corpus.
 The proposed system allows to identify opinion features from online reviews by exploiting the
difference in opinion feature statistics across two corpora.
 It is done by using a measure called domain relevance. In this a list of candidate opinion features is
extracted from the domain review corpus by defining a set of syntactic dependence rules.
 For each extracted candidate feature, its intrinsic-domain relevance (IDR) and extrinsic-domain
relevance (EDR) scores are estimated on the domain-dependent and domain-independent corpora,
respectively.
 Candidate features that are less generic (EDR score less than a threshold) and more domain-specific
(IDR score greater than another threshold) are then confirmed as opinion features.

 Opinion mining (also known as sentiment analysis) aims to analyze people’s
opinions, sentiments, and attitudes toward entities such as products, services, and
their attributes.
 Consumers nowadays are no longer satisfied with just the overall opinion rating of a
product. They want to understand “why it receives the rating?”
 In opinion mining, an opinion feature, indicates an entity or an attribute of an entity
on which users express their opinions.
 Supervised learning model may be tuned to work well in a given domain.
Unsupervised natural language processing (NLP) approaches identify opinion
features by defining domain-independent syntactic templates or rules that capture the
dependence roles and local context of the feature terms.

 The domain relevance (DR) of an opinion feature across two corpora is proposed and
evaluated. The DR criterion measures how well a term is statistically associated with a
corpus.
 The method is summarized as follows: First, several syntactic dependence rules are
used to generate a list of candidate features from the given domain review corpus.
 Next, for each recognized feature candidate, its domain relevance score with respect
to the domain-specific and domain independent corpora is computed. These are
termed the intrinsic-domain relevance (IDR) score, and the extrinsic domain relevance
(EDR) score, respectively.
 Finally, candidate features with low IDR scores and high EDR scores are snipped.

 Hatzivassiloglou and Wiebe gave a supervised classification method to predict
sentence subjectivity.
 Pang proposed three machine learning methods, to classify whole movie reviews into
positive or negative sentiments.
 Pang and Lee proposed to first employ a sentence-level subjectivity detector to
identify the sentences in a document as either subjective or objective, and
subsequently discarding the objective ones.
 Bollegala proposed a cross-domain sentiment classifier using an automatically
extracted sentiment thesaurus.

 An opinion feature such as “screen” in cellphone reviews is typically domain-specific.
 So this feature appears frequently in the given review domain, and rarely outside the
domain such as in a domain-independent corpus about Culture.
 Thus, domain-specific opinion features will be mentioned more frequently in the
domain corpus of reviews, compared to a domain-independent corpus.
 From the given domain-dependent review corpus and a domain-independent corpus,
we first extract a list of candidate features from the review corpus via manually
defined syntactic rules

 Opinion features are generally nouns or noun phrases, which typically appear as the
subject or object of a review sentence.
 The subject opinion feature has a syntactic relationship of type subject-verb (SBV)
with the sentence predicate (usually adjective or verb).
 The object opinion feature has a dependence relationship of verb-object (VOB) on the
predicate.
 In addition, it also has a dependence relationship of preposition-object (POB) on the
prepositional word in the sentence.
Candidate Feature Extraction

The price of the cellphone is too expensive I like the exterior very much !!
SBV dependency relation VOB dependency relation

 From the mentioned dependence relations, i.e., SBV, VOB and POB, we present three
syntactic rules as follows-

 The candidate feature extraction process works in the following steps:
1. Dependence parsing (DP) is first employed to identify the syntactic structure of
each sentence in the given review corpus;
2. The three rules mentioned in table are applied to the identified dependence
structures, and the corresponding nouns or noun phrases are extracted as
candidate features whenever a rule is fired.

 Domain relevance characterizes how much a term is related to a particular corpus
(i.e., a domain) based on two kinds of statistics, namely, dispersion and deviation.
 Dispersion quantifies how significantly a term is mentioned across all documents
by measuring the distributional significance of the term across different
documents in the entire corpus.
 Deviation reflects how frequently a term is mentioned in a particular document by
measuring its distributional significance in the document.
 Domain Relevance is calculated by
Opinion Feature Identification

 The procedure for computing the domain relevance is summarized in this Algorithm-
Algorithm 1: Calculating Intrinsic / Extrinsic Domain Relevance (IDR/EDR)
Input: A domain specific / Independent corpus C
Output: Domain relevant scores (IDR or EDR)
for each candidate feature CFi do
for each document Dj in the corpus C do
Calculate weight Wij
Calculate standard deviation Si
Calculate dispersion dispi
for each document Dj in the corpus C do
Calculate deviation devii
Compute domain relevance dri
Return a list of domain relevance (IDR or EDR) features for all candidate features;

Algorithm 2: Identifying opinion features via IEDR
Input: Domain review corpus R and domain-independent corpus D
Output: A validated list of opinion features
Extract candidates from the review corpus R;
for each candidate feature CFi do
compute IDR score idri via algorithm 1 in review corpus R;
compute EDR score edri via algorithm 1 in domain-independent corpus D;
if (idri >= ith) AND (edri <= eth) then
confirm candidate CFi as a feature;
return a validate set of opinion features;

 IEDR performance is evaluated on two real-world review domains, cellphone and hotel reviews.
 The proposed IEDR is compared to several opponent methods as follows-
 Intrinsic-domain relevance (IDR), which uses only the given review corpus to extract opinion
features,
 Extrinsic-domain relevance (EDR), which uses only the domain-independent corpus to extract
opinion features,
 Latent Dirichlet allocation (LDA), which is a generative probabilistic graphical topic model,
 Association rule mining (ARM), which mainly discovers frequent nouns or noun phrases as opinion
features
 Mutual reinforcement clustering (MRC), and
 Dependency parsing (DP), which uses synthetic rules to extract features
Experiment Design

 IEDR feature extraction results are feed to an actual opinion mining system called
iMiner in which associated opinion words are recognized by using the IEDR
identified features.
 The evaluation results on both hotel and cellphone domain show the effectiveness
and robustness of IEDR in identifying opinion features across particular review
domains.
 Evaluation result demonstrate that the improved feature extraction via IEDR can
significantly boost the performance of feature-based opinion mining.
Feature-Based Opinion Mining Application

Choice of domain-independent corpus
IEDR performance on cellphone reviews versus choice of domain-independent corpus/topic.
(Topics are ranked in descending order of F-measure)

Choice of domain-independent corpus
IEDR performance on hotel reviews versus choice of domain independent corpus/topic.
(Topics are ranked in descending order of F-measure.)

 A novel inter-corpus statistics approach to opinion feature extraction based on the
IEDR feature-filtering criterion is proposed, which utilizes the disparities in
distributional characteristics of features across two corpora, one domain-specific and
one domain-independent.
 IEDR identifies candidate features that are specific to the given review domain and
yet not overly generic (domain-independent).
 The influence of corpus size and topic selection on feature extraction performance is
evaluated.

[1] Zhen Hai, Kuiyu Chang, Jung-Jae Kim, and Christopher C. Yang, “Identifying Features in Opinion
Mining via Intrinsic and Extrinsic Domain Relevance”.
[2] V. Hatzivassiloglou and J.M. Wiebe, “Effects of Adjective Orientation and Gradability on Sentence
Subjectivity,” Proc. 18th Conf. Computational Linguistics, pp. 299-305, 2000.
[3] B. Pang, L. Lee, and S. Vaithyanathan, “Thumbs up?: Sentiment Classification Using Machine
Learning Techniques,” Proc. Conf. Empirical Methods in Natural Language Processing, pp. 79-86, 2002.
[4] B. Pang and L. Lee, “A Sentimental Education: Sentiment Analysis Using Subjectivity
Summarization Based on Minimum Cuts,” Proc. 42nd Ann. Meeting on Assoc. for Computational
Linguistics, 2004.
[5] R. Mcdonald, K. Hannan, T. Neylon, M. Wells, and J. Reynar, “Structured Models for Fine-to-Coarse
Sentiment Analysis,” Proc. 45th Ann. Meeting of the Assoc. of Computational Linguistics, pp. 432439,
2007.

[6] D. Bollegala, D. Weir, and J. Carroll, “Cross-Domain Sentiment Classification Using a Sentiment
Sensitive Thesaurus,” IEEE Trans. Knowledge and Data Eng., vol. 25, no. 8, pp. 1719-1731, Aug. 2013.
[7] C. Zhang, D. Zeng, J. Li, F.-Y. Wang, and W. Zuo, “Sentiment Analysis of Chinese Documents: From
Sentence to Document Level,” J. Am. Soc. Information Science and Technology, vol. 60, no. 12, pp.
2474-2487, Dec. 2009.

Identifying features in opinion mining via intrinsic and extrinsic domain relevance

Identifying features in opinion mining via intrinsic and extrinsic domain relevance

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Viewers also liked

Viewers also liked (14)

Similar to Identifying features in opinion mining via intrinsic and extrinsic domain relevance

Similar to Identifying features in opinion mining via intrinsic and extrinsic domain relevance (20)

Recently uploaded

Recently uploaded (20)

Identifying features in opinion mining via intrinsic and extrinsic domain relevance