Extracting Implicit Aspects based on Latent Dirichlet Allocation

Ekin Ekinci and Sevinç İlhan Omurca

Computer Engineering Department, Kocaeli University, İzmit, Kocaeli, Turkey

{ekin.ekinci, silhan}@kocaeli.edu.tr

1 RESEARCH PROBLEM

“What others think?” is the central to almost all

human activities and is the key influencer of human

behaviors (Fersini et al, 2016). For this reason, when

people need to make a decision about which film is

good or companies need to evaluate their products’

weakness, they take into consideration the opinions

of others.

With the rapid advances of the Internet, people

have dramatically changed their way to gather

others’ opinion through the online review websites.

These mediums have showed themselves as one of

most popular channels with millions of users who

share their opinions about they have purchased. Due

to increase in user generated content on online

review sites an important soure for data analysis can

be provided. According to a survey conducted about

50% of the young population have been affected by

these word of mouth data (Ha et al, 2015). So, for

people there is no need to ask someone else about

product they want to purchase and for companies

there is no need to conduct surveys for their brands

(Pawar et al, 2016). Beside all of them, for the same

product there are huge amount of reviews on

different web sites and to get the information they

are interested in is very difficult and time consuming

for average humans. Therefore, an automatic system

is needed.

Sentiment analysis, which has been studied on

since 1990’s and in recent years has increasingly

gained importance, is a study that analyses people’s

opinions and feelings toward products, companies,

services and so on (Türkmen, 2016; Liu, 2012).

Despite this short history, analyzing people's opinion

and sentiment has been attracted to researchers and

since then, many studies have been done. In these

studies, sentiment analysis is realized at one of the

three levels: document level, sentence level and

aspect level sentiment analysis (Vinodhini and

Chandrasekaran, 2012).

 Document Level: Essentially, with this level

of analysis, the task of determining whether

the given review is positive or negative. For

instance given a restaurant review, the task is

to learn general sentiment (good or bad) about

this restaurant by using whole document.

 Sentence Level: Sentences are just short

documents so there is in fact no key difference

between sentence level and document level

sentiment analysis. In this analysis the task is

to learn general sentiment from every sentence

separately.

 Aspect Level: As mentioned above, document

and sentence level analysis give general

sentiments about a product. Along with it is

not correct to assume that if the result is

negative; all the product specifications are

weak. Accordingly it doesn't mean that all the

product specifications are fine if the result is

positive. So there is a need for finer-grained

analysis. Aspect, which expresses sentiment,

is anything that defines, completes a product;

sentiment is positive or negative feeling about

the aspect (Türkmen et al, 2016). In aspect

level sentiment analysis sentiments are given

separately for every aspect. For example given

a restaurant review, instead of learning general

sentiment for this restaurant, learning

sentiments for aspects such as service, meal,

etc. is proposed.

In sentiment analysis aspects are categorized as

two types: explicit aspects and implicit aspects. If an

aspect or its alternatives appears in a review this

aspect is called explicit aspect. Conversely, if an

aspect does not appear in the review but is implied

by a sentiment this aspect is implicit. In the example

below, the first sentence contains an explicit aspect

“tuna”. The second sentence the implicit aspect

“price” is implied by using “cheap”.

Example:

 The tuna is also very fresh, not so chewy.

 It's fairly cheap (for being so trendy).

For extracting explicit aspects there are four

main methods tried in the literature: frequent noun

and noun phrases based methods, rule based

methods, supervised learning and topic models. On

the other hand, implicit aspect extraction is tough

task and many studies in the literarure ignore this

phase. Regarding to aforementioned matters, we

Ekinci, E. and Omurca, S.

Extracting Implicit Aspects based on Latent Dirichlet Allocation.

In Doctoral Consortium (DCAART 2017), pages 17-23

propose a implicit aspect extraction framework by

using semantic similarity based topic model.

2 OUTLINE OF OBJECTIVES

The aspect based sentiment analysis studies are

fundamentally focused on product aspects. Although

in previous studies explicit aspect extraction has

been studied intensely, on the contrary implicit

aspect extraction has been studied limitedly.

However implicit aspect extraction is the one of the

important phase of the sentiment analysis studies for

the following reasons. The sentence which only

contains sentiment and aspect is unknown is not

useful. More importantly, when the reviews are

examined it can be clearly seen that number of

implicit aspects in the sentences are enough to be

considered for better sentiment analysis (Xu et al,

2015).

Probabilistic topic models are based on the idea

that documents are mixtures of topics and a topic has

a probability distribution over words (Stevyers and

Griffiths, 2007). Actually, probabilistic topic model

methods are defined as a group of algorithms that

discover hidden thematic structure of document

collections by mapping them into low dimensional

space (Boyd-Graber and Blei, 2009). And a topic

can be defined as a collection of words that

frequently occur together and are related to the same

subject. Latent Dirichlet Allocation (LDA), which is

one of the simplest topic model, is emerging field in

machine learning and text mining (Blei et al, 2004;

Mei et al, 2007). The intuition behind LDA is that

documents exhibit multiple topics. LDA, as a

completely unsupervised method, is based on bag-

of-words assumption. And the LDA model does not

consider the semantic structure of the documents.

This lack motivates our research.

The effectiveness and better generalization of the

proposed framework we propose an unsupervised

approach for implicit aspect extraction that is

focused on the use of semantic similarity of

documents for topic proportions and topic

assignment. Furthermore, to encourage this Babelfy

which carries out both multilingual word sense

disambiguation and entity linking is used (Moro et

al, 2014). Babelfy is based on the Babelnet, which is

a integration of Wikipedia and WordNet, a

multilingual semantic network (Navigli and

Ponzetto, 2012). Babelfy is preferred for the concept

extraction and by using these concepts semantic

similarity of documents are calculated. The proposed

framework will be tested on reviews of restaurants;

also Jo and Oh (2011) used these reviews in English

for ASUM.

It is planned to achieve the following objectives

with our framework enabling aspect based sentiment

analysis based on semantic similarity based topic

model for implicit aspect extraction:

 Obtaining the reviews of restaurants and pre-

processing these reviews using Stanford

Natural Language Processing Tool,

 Determining noun phrases for extracting

multi-word aspects from reviews by using

Babelfy,

 Deciding parameters of LDA, applying the

model to the reviews and extracting product

aspects,

 Evaluating generalization performance of the

model on the test data,

 Extracting concepts by using Babelfy and

expanding reviews with these concepts,

 Calculating semantic similarity of reviews and

using these obtained values for topic

proportions and topic assignment in LDA,

 Evaluating generalization performance of the

new model on the test data and comparing

with LDA

 Extracting aspect sentiment pairs from

reviews,

 Extracting implicit aspects by using aspect

sentiment pairs,

 The efficiency of the proposed system will be

examined.

3 STATE OF THE ART

In recent years, aspect based sentiment analysis

studies have attracted more and more attention

because for the first time in human history, a huge

volume of opinionated data is obtained with new

resources that have millions of users such as

ecommerce and social media websites, blogs,

dictionaries, news portals (Liu, 2012).

When the literature is evaluated it is shown that

aspect extraction is one of the cornerstones of the

sentiment analysis studies. To design a powerful

sentiment analysis system, aspect extraction process

should be carried out successfully (Ekinci et al,

2016). The first studies about this topic were made

by Hu and Liu (2004). In their studies differences

between explicit and implicit aspects were explained

and only explicit aspect extraction was performed.

Explicit aspects which are noun or noun phrases

were extracted by using association rule mining.

Popescu and Etzioni (2005) developed OPINE an

DCAART 2017 - Doctoral Consortium on Agents and Artiﬁcial Intelligence

unsupervised information extraction system to

extract explicit aspects. They benefited from

Pointwise Mutual Information for pruning aspects.

Hu and Liu’s approach was further improved by Wei

et al (2010) incorporating Semantic Based

Refinement. The approach aims to use sentiments

for successfully extracting explicit aspects. They

used General Inquirer and Co-occurrence-based

pruning, Opinion-based infrequent feature

identification, Conjunction-based infrequent feature

identification rules for aspect extraction and pruning.

Brody and Elhadad (2010) devised an

unsupervised method, called Local LDA, for aspect

extraction. For each aspect representative words

were found by using Mutual Information. For

example representative words for “meal” are “menu,

fish, cuisine, and so on”. Conjunctions and negations

were preferred for adjective extraction and they

benefited from Conjunction Graph for determining

polarities of adjectives. Li et al (2010) proposed two

new methods, called Sentiment-LDA and

Dependency-Sentiment-LDA. Sentiment-LDA is

based on the idea that sentiments are related to topic.

In Dependency-Sentiment-LDA, they integrated

sentiment dependency to the topic model. Wang et al

(2010) developed semi-supervised topic model Co-

LDA. In Co-LDA, aspects and sentiments were

modeled simultaneously. For this purpose the model

is divided into two parts; sentiment LDA and topic

LDA. Jo and Oh (2011) assumed that words in the

same sentence are under the same topic with

Sentence LDA. They after developed an advanced

version of Sentence LDA, called Aspect and

Sentiment Unification Model (ASUM). With ASUM

aspects and sentiments were modeled together and

aspect sentiment pairs were obtained. Xianghua et al

(2013) utilized LDA to extract global topics for

reviews in Chinese. They also utilized sliding

window for local topics. For sentiment polarity they

used Hownet lexicon. Ding et al (2013) composed

Hierarchical Dirichlet Process-LDA (HDP-LDA).

HDP-LDA differed from the LDA by automatic

determination of topic counts. For determining

sentiments they utilized lexicon. Bagheri et al (2013)

devised Aspect Detection Model based LDA (ADM-

LDA) which ignored bag of words and was based on

Markov Chain. Wang et al (2014) proposed two new

semi-supervised methods, called Fine-grained Label

LDA (FL-LDA) and Unified Fine-grained Label

LDA (UFL-LDA). FL-LDA utilized seed lexicon for

aspects to extract aspects in the reviews. In UFL-

LDA unlabeled documents were considered for high

frequency aspects. Zheng et al (2014) devised

Appraisal Expression Patterns LDA (AEP-LDA) for

extracting product aspects from restaurant, hotel,

MP3 player and camera reviews. The basic idea

behind this model was that words in the same

sentence were under the same topic. Aspects and

sentiments were extracted simultaneously in this

model. Like Bagheri et al (2013), Yin et al (2014)

ignored bag of words in their LDA based approach,

called Dependency-Topic-Affects-Sentiment-LDA

(DTAS). Instead of bag of words they preferred

Markov Chain. They assumed that sentiments in a

sentence affected by sentence that contains sentence

and the previous sentence.

In our thesis, we aim to extract implicit aspects

so we also delve into literature in this subject area.

Su et al (2008) used Mutual Reinforcement to

expose hidden relation between aspect categories

and group of sentiments. The hidden relation was

presented by using bipartite graph. It was enough to

create a connection between aspect and sentiment to

be in the same sentence. The connection weight was

determined by the total number of co-occurrence of

them. Zhang et al (2012) preferred statistical

methods for implicit aspects. They used PMI and

frequency based collocation selection method for

this purpose. Wang et al (2013) benefited from

association rules for extraction of implicit aspects.

By using five different association rules for aspects

and sentiments new rules were extracted. From these

rules implicit aspects were extracted by using

frequency and PMI. Bagheri et al (2013) proposed

graph-based scoring for implicit aspects. Relation

between explicit aspects and sentiments was

demonstrated with this graph. Xueke (2013) devised

Joint Aspect/Sentiment Model (JAS) to remove

deficiencies of LDA. By using aspect based

sentiments were used for extraction of implicit

aspects. Lau et al (2014) proposed LDA based fuzzy

product ontologies for aspect based sentiment

analysis. Both taxonomic (memory is a hardware)

and non-taxonomic (bright flash) relations were

extracted with this method. For non-taxonomic

relations in sentences they benefited from Mutual

Information. Poria et al (2014) presented rule based

method for implicit aspects. They used explicit

aspects and implicit aspect clues (IACs) for

extraction of these aspects. IACs were adjectives

and mapped to associated aspects category. Xu et al

(2015) devised LDA based Explicit Topic Model,

which was semi-supervised, for implicit aspects. The

obtained results from this model was used in

Support Vector Machines to extract implicit aspects

from sentences.

Extracting Implicit Aspects based on Latent Dirichlet Allocation

4 METHODOLOGY

The thesis consists of ten steps as mentioned in the

outline of the objectives section and is in the fourth

step right now. The steps are explained in detailed

and results are given for every steps below.

4.1 Dataset and Preprocessing

Reviews concerning 320 different restaurants from 4

different cities (Atlanta, Chicago, Los Angeles and

New York City) are obtained from a web site

. This

data set was used in Jo and Oh's (2011) study. The

whole dataset consists of 25459 reviews in English

but only 2647 reviews are used for this study. The

summary of dataset depicted in Table 1.

Table 1: The summary of dataset.

Domain Number of

reviews

Average

sentence

count

Average

word count

Restaurant 2647 12 187

After the dataset is obtained, for each of the word

spell correction is applied. Spell correction is an

important phase in sentiment analysis of word-of-

mouth data and then stop words are eliminated.

Stemming is also a crucial step. Stemming is used

for reduces the different form of word to a single

form. As the final phase of preprocessing, POS

tagging is performed to assign parts of speech to

token such as noun, adjective, verb, etc. For

preprocessing steps Stanford NLP tool is used.

There are 8603 different words in the reviews.

Figure 1 is a part of an actual review and Figure 2 is

the preprocessed version of this review.

Had brunch with the girls today and we ordered

the bottomless bubbly, mac and cheese, crab

cakes benedict, penne, grilled cheeseee with

tomato sauce and breakfast panini...

Figure 1: The actual review.

Figure 2: The preprocessed review.

http://uilab.kaist.ac.kr/research/WSDM11

4.2 Multi-word Aspects

In the reviews, some aspects have more than one

single word. Aspect extraction is very important

sentiment analysis, therefore, for an overall

sentiment analysis of the reviews, multi-word

aspects of the products must be detected. For this

purpose, Babelfy, which is unified graph for word

sense disambiguation and entity linking, is used

(Moro et al, 2014). From reviews 2640 different

multiword aspects are obtained such as, chicken

salad, burgundy wine, flat iron steak and so on.

4.3 Latent Dirichlet Allocation

LDA is described as generative probabilistic model

for collections of discrete data such as text corpora

by Blei et al (2004). In LDA, generative model,

which is a simple probabilistic procedure, specifies

document creation by using latent variables

(Stevyers and Griffiths, 2007; Jadhav 2014). With

latent, it is desired to mean learning the meaning of

the document by discovering latent topics (Mei et al,

2007). The basic intuition behind LDA is that

documents exhibit multiple topics and topic has a

probability distribution over words. LDA is

completely unsupervised and is based on bag-of-

words assumption. The generative model and

posterior distribution of LDA is shown in Figure 3.

Figure 3: Generative model for LDA.

Each documents are mixtures of topics and words in

these documents are chosen one of these topics. And

each topic has a distribution over words from a fixed

vocabulary. Distribution over words and topic

proportions are obtained with Dirichlet distribution.

Dirichlet distribution is a the conjugate prior for the

parameters of the multinomial distribution (Bishop,

2016).

The graphical model for LDA is given by using

plate notation and is represented in Figure 4.

brunch girl today order bottomless bubbly

mac cheese crab cake benedict penne

grilled cheese tomato sauce breakfast

panini

DCAART 2017 - Doctoral Consortium on Agents and Artiﬁcial Intelligence

Figure 4: Graphical model for LDA.

In Figure 4 each nodes are random variables and

directed edges are used to explain how these random

variables are generated along with these edges. The

observed node is shaded (words in the document)

and hidden nodes are unshaded. In this model

the total number of documents,

is the total

number of topics. Number of words in document

is represented with

N .

and

are Dirichlet

parameters.

is topic proportion in the documents

and

is distribution over words. According to

graphical model the joint probability of observed

and hidden random variables is given with Equation

() ()

()( )





































∏∏∏

===

knmnm

mnm

zwpzppp

ϕθαθβϕ

,||||

(1)

The main purpose of GDA is to obtain model

parameters and for model parameters posterior

distribution in Equation 2 is used.

()

MMMK

wzp

:1:1:1:1

,,,

|,,

θϕ

(2)

Collapsed Gibbs Sampling algorithm is preferred

for this posterior distribution.

4.4 Experimental Results

To apply LDA to reviews of restaurants firstly

model parameters have to be specified. For

K/50=

and

01.0=

, which are recommended

by Stevyers and Griffiths (2007), values are usually

used. The number of topics is determined

100=K and 1000 iteration of the Collapsed

Gibbs sampling algorithm have been performed. The

extracted product aspects are given in Table 2.

Table 2: The extracted product aspects.

Breakfast Dessert Drink Salad

toast dessert drink salad

breakfast chocolate bar goat cheese

egg cream cocktail vinaigrette

banana cookie bartender chicken

coffee ice cream martini chicken salad

pancake vanilla round protein

berry rice alcohol chipotle

brioche peanut specialty cucumber

fruit mousse tab side

syrup cod vodka lettuce

In Table 2 there are product aspects which are

extracted from restaurant reviews. The extracted

product features are examined for the validity of the

proposed method. Three criteria are considered in

the evaluation of these aspects: i) The aspects under

the same topic should be compatible with each other,

ii) aspects can be capture details in the reviews and

iii) the most frequently discussed aspects in

comments can be captured (Jo and Oh, 2011).

In order to measure the generalization

performance of the model a measure of perplexity

which is given in Equation 3. Perplexity is

calculated over test data so 10% of reviews (260

reviews) are used as test data. The perplexity, used

by convention in language modeling, is

monotonically decreasing in the likelihood of the

test data, and is algebraicly equivalent to the inverse

of the geometric mean per-word likelihood (Blei et

al 2004).

()





















−=



test

Dperplexity

log

exp)(

(3)

The obtained results are given in Figure 5.

Figure 5: Perplexity results for reviews of restaurants.

Extracting Implicit Aspects based on Latent Dirichlet Allocation

As the obtained result is evaluated it can be clearly

seen that the model has a good generalization

performance.

5 EXPECTED OUTCOME

The outcomes of this study present implicit aspect

extraction from reviews of restaurants in English.

For this purpose a novel framework will be designed

for implicit aspect extraction by using semantic

similarity based LDA. For semantic similarity of

reviews concepts, which are obtained by using

Babelfy, will be extracted and these concepts will be

represented in high dimensional space. The

generalization performance of the proposed model

will be compared with LDA.

6 STAGE OF THE RESEARCH

This paper provides with the background of the

research that implicit aspect extraction from reviews

in English. In this paper motivation and objectives of

the research, literature review is given. The current

stage of the research is focusing initially on the first

forth stages.

The next stage we will plan to extract concepts

by using Babelfy. These concepts will be used for

semantic similarity of reviews. As a result, the goal

of this stage is to organize topic proportions based

on these similarity results. In this way, we aim to

improve generalization performance of the LDA.

REFERENCES

Fersini, E., Pozzi, F. A., Messina, E., 2016. Approval

network: a novel approach for sentiment analysis in

social networks. World Wide Web, Springer, 1-24.

Ha, S. H., Bae, S. Y., Son, L. K., 2015. Impact of Online

Consumer Reviews on Product Sales: Quantitative

Analysis of the Source Effect. Applied Mathematics &

Information Sciences, 9(2L), 373-387.

Pawar, A. B., Jawale, M. A., Kyatanawar, D. N., 2016.

Fundamentals of Sentiment Analysis: Concepts and

Methodology. Sentiment Analysis and Ontology

Engineering, 25-48.

Türkmen, H., Ekinci, E., İlhan Omurca, S., 2016. A Novel

Method for Extracting Feature Opinion Pairs for

Turkish. Artificial Intelligence: Methodology, Systems,

and Applications, 162-171.

Liu, B., 2012. Sentiment Analysis and Opinion Mining,

Morgan & Claypool Publishers. California.

Vinodhini, G., Chandrasekaran, R. M., 2012. Sentiment

Analysis and Opinion Mining: A Survey. International

Journal of Advanced Research in Computer Science

and Software Engineering, 2(6), 282-292.

Türkmen, H., İlhan Omurca, S., Ekinci, E., 2016. An

aspect based sentiment analysis on Turkish hotel

reviews. Girne American University Journal of Social

and Applied Sciences, 6, 9-15.

Xu, H., Zhang, FF, Wang, W., 2015. Implicit feature

identification in Chinese reviews using explicit topic

mining model. Knowledge-Based Systems, 76, 166-

175.

Stevyers, M., Griffiths, T., 2007. Probabilistic Topic

Models. Handbook of Latent Semantic Analysis, 1-15.

Boyd-Graber, J., Blei, D., 2009. Syntactic Topic Models.

Neural Information Processing Systems, 185-192.

Blei, D.M., Ng, A.Y., Jordan, M.I., 2004. Latent dirichlet

allocation. The Journal of Machine Learning

Research, 3, 993-1022.

Mei, Q., Shen, X., Zhai, C., 2007. Automatic Labeling of

Multinomial Topic Models. In Proceedings of ACM

KDD, 490-499.

Moro, A., Cecconi, F., Navigli, R., 2014. Multilingual

Word Sense Disambiguation and Entity Linking for

Everybody. In Proceedings of the 13th International

Semantic Web Conference, 25-28.

Navigli, R., Ponzetto, S. P., 2012. BabelNet: The

automatic construction, evaluation and application of a

wide-coverage multilingual semantic network.

Artificial Intelligence, 193, 217-250.

Jo, Y., Oh, A., 2011. Aspect and Sentiment Unification

Model for Online Review Analysis. In Proceedings of

4th ACM International Conference on Web Search

and Data Mining, 815-824.

Ekinci, E., Türkmen, H., İlhan Omurca, S., 2016. Multi-

word Aspect Extraction from user reviews. In World

Conference on Innovation & Computer Science, 1-11.

Hu, M., Liu, B., 2004. Mining Opinion Features in

Customer Reviews. In Proceedings of 19th National

Conference on Artificial Intelligence, 755-760.

Hu, M., Liu, B., 2004. Mining and Summarizing Customer

Reviews. In Proceedings of International Conference

on Knowledge Discovery and Data Mining,

Publishing, 168-177.

Popescu, A. M., Etzioni, O., 2005. Extracting product

features and opinions from reviews. In Proceedings of

Conference on Empirical Methods in Natural

Language Processing, 339-346.

Wei, C. P., Chen, Y. M., Yang, C. S., Yang, C. C, 2010.

Understanding What Concerns Consumers A Semantic

Approach to Product Feature Extraction from

Consumer Reviews. Information Systems and E-

Business Management, 8 (2), 149-167.

Brody, S., Elhadad, N., 2010. An Unsupervised Aspect-

Sentiment Model for Online Reviews. In Proceedings

of Annual Conference of the North American Chapter

of the Association for Computational Linguistics,

Publishing, 804-812.

DCAART 2017 - Doctoral Consortium on Agents and Artiﬁcial Intelligence

Wang, W., 2010. Sentiment Analysis of Online Product

Reviews with Semi-supervised Topic Sentiment

Mixture Model, In Proceedings of 7th International

Conference on Fuzzy Systems and Knowledge

Discovery, 2385-2389.

Jo, Y., Oh, A., 2011. Aspect and Sentiment Unification

Model for Online Review Analysis. In Proceedings of

4th ACM International Conference on Web Search

and Data Mining, 815-824.

Xianghua, F., Guo, L., Yanyan, G., Zhiqiang, W., 2013.

Multi-aspect sentiment analysis for Chinese online

social reviews based on topic modeling and HowNet

Lexicon. Knowledge-Based Systems, 37, 186-195.

Ding, W., Song, X., Guo, L., Xiong, Z., Hu, X., 2013. A

Novel Hybrid HDP-LDA Model for Sentiment

Analysis, In Proceedings of 2013 IEEE/WIC/ACM

International Conferences on Web Intelligence and

Intelligent Agent Technology, 329-336.

Bagheri, A., Saraee, M., Jong, F., 2013. Care more about

customers: Unsupervised domain-independent aspect

detection for sentiment analysis of customer reviews,

Knowledge-Based Systems, 52, 201–213.

Wang, T., Cai, Y., Leung, H., Lau, R.Y.K., Li, Q., Min, H.

2014, Product aspect extraction supervised with online

domain knowledge, Knowledge-Based Systems, 71,

86-100.

Zheng, X., Lin, Z., Wang, X., Lin, K.J., Song, M., 2014.

Incorporating appraisal expression patterns into topic

modeling for aspect and sentiment word identification.

Knowledge-Based Systems, 61, 29-47.

Yin, S., Han, J., Huang, Y., Kumar, K., 2014.

Dependency-Topic-Affects-Sentiment-LDA Model for

Sentiment Analysis. In Proceedings of 2014 IEEE

26th International conference on Tools with Artificial

Intelligence, 413-418.

Su, Q., Xu, X., Guo, H., Wu, X., Zhang, X., Swen, B., Su,

Z., 2008. Hidden Sentiment Association in Chinese

Web Opinion Mining, In Proceedings of International

Conference on World Wide Web, 959-968.

Zhang, W., Xu, H., Wan, W., 2012. Weakness Finder:

Find product weakness from Chinese reviews by using

aspects based sentiment analysis. Expert Systems with

Applications, 39(11), 10283–10291.

Wang, W., Xu, H., Wan, W., 2013. Implicit feature

identification via hybrid association rule mining,

Expert Systems with Applications, 40(9), 3518–3531.

Bagheri, A., Saraee, M., Jong, F., 2013. Care more about

customers: Unsupervised domain-independent aspect

detection for sentiment analysis of customer reviews.

Knowledge-Based Systems, 52, 201–213.

Xueke, X., Xueki, C., Songbo, T., Yue, L., Huawei, S.,

2013. Aspect-Level Opinion Mining of Online

Customer Reviews. China Communications, 10(3),

25-41.

Lau, R.Y.K., Li, C., Liao, S.S.Y., 2014. Social analytics:

Learning fuzzy product ontologies for aspect-oriented

sentiment analysis. Decision Support Systems, 65, 80-

94.

Poria, S., Cambria, E., Winterstein, G., Huang, G.B.,

2014. A Rule-Based Approach to Aspect Extraction

from Product Reviews, In: Workshop Proceedings of

the 25th International Conference on Computational

Linguistics.

Moro, A., Raganato, A., Navigli, R., 2014. Entity Linking

meets Word Sense Disambiguation: Uniﬁed

Approach. Transactions of the Association for

Computational Linguistics, 2, 231-244.

Stevyers, M., Griffiths, T., 2007. Probabilistic Topic

Models, Handbook of Latent Semantic Analysis.

Jadhav N., 2014. Topic Models for Sentiment analysis: A

Literature Survey. Technical Report, 1-11.

Bishop, C. M., 2006. Pattern Recognition and Machine

Learning, Springer.

Extracting Implicit Aspects based on Latent Dirichlet Allocation