helpfn

704 IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 27, NO. 4, APRIL 2019

Sentiment Lexicon Construction With Hierarchical Supervision Topic Model

Dong Deng , Liping Jing , Jian Yu, Shaolong Sun , and Michael K. Ng

Abstract—In this paper, we propose a novel hierarchical su- pervision topic model to construct a topic-adaptive sentiment lexicon (TaSL) for higher-level classification tasks. It is widely rec- ognized that sentiment lexicon as a useful prior knowledge is crucial in sentiment analysis or opinion mining. However, many existing sentiment lexicons are constructed ignoring the variability of the sentiment polarities of words in different topics or domains. For example, the word “amazing” can refer to causing great surprise or wonder but can also refer to very impressive and excellent. In TaSL, we solve this issue by jointly considering the topics and sen- timents of words. Documents are represented by multiple pairs of topics and sentiments, where each pair is characterized by a multinomial distribution over words. Meanwhile, this generating process is supervised under hierarchical supervision information of documents and words. The main advantage of TaSL is that the sentiment polarity of each word in different topics can be sufficiently captured. This model is beneficial to construct a domain-specific sentiment lexicon and then effectively improve the performance of sentiment classification. Extensive experimental re- sults on four publicly available datasets, MR, OMD, semEval13A, and semEval16B were presented to demonstrate the usefulness of the proposed approach. The results have shown that TaSL performs better than the existing manual sentiment lexicon (MPQA), the topic model based domain-specific lexicon (ssLDA), the expanded lexicons(Weka-ED, Weka-STS, NRC, Liu’s), and deep neural network based lexicons (nnLexicon, HIT, HSSWE).

Index Terms—Sentiment analysis, topic model, sentiment lexicon construction, opinion mining, text mining.

Manuscript received April 27, 2018; revised September 22, 2018 and De- cember 15, 2018; accepted December 24, 2018. Date of publication January 10, 2019; date of current version February 15, 2019. This work was supported in part by the National Natural Science Foundation of China under Grants 61822601, 61773050, and 61632004, in part by the Beijing Natural Science Foundation under Grant Z180006, and in part by the Beijing Municipal Science & Technology Commission under Grant Z181100008918012. The associate ed- itor coordinating the review of this manuscript and approving it for publication was Prof. Carlos Busso. (Corresponding author: Liping Jing.)

D. Deng, L. Jing, and J. Yu are with the School of Computer and Informa- tion Technology, Beijing Jiaotong University, Beijing 100044, China (e-mail:, [email protected]; [email protected]; [email protected]).

S. Sun is with the Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing 100190, China, with the School of Economics and Management, University of Chinese Academy of Sciences, Beijing 100190, China, with the Department of Systems Engineering and Engineering Manage- ment, City University of Hong Kong, Hong Kong, and also with the Center for Forecasting Science, Chinese Academy of Sciences, Beijing 100190, China (e-mail:, [email protected]).

M. K. Ng is with the Center for Mathematical Imaging and Vision, and Department of Mathematics, Hong Kong Baptist University, Hong Kong (e-mail:, [email protected]).

Digital Object Identifier 10.1109/TASLP.2019.2892232

I. INTRODUCTION

W ITH the rapid growth of social media, delivering or shar-ing messages online has become a very popular activity in our daily life. Thus, a large amount of opinionated data like movie reviews, product comments and microblogs, are growing at an exponential rate every day. Nowadays, such big opinion- ated data is valuable to provide a great convenience for our diverse decision-making processes [1], [2]. For instance, when trying to buy a dress, the lady would like to review the online comments on the dresses before asking their friends for opin- ions. Governments or companies also do not need to conduct opinion polls toward their services or policies because there is an abundant of publicly available information.

Sentiment analysis, as a typical technique to extract and iden- tify opinions from the text, has attracted much attention and is being widely used in many domains [2]–[8]. So far, many ap- proaches have been proposed to do sentiment analysis, which can be roughly classified into document level, sentence level and entity and aspect level classification [1], [9]. Sentiment lex- icon plays a crucial role in these problems because it contains the sentiment polarities of words or phrases which can be used as the supplementary or prior feature of text data [10], such as beautiful and good are commonly used positive words, bad and awful are usually negative words.

Nowadays, many sentiment lexicons have been proposed which can be roughly classified into domain-independent and domain-specific lexicons. Domain-independent lexicons are usually manually constructed [11], [12] or using diverse statistical co-occurrence information between candidates and sentiment labels [13]–[19]. However, manually constructed lex- icons are usually time-consuming, high cost and low coverage, which cannot be satisfied in practice. Other lexicons constructed by using co-occurrence information cannot discover some im- plicit and connotative sentiment words which are important in some specific domains. Different from domain-independent lex- icons, domain-specific lexicons can sufficiently utilize the spe- cific domain information. The methods in [20]–[24] use diverse characters among documents to construct sentiment lexicons. However, these methods do not consider the variability of words’ sentiment polarities in different domains or topics. As we know, a word’s sentiment may vary according to the domains or topics in which this word occurs. For example, the word positive may convey positive opinion when associated with the word feeling, but negative opinion when associated with word HIV. Therefore, it is important and practical to identify the variability of words’ sentiment polarities according to the domains where they occur.

2329-9290 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

Authorized licensed use limited to: University of the Cumberlands. Downloaded on July 24,2021 at 03:32:52 UTC from IEEE Xplore. Restrictions apply.

DENG et al.: SENTIMENT LEXICON CONSTRUCTION WITH HIERARCHICAL SUPERVISION TOPIC MODEL 705

In this paper, we propose a novel hierarchical supervision topic model to construct a topic-adaptive sentiment lexicon (TaSL). Different from other existing approaches, our model is designed to explicitly consider the effects of domains or top- ics to words’ sentiment polarities. Specifically, In TaSL, each document is represented by multiple pairs of latent topics and sentiments, where each pair is characterized by a multinomial distribution over words. Additionally, it is convenient for TaSL to integrate the hierarchical supervision information of doc- uments and words. The final goal of TaSL is to capture the sentiment polarities of words under different topics, provide a domain-specific lexicon, and improve the performance of senti- ment classification.

The main contribution of this work can be summarized as follows.

� A new probabilistic graphical model, TaSL, is proposed to construct a domain-specific sentiment lexicon.

� TaSL has the ability to take advantage of the hierarchical supervision information of documents and words, e.g., the sentiment label of documents and some commonly used sentiment words.

� An effective algorithm based on gibbs sampling is designed to infer the proposed graphical model.

� A series of experiments are conducted on benchmark datasets (MR and OMD) and the challenging datasets (Se- mEval13A and SemEval16B) to show the performance of TaSL by comparing with baselines.

The rest of this paper is organized as follows. Section II in- troduces some related works. Section III describes the proposed TaSL model and its inference process. Section IV shows how to construct the sentiment lexicon and classify documents into positive or negative through our proposed lexicon. A series of experiments on real-world datasets (Movie Reviews, Obama- McCain Debate Comments, Twitter benchmark (SemEval13A and SemEval16B)) are conducted in Section V. The results have shown that the proposed TaSL lexicon achieves better per- formance comparing with the state-of-the-art lexicons. A brief conclusion and future work are given in Section VI.

II. RELATED WORK

In order to construct sentiment lexicons, many approaches have been proposed. Existing lexicons can be roughly classified into domain-independent and domain-specific lexicons. Both of them are very important and widely used in many applications.

For domain-independent sentient lexicons, at the beginning, the experts manually determine the sentiment polarities of words or phrases [11], [12] or use Mechanical Turk [25]. However, this strategy is usually time-consuming, high cost and low coverage. Meanwhile, it may result in large variance due to the personal opinions. In order to solve the above issues, some lexicons are constructed by using semi-automatic ways. Related works in [13], [15]–[17], [26], [27] take advantage of thesauri infor- mation (e.g., WordNet) to expand the seed list. However, this requirement limits their applications. Bravo-Marquez et al. [28] used machine learning techniques, e.g., classification, to build a word sentiment classifier based on the existing lexicons, and automatically determine the sentiment polarities of new com-

ing words. In addition, Statistical co-occurrence information between words (e.g., point-wise mutual information) [18], [29], neural network based techniques [19], [30]–[35] are also used to construct sentiment lexicons. Although such methods can gen- erate a large scale sentiment lexicon, they did not consider the variability of words’ sentiment polarities under different topics or domains. Thus, these lexicons are usually more general which make them less accurate in some specific domains.

To meet the requirements of constructing domain-specific lexicons, researchers have proposed many works. The methods in [20]–[22] consider the relationship (such as mutual infor- mation, contextual or semantic similarity) between words in the target domain and the words in the seed sentiment lexicon. Then the candidates are chosen which are close enough to the seed words. Similarly, linear programming [36], graph prop- agation algorithm [23], integrating lexicon-based and corpus- base system [24], context information [37] are also used to construct domain-specific sentiment lexicons. Although these studies have considered different information among documents and use diverse models to discover sentiment words, they did not consider the variability of words’ sentiment polarities under latent topics. Xie et al. [38] used probabilistic topic model [39] to generate sentiment polarities of words. However, this model just constructs a domain-independent lexicon. In order to con- struct domain-specific sentiment lexicon depending on differ- ent topics or domains, Yang et al. [40] proposed a supervised sentiment-aware LDA (ssLDA) model (as shown in Fig. 2(a)) by extending the traditional topic model (LDA). It assumes that the whole document set has two kinds of topics: sentiment and non- sentiment topics. In ssLDA, the sentiment topics contain two topics (positive topic and negative topic). Meanwhile, ssLDA assumes that all words can be separated into two parts, one part of words convey the sentiment polarities, and the other part does not. In other words, this strategy only adopted two senti- ment topics to model words which contain sentiment polarities. Actually, it’s not enough for ssLDA to model the variability of words’ sentiment polarities under different topics.

In this paper, in order to sufficiently characterize the rela- tionship among documents, topics, words, and sentiments, we propose a new topic-adaptive sentiment lexicon construction model (TaSL) by introducing latent sentiment and topic vari- ables. In TaSL, each word can be generated over both topics and sentiments and each document follows a multinomial distribu- tion on pairs of topics and sentiments. The proposed model will be described in the next section.

III. TASL MODEL

The proposed topic-adaptive sentiment lexicon model aims to characterize the property that each word may have multiple sen- timent polarities under different topics. In order to implement this, we jointly model the latent topics and sentiments, where each word is generated via a multinomial distribution condi- tioned on pairs of topics and sentiments, as shown in Fig. 2(b).

A. Primary Notation

Let D = {d1 , d2 , . . . , dm , . . . , dM } be the collection of M documents, where each document is represented via a sequence

Authorized licensed use limited to: University of the Cumberlands. Downloaded on July 24,2021 at 03:32:52 UTC from IEEE Xplore. Restrictions apply.

706 IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 27, NO. 4, APRIL 2019

TABLE I NOTATION USED THROUGHOUT THE ARTICLE

of Nm words denoted as dm = {w1 , w2 , . . . , wN m }. Each word in the document is taken as an item. V represents the num- ber of distinct terms in the vocabulary which is denoted by {1, 2, . . . , V }. K is the number of topics, and O be the number of sentiment polarities of words. The notation is summarized in Table I.

B. Model Formulation

TaSL is a generative probabilistic model. The basic idea is that documents are represented as random mixtures over pairs of topics and sentiments, where each pair of topic and senti- ment is characterized by a distribution of words. In this paper, we assume K topics ({z1 , . . . , zk . . . , zK }) and O sentiments ({e1 , . . . , eo . . . , eO }) in each document. The overall procedure to generate n-th word in m-th document can be implemented via two steps: generating a pair of topic z and sentiment e from the multinomial distribution on (θm ), and then choosing the word from the multinomial distribution on ϕ(z ,e) . The generative pro- cess of TaSL is summarized as follows.

1) For each pair of topic and sentiment (z,e), z ∈ {1, 2, . . . , K}, e ∈ {1, 2, . . . , O}, choose the parameters ϕ(z ,e) ∈ RV over all words from Dirichlet distribution, ϕ(z ,e) ∼ Dir(β).

2) For each document dm , m ∈ {1, . . . , M} a) generate its’ sentiment label ym via ym ∼ Bern(ρ). b) generate its’ joint distribution θm over topic and

sentiment pair (z,e) via θm ∼ Dir(α(z ,e)y m ). 3) For each word (represented as n-th word wn in corpus) in

document dm a) generate a pair of topic and sentiment (z,e) from

joint distribution θm via (z, e) ∼ M ult(θm ). b) generate word wn from distribution ϕ

(z ,e) via wn ∼ M ult(ϕ(z ,e) )

In TaSL, it permits that each word has O sentiment po- larities. Meanwhile, in order to utilize the supervision in- formation of documents, the hyperparameter α is selected according to the sentiment labels y of documents. More specif- ically, α

(z ,e) y m := η

e y m

for e ∈ {1, 2, . . . , O}, where the m-th document’s sentiment label ym is generated from a Bernoulli distribution with hyperparameter ρ, i.e., ym ∼ Ber(ρ). In other words, if the document has sentiment (ŷm ), η

e ŷ m

is set to be greater than ηey̌ m (y̌m is the sentiment that the current docu- ment does not have). This setting prefers to assigning the words to the same sentiment as the current document. Similarly, the hyperparameter β can be set according to the prior knowledge about the word sentiment, e.g., the existing lexicons. In detail, β

(z ,e) w n = γ

e lw n

, e ∈ {1, 2, . . . , O}, lw n is the sentiment polarity of word wn which can be pre-obtained from the existing senti- ment lexicons. Furthermore, γêlw n is set to be greater than γ

ě lw n

(ê is equal to lw n and ě represents other sentiment polarities except for lw n ) which makes the word to a large extent be sampled from the lw n sentiment.

In TaSL, each pair of latent topic and sentiment in m-th doc- ument θm is characterized by the (K*O)-dimensional Dirichlet distribution. It has the following probability density.

p(θm ; α) = Γ

(∑K z = 1

∑O e= 1 α

(z ,e) )

∏K z = 1

∏O e= 1 Γ(α

(z ,e) ) θα

( 1 , 1 ) −1 (1,1) , . . . , θ

α ( K , O ) −1 (K ,O )

(1)

for the n-th word wn in m-th document, its’ corresponding pair of topic and sentiment (zn , en ) is generated from the multino- mial distribution of θm through P ((zn , en )|θm ). Once deter- mining the topic and sentiment, word wn is generated from the multinomial distribution ϕ(zn ,en ) by P (wn |(zn , en ), ϕ). In this case, the joint distribution of hidden and observed variables in TaSL can be formulated as follows:

p( wm , ( zm , em ), θm ; α, β) =

P (θm ; α) N m∏ n = 1

P ((zn , en )|θm )P (wn |(zn , en ), ϕ; β) (2)

( zm , em ) indicates pairs of topic and sentiment assignments for words in m-th document. wm represents the word sequences in m-th document. Then, by margining over the latent variables z and e, words wm in m-th document are generated via a multi- nomial distribution conditioned on the multiple pairs of topics

Authorized licensed use limited to: University of the Cumberlands. Downloaded on July 24,2021 at 03:32:52 UTC from IEEE Xplore. Restrictions apply.

DENG et al.: SENTIMENT LEXICON CONSTRUCTION WITH HIERARCHICAL SUPERVISION TOPIC MODEL 707

Fig. 1. Framework of (a) the existing ssLDA model and (b) the topic-adaptive model TaSL.

and sentiments ((z, e)), their probabilities can be calculated as

p( wm ; α, β) = ∫

θm

{ N m∏ n = 1

K ,O∑ (zn ,en )

p(wn |(zn , en ), ϕ; β)P ((zn , en )|θm ) ⎫ ⎬ ⎭ p(θm ; α)dθm

(3)

Among them, the topic and sentiment assignment (zn , en ) of the n-th word in the m-th document is simultaneously gener- ated by a multinomial distribution, i.e., (zn , en ) ∼ M ult(θm )), where θm ∈ RK ×O is chosen from Dirichlet distribution with hyperparameter α, i.e., θm ∼ Dir(α) under the document’s sentiment label ym . Once obtaining the sentiment en and topic zn , the n-th word is generated by a multinomial distri- bution on the pair of latent topic and sentiment, i.e., wn ∼ M ult(ϕ(z ,e)|(zn , en )). ϕ(z ,e)|(zn , en ) ∈ RV is chosen from Dirichlet distribution with hyperparameter β, i.e., ϕ(z ,e) ∼ Dir(β). This is the main contribution of TaSL, different from ssLDA where each word is only drawn from different kinds of topics.

C. Comparing With ssLDA

TaSL and ssLDA are both generated through a probabilistic topic model which considers the latent sentiments and topics simultaneously. As shown in Fig. 1(a), ssLDA divides the top- ics and words into two types: sentiment (θ(s) , w(s) ) and non- sentiment (θ(n ) , w(n ) ). Before generating the topics and words, ssLDA firstly determines the sentiment (s). ssLDA is similar to TaSL which also utilizes topic model and supervision infor- mation. However, TaSL differs from ssLDA in many aspects. Firstly, TaSL defines a unified joint distribution which models the latent topics and sentiments simultaneously. It differs from ssLDA which determines the sentiment polarities and topics in

two separate steps. Secondly, ssLDA just uses two sentiment topics to model the variability of words’ sentiment polarities. In contrast, TaSL fully considers the sentiment polarities of words under each topic. Finally, TaSL utilizes both the document-level and word-level supervision information. However, ssLDA just uses the document-level supervision information.

D. Model Inference

To learn the parameters in TaSL, the Collapsed Gibbs Sam- pling algorithm is adopted. In this paper, we need to specify the conditional probability of hidden variables (z, e), which are the hidden pairs of topics and sentiments associated with the observed words w. However, we cannot directly sample this conditional probability. The collapsed sampler runs a markov chain that uses the full conditional distribution in order to sim- ulate it. In our case, for a given word wn , we need to sample a pair of topic and sentiment (zn , en ) by fixing the topic and sentiment information of other words as follows.

P ((zn = k, en = o)|( z¬n , e¬n ), w). (4) Here w is the word vector in the whole corpus, z¬n indicates the topic vector of all words except for the n-th word in the corpus sequence, and e¬n indicates the sentiment polarity vector of all words except for the n-th word in the corpus sequence. (4) can be reformulated as follows:

P ((zn = k, en = o)|( z¬n , e¬n ), w)

= P (zn = k, en = o) ∗ P (( z¬n , e¬n ), w)

P (( z¬n , e¬n ), w)

(5)

the n-th word’s topic and sentiment (zn = k, en = o) is inde- pendent from other words’ topics and sentiments ( z¬n , e¬n ). Thus, we have

P (zn = k, en = o) ∗ P (( z¬n , e¬n ), w) = P (( z, e), w) (6) P (( z¬n , e¬n ), w) = P (( z¬n , e¬n ), w¬n )P (wn ) (7)

Authorized licensed use limited to: University of the Cumberlands. Downloaded on July 24,2021 at 03:32:52 UTC from IEEE Xplore. Restrictions apply.

708 IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 27, NO. 4, APRIL 2019

By substituting (6) and (7) to (5), we have

P ((zn = k, en = o)|( z¬n , e¬n ), w)

= P (( z, e), w)

P (( z¬n , e¬n ), w¬n )P (wn )

(8)

meanwhile, wn is observed, we can reformulate the (8) as fol- lows:

P ((zn = k, en = o)|( z¬n , e¬n ), w)

∝ P (( z, e), w) P (( z¬n , e¬n ), w¬n )

(9)

here w¬n indicates the word vector except for the n-th word in the corpus. From the proposed TaSL model, the probability of joint distribution on ( z, e) (i.e., the pair sequence on topic and sentiment) and w can be counted as follows.

P (( z, e), w) = P ( w|( z, e))P (( z, e))

= ∏ (k ,o)

�( −−−→ c(k ,o) + β)

�( β) ∏ m

�( −−−→ m(k ,o) + α) �( α)

(10)

where �( α) = ∏

i Γ(α i ) Γ(

∑ i α i )

, �( β) = ∏

j Γ(βj ) Γ(

∑ j βj )

, Γ(αi ) is a Gamma

function with parameter αi . −−−→ c(k ,o) is the number of times that

each word (belonging to the k-th topic and o-th sentiment)

appears in the whole corpus, −−−→ m(k ,o) is the number of words

(belonging to the k-th topic and o-th sentiment) in the m-th document.

Similarly, the probability of the joint distribution on ( z¬n , e¬n ) and w¬n can be computed with

P (( z¬n , e¬n ), w¬n ) = P ( w¬n |( z¬n , e¬n ))P (( z¬n , e¬n ))

= ∏ (k ,o)

�( −−−→ c (k ,o) ¬n + β)

�( β) ∏ m

�( −−−→ m

(k ,o) ¬n + α) �( α) .

(11)

By substituting �(·) to (10) and (11), we can get P ((zn = k, en = o)|( z¬n , e¬n ), w) =

c (k ,o) w n − 1 + βw n( ∑V

t= 1 (c (k ,o) t + βt )

) − 1

× m (k ,o) − 1 + α(k ,o)y m( ∑

(k ,o) (m (k ,o) + α(k ,o)y m )

) − 1 (12)

where the denominator of (12) is to normalize the conditional probability.

After generating the topics, sentiments and words, the param- eters θ

(k ,o) m and ϕ

(k ,o) n can be estimated by

θ(k ,o)m = m(k ,o) + α(k ,o)y m∑

(k ,o)

( m(k ,o) + α(k ,o)y m

) (13)

ϕ(k ,o)w n = c (k ,o) w n + βw n∑V

t= 1 (c (k ,o) t + βt )

(14)

Here ϕ (k ,o) w n will be used to construct the sentiment lexicon.

More details will be given in the next section.

IV. LEXICON-BASED SENTIMENT CLASSIFICATION

A. Constructing Sentiment Lexicon

Once estimating the parameters of TaSL, we can generate the topic-adaptive sentiment lexicon from ϕ(z ,e) . As we know, two kinds of sentiment lexicons are commonly used. The first one only assigns a sentiment score for each word. In this kind of lexi- cons, positive words have positive sentiment scores and negative words have negative sentiment scores. Furthermore, the senti- ment strength of both positive and negative words in the lexicon can be calculated by the absolute value of their corresponding sentiment scores, just like nnLexicon lexicon [30], we call this lexicon TaSL-I. The second one is to assign a probability distri- bution over sentiment polarities (i.e. positive, negative, neutral) for each word. From the lexicon, we can get the sentiment po- larity and the sentiment strength of each word according to the corresponding probabilities, just like the sentiWordnet lexi- con [17]. We call this lexicon TaSL-II. In this paper, our method can construct both kinds of sentiment lexicons.

In order to effectively capture the variability of words’ senti- ment polarities and construct topic-adaptive sentiment lexicons TasL-I and TaSL-II, we need to consider the sentiment polari- ties and strength of words over all topics. Thus, we intuitively assume that if a word has a higher probability on positive senti- ment over all weighted topics, it should be part of the positive sentiment lexicon. This can be verified in Fig. 2. In Fig. 2, the absolute value of y-axis indicates the topic weights of words in each topic. The sign of y-axis represents the sentiment po- larities of words in the corresponding topic. The positive sign indicates the positive sentiment and the negative sign indicates the negative sentiment. The x-axis indicates each topic from 1 to 10. From this figure, we can easily observe that the posi- tive word “great” has a higher probability on positive sentiment over all weighted topics and negative word “bad” has a higher probability on negative sentiment over all weighted topics.

More specifically, the sentiment strength of the n-th word (wn ) over O sentiments can be obtained by marginalizing over all weighted topics, qn = {qen }Oe= 1 where qen =

∑K z = 1 ϕ

(z ,e) w n .

Then, the sentiment polarity of word wn can be determined by ẽn = arg maxe{q1n , . . . , qen , . . . , qOn }. In this paper, for TaSL- I lexicon, the number of sentiment polarities O is set to be 2 (one for positive and others for negative) and the sentiment strength for word wn is determined by qn = qen |e=positiv e − qen |e=n eg ativ e . For TaSL-II lexicon, O is set to be 3 (O = 3 indicates that there are three sentiment polarities of words: pos- itive, negative and neutral).

At the same time, except for the supervision information of documents, TaSL model can also efficiently utilize the prior knowledge of words, such as existing manually constructed sen- timent lexicons. In order to utilize such information, hyperpa- rameter β is constructed according to these existing lexicons, as described in Section III-B. The sentiment lexicons constructed with the aid of prior knowledge of existing sentient lexicons

Authorized licensed use limited to: University of the Cumberlands. Downloaded on July 24,2021 at 03:32:52 UTC from IEEE Xplore. Restrictions apply.

DENG et al.: SENTIMENT LEXICON CONSTRUCTION WITH HIERARCHICAL SUPERVISION TOPIC MODEL 709

Fig. 2. The topic weights and sentiment polarities of words “great” and “bad” under each topic. The number of topic number is 10.

(such as MPQA [12]) are called as TaSL-I-prior for TaSL-I lexicon and TaSL-II-prior for TaSL-II lexicon. Finally, in this paper, we can totally construct four sentiment lexicons: TaSL-I, TaSL-II, TaSL-I-prior and TaSL-II-prior. In the experiments, four lexicons will be evaluated by comparing with the-state-of- the-art existing lexicons.

B. Lexicon-Based Sentiment Classification

The resulting lexicons can be used to solve the sentiment clas- sification tasks. In this paper, we only consider the binary senti- ment classification task. Multi-sentiment classification tasks are left for our future work. In detail, we use two methods to eval- uate the classification performance of the sentiment lexicons: count-based [40] and score-based [30], [41] methods.

The count-based method aims to compare the number of positive and negative words in each document [40]. If the num- ber of positive words are bigger than that of negative words, the document will be assigned into positive sentiment, otherwise the document is negative, i.e.,

ẽm =

{ P ositive if P Wm > N Wm

N egative otherwise. (15)

where

P Wm = |{wn |wn ∈ wm & ẽn = positive}| and

N Wm = |{wn |wn ∈ wm & ẽn = negative}| ẽm indicates the sentiment polarity of m-th document. P Wm and N Wm represent the number of positive and negative words in m-th document respectively. Thus, if P Wm > N Wm , the sentiment polarity of m-th document is positive, otherwise the sentiment polarity of m-th document is negative.

The score-based method [30], [41] aims to evaluate the total sentiment score of documents. If the total sentiment score of documents is bigger than 0, the documents will be assigned into positive sentiment, otherwise the document is negative. Specifi-

cally, the difference between the positive sentiment strength and negative sentiment strength for each word is used as the word’s attribute via Sn = q

(positiv e) n − q(n eg ativ e)n . Then, the sentiment

polarity of the document (d) is obtained by checking the total sentiment strength of all words in the current document.

ẽm =

{ P ositive if Sm > 0

N egative otherwise. (16)

where

Sm = N m∑ n = 1

Sn .

ẽm indicates the sentiment polarity of m-th document. Sm indicates the overall sentiment score of the m-th document. If Sm > 0, the sentiment polarity of m-th document is pos- itive, otherwise, the sentiment polarity of m-th document is negative.

V. EXPERIMENTS

In this section, a series of experiments are conducted on four real-world datasets to evaluate the performance of the proposed TaSL model for binary sentiment classification.

A. Datasets

In experiments, in order to evaluate our lexicon in different domains, four public-available datasets are used: Movie Review (MR), Obama-McCain Debate (OMD), The Twitter benchmark of SemEval13 task A (SemEval13A) and The Twitter benchmark of SemEval16 task B (SemEval16B). Since we focus on binary sentiment classification, following [42], all neutral samples are removed. The reason why we use these datasets is that, firstly, we want to evaluate our proposed method, TaSL, in different domains. Thus, we select tweets and movie reviews as our datasets. Secondly, these datasets are also utilized in some base- lines, such as ssLDA use MR and OMD to evaluate their lex- icons, HSSWE utilizes SemEval13A and SemEval16B. Thus,

Authorized licensed use limited to: University of the Cumberlands. Downloaded on July 24,2021 at 03:32:52 UTC from IEEE Xplore. Restrictions apply.

710 IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 27, NO. 4, APRIL 2019

TABLE II STATISTICS OF SEMEVAL13A AND SEMEVAL16B

we use these datasets to provide fair comparison with these baselines.

MR contains 2000 reviews and each review is related to the rating star or a numerical score. Following [29], 1000 reviews with larger numerical value or ratings (e.g., having more than 3 stars) are taken as positive examples, and 1000 reviews with smaller numerical value or ratings (e.g., less than 3 stars) as negative examples.

OMD is constructed from 3269 tweets crawled during the 2008 Presidential Debates [43]. The sentiment label of each tweet is acquired through Amazon Mechanical Turk and the final label is determined via the majority voting method. Finally, OMD dataset covers 742 positive and 1246 negative tweets.

SemEval13A comes from the SemEval-2013 Workshop [44] which focuses on semantics and sentiment analysis in the text. Due to that some twitters can not be downloaded, only the available 3827 training and 1637 testing tweets are used in ex- periments, including 3968 positive examples and 1496 negative examples.

SemEval16B comes from the SemEval-2016 Workshop [45]. Among them, we collect 3615 training and 10551 testing tweets. More details about SemEval13A and SemEval16B are shown in Table II.

In the experiments, MR and OMD are tokenized with a natu- ral language processing toolkit NLTK1 and stemmed by Word- Net stemmer.2 In these two datasets, rare words that occur less than 5 times in corpus are removed. For MR and OMD datasets, we use 10-fold cross validation. The original data in OMD are randomly portioned into 10 equal size subsets. For MR, we use authors’ published pre-split 10-folds dataset. Of the 10 subsets, a single subset is retained as the validation data for testing the model and the remaining 9 subsets are used as training data. We repeat this process 10 times with each of the 10 subsets used exactly once as the validation data. For SemEval13A and 16B, we just use the training and testing data which have been pre-split, not the validation dataset.

SemEval13A and SemEval16B are firstly processed by re- placing the url link, name and number with special word “URL”, “NAME” and NUMBE respectively. Then they are tokenized with the stanford natural language processing toolkit.3 When dealing with SemEval13A, the model is trained and tested according the pre-split dataset as shown in Table II. But for SemEval16B, in order to make use of the existing resource,

1http://www.nltk.org 2http://wordnet.princeton.edu 3http://nlp.stanford.edu/software/

we combine the training data of both SemEval13A and Se- mEval16B to train ssLDA and TaSL model.

B. Methodology

Nine existing sentiment lexicons are used as baselines. Among them, MPQA [12] are manually constructed. Liu’s [26], NRC [18], HSSWE [31], HIT [19], Weka-ED and Weka- STS [28] are lexicons which are constructed by expanding the existing manual lexicon or using thesauri information and docu- ments. Since the existing sentiment lexicons, MPQA and Liu’s, only contain the sentiment label of words (such as negative label for word bad, positive label for word good and etc.), we cannot conduct experiments for these two lexicons in the score-based evaluation measurement. In our proposed method, the manually constructed lexicon MPQA is used to set the parameter β, but it’s worth noting that other domain-independent sentiment lexicons can also be integrated into our model except for MPQA.

HSSWE, HIT, nnLexicon and ssLDA are automatically con- structed lexicons. ssLDA is constructed through a supervised sentiment-aware LDA model which is trained on the training data provided by the learning task. In this case, the lexicon is domain-specific. However, the words in nnLexicon, HIT and HSSWE are extracted from a large amount of tweet data in- stead of the limited training data. In this case, we can say that nnLexicon, HIT and HSSWE are more general than ssLDA and TaSL.

In the experiments, we implement ssLDA by ourselives and set its’ parameters following [46]. For HSSWE and HIT,4

nnLexicon,5 Weka-ED and Weka-STS lexicons6 and NRC,7 we directly use these lexicons provided by the authors.

All sentiment classification tasks are evaluated via four met- rics: precision (Pre), recall (Rec), F-score (F1) and Accuracy (Acc). Meanwhile, these results are the averaged values by train- ing the model three times. Finally, three metrics (precision, re- call, and F-score) for each category (positive and negative) are listed and the overall accuracy for two categories are also listed.

C. Performance of Document Sentiment Classification

For TaSL, the hyper-parameters β and α, and the number of latent topics K are tuned via cross-validation technique. In detail, α

(z ,e) y m = η

e y m

|e ∈ {1, 2, . . . , O} and βz ,ew n = γelw n , e ∈ {1, 2, . . . , O}. For convenience, we introduce four extra param- eters ε, π, and υ to set α and β (ε, π, , υ ≥ 0). For lexicons TaSL-I and TaSL-I-prior, the number of sentiment polarities of words O is set to 2. Thus, we set η1 = (ε + π, ε), η2 = (ε, ε + π), γ1 = ( + υ, ) and γ2 = ( , + υ). For lexicons TaSL-II and TaSL-II-prior, the value of O is set to 3. Thus, we set η1 = (ε + π, ε, ε), η2 = (ε, ε + π, ε), γ1 = ( + υ, , ) and γ2 = ( , + υ, ). The proper value of (ε, π, , υ) can be tuned with cross-validation on training data.

4https://github.com/NUSTM/HSSWE 5https://github.com/duytinvo/acl2016 6http://www.cs.waikato.ac.nz/ml/sa/lex.html#ijcai15 7www.saifmohammad.com/WebPages/Abstracts/NRC-SentimentAnalysis.

htm

Authorized licensed use limited to: University of the Cumberlands. Downloaded on July 24,2021 at 03:32:52 UTC from IEEE Xplore. Restrictions apply.

DENG et al.: SENTIMENT LEXICON CONSTRUCTION WITH HIERARCHICAL SUPERVISION TOPIC MODEL 711

TABLE III COMPARING TASL (K = 95) WITH SSLDA ON MR

In this paper, we use different strategies to set these hyper- parameters. For topic number K in ssLDA and TaSL, we con- duct experiments by varying K from 5 to 95 with a step of 10. Then we choose the value which corresponds to the best performance. For hyperparameter α, it is determined by ε and + π. We firstly fix the ε + π from 0.05 to 1 with a step of 0.05. When ε + π is chosen, for example 0.95, we then adjust the parameter π from 0.01 to 0.95 with a step of 0.05. For hyperparameter β, it is determined by and + υ. Different from hyperparameter α, we choose the parameter and + υ in a set [0.001, 0.0045, 0.008, 0.045, 0.01, 0.08, 0.09, 0.1, 0.8]. It’s worth noting that the value of + υ is bigger than that of . The values of and υ are chosen when they generate the best performance.

For MR dataset, the hyperparameters of TaSL are ε = 0.01, π = 0.94, = 0.0045, υ = 0.045 and K = 95. Since HSSWE, HIT, NRC, nnLexicon and Weka-ED/Weka-STS are constructed from the social media data (twitters), which are different from the review data, they are not used to analyze MR data. Finally, three existing methods (MPQA, Liu’s, ssLDA) are used as base- lines for count-based classification, and ssLDA as a baseline for score-based classification. Table III lists their prediction perfor- mance on testing data, where the best and second best results are highlighted in bold and underlined italics respectively.

From Table III, we can observe that TaSL and TaSL with prior knowledge (TaSL-prior), significantly outperform ssLDA and other existing lexicons in most cases including both count-based and score-based sentiment classification tasks. In terms of the overall evaluation metrics, F-score and accuracy, the proposed lexicon has a significant improvement over the baselines. For instance, on count-based classification task, TaSL-type lexicon has 19.73% improvement (81.71% vs. 68.24%) in terms of F- score on positive documents, 8.39% improvement (82.22% vs. 75.85%) in terms of F-score on negative documents and 13.83% improvement (81.82% vs. 71.88%) in terms of Accuracy on all testing data. On score-based classification task, TaSL-type lexi- con has 21.56% improvement (85.09% vs. 70.00%) in terms of F-score on positive documents, 10.32% improvement (84.99% vs. 77.04%) in terms of F-score on negative documents and

14.89% improvement (85.05% vs. 74.03%) in terms of Accu- racy on all testing documents.

For OMD dataset, we set ε = 0.01, π = 0.94, = 0.08, υ = 0.02 and K = 20. Nine baselines (MPQA, Liu’s, NRC, ssLDA, Weka-ED, Weka-SS, nnLexicon, HIT and HSSWE) are used to compare the proposed TaSL lexicon. The results are shown in Table IV. As expected, the lexicons constructed from the current dataset, e.g., ssLDA and TaSL, outperform the general manually built lexicons: MPQA, Liu’s, NRC and lex- icons constructed from large-scale Wikipedia and twitter data set: Weka-ED, Weka-STS, nnLexicon, HIT and HSSWE. Even though the later five lexicons take advantage of a larger dataset, they prefer to the general purpose. In other words, they cannot effectively capture the domain-specific information like ssLDA and TaSL. Recall that the OMD data is collected from twit- ter, but it is focused on a social event (the 2008 Presidential Debates between Obama and McCain), thus, domain-specific lexicons generated from ssLDA and TaSL are more proper for OMD dataset than the general lexicons, Weka-ED, Weka-STS and nnLexicon. Again, TaSL performs better than ssLDA in most cases because TaSL introduces more topics to characterize the content of words, and uses pairs of topics and sentiments to model the sentiment polarity of each word.

However, compared with the improvement of TaSL in MR dataset, our lexicon did not get a significant improvement (TaSL has 0.2% and 4.54% improvement in terms of overall accuracy of the count-based and score-based evaluation respectively) than ssLDA in this dataset, it’s mainly because i) number of words contained among a single review in MR is much bigger (nearly 10 times) than that of a single twitter and it has been known that the probabilistic topic model performs better in long docu- ments [47]. ii) OMD dataset is collected during the 2008 Presi- dential Debates whose topics among texts are intrinsically less than that in MR dataset. The variability of sentiment polarities of words under different topics are also less than that of words in MR dataset.

Similarly, the comparison results on SemEval13A and Se- mEval16B are given in Table V and VI respectively. For Se- mEval13A dataset, we set ε = 0.01, π = 0.94, = 0.0045,

Authorized licensed use limited to: University of the Cumberlands. Downloaded on July 24,2021 at 03:32:52 UTC from IEEE Xplore. Restrictions apply.

712 IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 27, NO. 4, APRIL 2019

TABLE IV COMPARING TASL (K = 20) WITH SSLDA ON OMD

TABLE V COMPARING TASL (K = 45) WITH FOUR BASELINES ON SEMEVAL13A

Authorized licensed use limited to: University of the Cumberlands. Downloaded on July 24,2021 at 03:32:52 UTC from IEEE Xplore. Restrictions apply.

DENG et al.: SENTIMENT LEXICON CONSTRUCTION WITH HIERARCHICAL SUPERVISION TOPIC MODEL 713

TABLE VI COMPARING TASL (K = 50) WITH FOUR BASELINES ON SEMEVAL16B

TABLE VII PERFORMANCE OF TASL IN SEMEVAL 2016 TASK B

υ = 0.045 and K = 45. For SemEval16B, we set ε = 0.01, π = 0.94, = 0.008, υ = 0.072 and K = 50. Since both datasets are from twitters, all nine baselines can be used.

From Table V and VII, we observe that TaSL-type lexicons are superior or competitive to all the baselines in most cases on both datasets. An interesting observation is that there is a big gap between positive documents prediction results and nega- tive documents prediction results. The main reason is that there are much more positive documents than negative documents in training data, which leads to that the majority class (i.e., posi- tive) dominates the final performance [48].

It is expected that the large-scale data-based lexicons (such as Weka-ED, Weka-STS, nnLexicon, HIT and HSSWE) out- perform the manually generated lexicons on these two datasets, which is consistent with the results obtained in [31]. However, the general purpose lexicons, nnLexicon, HIT and HSSWE, re- sult in its inferiority by comparing with the proposed TaSL- type lexicons, especially on the overall accuracy evaluation metric.

It is interesting that the pre-constructed lexicons (such as MPQA, Liu’s, NRC, Weka-ED, Weka-STS, nnLexicon, HIT and HSSWE) almost perform worse on SemEval16B than Se- mEval13A, while ssLDA and TaSL-type lexicons obtain con- trary results. Actually, SemEval16B task is more difficult than

SemEval13A because the former contains much more testing documents. Fortunately, ssLDA and TaSL-type lexicons take advantage of the given training data to learn domain-specific lexicons, where both SemEval13A and SemEval16B training data are used to build the corresponding lexicons, and predict the sentiment polarities of SemEval16B testing documents. Again, TaSL outperforms ssLDA in most cases.

For SemEval16B, we also compare TaSL-type lexicons with the published 19 methods in SemEval16B competition. Note that most published methods in the competition adopted more complicated framework, e.g., ensemble learning. Following the competition setting [45], three evaluation metrics ρP N , F P N1 and Acc are adopted and they are defined as follows.

F P N1 = F 1P + F 1N

2

ρP N = RecP + RecN

2

Acc is the overall accuracy on both positive and negative sam- ples. The comparison results are shown in Table VII. Each value has a superscript and subscript. The superscript indicates the to- tal number of methods participated in this completion, including our proposed TaSL Lexicon, the subscript records the rank of the proposed method in all compared methods (smaller subscript indicates better performance). It can be seen that, the proposed TaSL lexicons rank ahead relative to the complicated methods, which further confirm that topic-adaptive sentiment lexicons are helpful for sentiment classification.

D. Topic-Adaptive Word Sentiment Analysis

In order to investigate how TaSL determines the topic- adaptive sentiments of words, we check the estimation of ϕ

(z ,e) i

Authorized licensed use limited to: University of the Cumberlands. Downloaded on July 24,2021 at 03:32:52 UTC from IEEE Xplore. Restrictions apply.

714 IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 27, NO. 4, APRIL 2019

Fig. 3. Proportion of words whose sentiment polarities are changed between two topics.

of the i-th word. Take SemEval13A dataset as an example, in the k-th topic, we determine the sentiment of the i-th word via

ẽki = arg max e

{ϕ(k ,e)w i |e ∈ {1, 2, 3}}. Then, the number of words whose sentiments are varying in two topics (k1 and k2 ) can be counted by

WS C = V∑

i= 1

δ ( ẽk 1i , ẽ

k 2 i

) (17)

where

δ(ẽk 1i , ẽ k 2 i ) =

{ 1 if ẽk 1i = ẽk 2i 0 otherwise.

(18)

Fig. 3 lists the fraction of words whose sentiments vary in each pair of topics (i.e., WS C /V , here V = 10693), here the first ten topics are demonstrated. It can be seen that in each pair of topics, around 1.7% words (i.e., 181) changed their sentiment polarities along topics.

Table VIII shows four words (“great”, “war”, “eye” and “pos- itive” ) in SemEval13A whose sentiment polarities change in different topics and the corresponding documents. For instance, word “positive” conveys positive opinion in the topic about moods but conveys negative opinion in the topic related to drug tests. For the word “war”, it is a positive word when it appears in the title of the book that the user is interested in, but it becomes a negative word when it describes the massive stuff or real war. The same situation can also be observed for words “eye” and “great” from the listed document examples. Thus, it is neces- sary to identify words’ sentiment polarities according to the topics.

E. Evaluation of Lexicon Quality

The difference between probabilities that the n-th word be- longing to the positive and negative sentiment, i.e., qpositiv en − qn eg ativ en are shown in Table X. Here five positive words and five negative words are listed. The large absolute value marked in bold indicates that there is no doubt that the word belongs to the corresponding sentiment. For such kind of words with clear sentiment polarities, both TaSL and ssLDA can obtain competitive results.

TABLE VIII WORD EXAMPLES EXTRACTED FROM SEMEVAL13A

Meanwhile, according to the value of qen = ∑K

z = 1 ϕ (z ,e) w n ,

i.e., the probability that the n-th word belonging to the e- th sentiment, we can rank all positive or negative words ac- cording to this polarity score. Table IX lists the top 20 pos- itive and negative words from SemEval13A identified by TaSL-II, TaSL-II-prior and the baseline ssLDA. As expected, the proposed methods output more reasonable results than ssLDA.

F. Effect of Topic Size in TaSL

In this paper, we compare the performance of many exist- ing sentiment lexicons with ours. Among them, we directly use the lexicons, MPQA, Liu’s, NRC, Weka-ED, Weka-STS, HIT, nnLexicon and HSSWE, which are released by the authors. Thus, the value of K has no effect on the above sentiment lexi- cons. But for ssLDA, the authors did not release the lexicon. In order to demonstrate the effect of K on the learnt sentiment lex- icons, we conduct experiments by varying K from 5 to 95 with the step of 10 on SemEval13A and keep other hyper-parameters same as in the above described.

Fig. 4 gives the effect of K on ssLDA, TaSL-II and TaSL-II- prior on the score-based sentiment classification task in terms of accuracy and the corresponding standard deviation. It can be seen that with the increase of K, the performance increases and approaches to the best point at around K = 45. However, the accuracy descends when K becomes larger. The other datasets have the similar properties, and the optimal K for each dataset is given in the caption of the corresponding Tables (III∼VI), where the results are the average results by running the lexicons for three times.

Authorized licensed use limited to: University of the Cumberlands. Downloaded on July 24,2021 at 03:32:52 UTC from IEEE Xplore. Restrictions apply.

DENG et al.: SENTIMENT LEXICON CONSTRUCTION WITH HIERARCHICAL SUPERVISION TOPIC MODEL 715

TABLE IX TOP 20 SENTIMENT WORDS LIST ON SEMEVAL13A WORDS

TABLE X DETECTING DOMAIN-INDEPENDENT SENTIMENT WORDS ON SEMEVAL13A

Fig. 4. Effect of topic size on TaSL in terms of Acc on SemEval13A.

G. Effect of Supervision Information

In this section, we evaluate the effect of supervision infor- mation on hyper-parameter α. As described above, in order to capture the sentiment distribution of documents, we set the parameter α according to the sentiment labels of training docu- ments. Take O = 2 as an example, α can be set as

α(z ,e)y m = η e y m

|e ∈ {1, 2} η1 = (ε + π, ε), if the document is positive

η2 = (ε, ε + π), if the document is negative

Obviously, the value of π determines the difference gap between positive and negative polarities. Thus, we focus on the effect of π on the final classification performance in terms of accuracy (Acc). π = 0 indicates that we did not add any supervision information about documents’ sentiment labels, while π > 0 means that the documents’ sentiment information guides the learning process.

Fig. 5 shows the effect of π. It can be seen that the proposed models perform bad without the aid of supervision information, i.e., π = 0. Note that TaSL-II-prior is better than TaSL-II in this case, because even though there is no documents’ sentiment information, the words’ sentiment information (extracted from manual lexicon MPQA) is used in TaSL-II-prior to supervise the learning process. Fortunately, with the increase of π, the proposed models get better and better performances until they approach to the stable point.

H. Convergency Analysis of TaSL

In order to perform the convergence analysis of the pro- posed algorithm to solve the TaSL-type models, we conduct

Authorized licensed use limited to: University of the Cumberlands. Downloaded on July 24,2021 at 03:32:52 UTC from IEEE Xplore. Restrictions apply.

716 IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 27, NO. 4, APRIL 2019

Fig. 5. Effect of supervision information on α in terms of Acc on SemEval13A.

Fig. 6. Convergency analysis of TaSL and ssLDA in terms of Acc on SemEval13A.

experiments to compare the convergence rate with ssLDA model on SemEval13A dataset. Here ssLDA is used as a base- line because both TaSL-type and ssLDA lexicons are gener- ated by an iterative algorithm with collapsed gibbs sampling technique.

Fig. 6 demonstrates relationship between the sentiment classi- fication accuracy and the iterations obtained by TaSL-II, TaSL- II-prior and ssLDA. The results show that TaSL-type almost converges in the first several steps. Meanwhile, it converges faster and more stable than ssLDA. Again, the performance of TaSL-type lexicon at the stable point is better than that of ssLDA lexicon. This result further indicates that the proposed method is effective and efficient in real applications.

VI. CONCLUSION AND FUTURE WORK

In this paper, we propose an effective supervision topic- adaptive model (TaSL) to explore the sentiment polarity of

words in text data. TaSL is able to simultaneously model the topics and sentiments under the supervision of documents’ sen- timent and words’ prior sentiment information. This generative model not only captures the sentiment attribute of word under different topics, but also constructs the topic-adaptive sentiment lexicon which is helpful to improve the lexicon-based senti- ment classification performance. A series of experiments on four real-world datasets have shown that the proposed TaSL-type sentiment lexicon consistently outperforms the state-of-the-art semantic lexicons.

In the future work, we still have much work to do. Firstly, because all the datasets used in this paper are already anno- tated and labelled, we’re also interested to take advantage of the unlabeled data for making up the lack of supervision in- formation. For example, we can design semi-supervised model to construct a sentiment lexicon which can fully utilize the la- beled and unlabeled data or automatically construct a sentiment lexicon unsupervised.

Secondly, imbalance is a challenging problem in sentiment lexicon construction. In the experiment, existing methods, in- cluding TaSL and all baselines, are performing well in both positive and negative class in balanced datasets, MR and OMD but not that good in imbalanced datasets, SemEval13A and Se- mEval16B. Meanwhile, negation also have an impact on the sentiment lexicon. Thus, we’re interested to solve this problem in our future work.

Thirdly, classifying documents or sentences into positive, neutral, and negative is more piratically. Thus, we want to do multi-class sentiment classification in our future work.

Finally, we’re also interested to convert a sentiment lexicon from source domain to target domain. In this way, we can utilize large source domain data,but a relatively less target domain data.

REFERENCES

[1] B. Liu, “Sentiment analysis and opinion mining,” Synthesis Lectures Hu- man Lang. Technol., vol. 5, no. 1, pp. 1–167, 2012.

[2] Y. Dang, Y. Zhang, and H. Chen, “A lexicon-enhanced method for sen- timent classification: An experiment on online product reviews,” IEEE Intell. Syst., vol. 25, no. 4, pp. 46–53, Jul./Aug. 2010.

[3] W. Medhat, A. Hassan, and H. Korashy, “Sentiment analysis algorithms and applications: A survey,” Ain Shams Eng. J., vol. 5, no. 4, pp. 1093– 1113, 2014.

[4] F. A. Pozzi, E. Fersini, E. Messina, and B. Liu, Sentiment Analysis in Social Networks. San Mateo, CA, USA: Morgan Kaufmann, 2016.

[5] L. Kaushik, A. Sangwan, and J. H. L. Hansen, “Automatic sentiment detection in naturalistic audio,” IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 25, no. 8, pp. 1668–1679, Aug. 2017.

[6] Z. Wang, S. Y. M. Lee, S. Li, G. Zhou, “Emotion analysis in code-switching text with joint factor graph model,” IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 25, no. 3, pp. 469–480, Mar. 2017.

[7] J. Wiebe and E. Riloff, “Creating subjective and objective sentence clas- sifiers from unannotated texts,” in Proc. Int. Conf. Intell. Text Process. Comput. Linguistics, 2005, pp. 486–497.

[8] E. Riloff and J. Wiebe, “Learning extraction patterns for subjective expres- sions,” in Proc. Conf. Empirical Methods Natural Lang. Process., 2003, pp. 105–112.

[9] D. Tang, B. Qin, F. Wei, L. Dong, T. Liu, and M. Zhou, “A joint segmen- tation and classification framework for sentence level sentiment classifi- cation,” IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 23, no. 11, pp. 1750–1761, Nov. 2015.

[10] S. Kiritchenko, X. Zhu, and S. M. Mohammad, “Sentiment analysis of short informal texts,” J. Artif. Intell. Res., vol. 50, pp. 723–762, 2014.

Authorized licensed use limited to: University of the Cumberlands. Downloaded on July 24,2021 at 03:32:52 UTC from IEEE Xplore. Restrictions apply.

DENG et al.: SENTIMENT LEXICON CONSTRUCTION WITH HIERARCHICAL SUPERVISION TOPIC MODEL 717

[11] P. Stone, D. C. Dunphy, M. S. Smith, and D. M. Ogilvie, “The general inquirer: A computer approach to content analysis,” J. Reg. Sci., vol. 8, no. 1, pp. 113–116, 1968.

[12] T. Wilson, J. Wiebe, and P. Hoffmann, “Recognizing contextual polarity in phrase-level sentiment analysis,” in Proc. Conf. Human Lang. Technol. Empirical Methods Natural Lang. Process., 2005, pp. 347–354.

[13] A. Esuli and F. Sebastiani, “SentiWordNet: A publicly available lexical resource for opinion mining,” in Proc. 5th Conf. Lang. Resour. Eval., 2006, pp. 417–422.

[14] L. Barbosa and J. Feng, “Robust sentiment detection on twitter from biased and noisy data,” in Proc. 23rd Int. Conf. Comput. Linguistics, Posters, 2010, pp. 36–44.

[15] L. Gatti, M. Guerini, and M. Turchi, “SentiWords: Deriving a high pre- cision and high coverage lexicon for sentiment analysis,” IEEE Trans. Affective Comput., vol. 7, no. 4, pp. 409–421, Oct.–Dec. 2016.

[16] J. Kamps, M. Marx, R. J. Mokken, and M. De Rijke, “Using WordNet to measure semantic orientations of adjectives,” in Proc. 4th Int. Conf. Lang. Resour. Eval., 2004, vol. 4, pp. 1115–1118.

[17] S. Baccianella, A. Esuli, and F. Sebastiani, “SentiWordNet 3.0: An en- hanced lexical resource for sentiment analysis and opinion mining,” in Proc. 7th Int. Conf. Lang. Resour. Eval., May 2010, pp. 2200–2204.

[18] S. M. Mohammad, S. Kiritchenko, and X. Zhu, “NRC-Canada: Building the state-of-the-art in sentiment analysis of tweets,” Atlanta, GA, USA, 2013, p. 321.

[19] D. Tang, F. Wei, B. Qin, M. Zhou, and T. Liu, “Building large-scale twitter- specific sentiment lexicon: A representation learning approach,” in Proc. 25th Int. Conf. Comput. Linguistics, 2014, pp. 172–182.

[20] J. Blitzer, M. Dredze, and F. Pereira, “Biographies, Bollywood, boom- boxes and blenders: Domain adaptation for sentiment classification,” in Proc. 45th Annu. Meeting Assoc. Comput. Linguistics, vol. 7, 2007, pp. 440–447.

[21] H. Saif, M. Fernandez, L. Kastler, and H. Alani, “Sentiment lexicon adaptation with context and semantics for the social web,” Semantic Web, vol. 8, no. 5, pp. 643–665, 2017.

[22] S. Deng, A. P. Sinha, and H. Zhao, “Adapting sentiment lexicons to domain-specific social media texts,” Decis. Support Syst., vol. 94, pp. 65– 76, 2017.

[23] S. Feng, J. S. Kang, P. Kuznetsova, and Y. Choi, “Connotation lexicon: A dash of sentiment beneath the surface meaning,” in Proc. Assoc. Comput. Linguistics, 2013, pp. 1774–1784.

[24] A. Andreevskaia and S. Bergler, “When specialists and generalists work together: overcoming domain dependence in sentiment tagging,” in Proc. Assoc. Comput. Linguistics, 2008, pp. 290–298.

[25] C. S. Khoo and S. B. Johnkhan, “Lexicon-based sentiment analysis: Com- parative evaluation of six sentiment lexicons,” J. Inf. Sci., pp. 0165–5515, 2017.

[26] M. Hu and B. Liu, “Mining and summarizing customer reviews,” in Proc. 10th ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining, 2004, pp. 168–177.

[27] A. Neviarouskaya, H. Prendinger, and M. Ishizuka, “SentiFul: A lexicon for sentiment analysis,” IEEE Trans. Affective Comput., vol. 2, no. 1, pp. 22–36, Jan.–Jun. 2011.

[28] F. Bravo-Marquez, E. Frank, and B. Pfahringer, “Positive, negative, or neutral: Learning an expanded opinion lexicon from emoticon- annotated tweets,” in Proc. 23rd Int. Joint Conf. Artif. Intell., 2015, pp. 1229–1235.

[29] B. Pang, L. Lee, and S. Vaithyanathan, “Thumbs up?: Sentiment clas- sification using machine learning techniques,” in Proc. Conf. Empirical Methods Natural Lang. Process., 2002, vol. 10, pp. 79–86.

[30] D. T. Vo and Y. Zhang, “Don’t count, predict! An automatic approach to learning sentiment lexicons for short text,” in Proc. 54th Annu. Meeting Assoc. Comput. Linguist., 2016, pp. 219–224.

[31] L. Wang and R. Xia, “Sentiment lexicon construction with representa- tion learning based on hierarchical sentiment supervision,” in Proc. Conf. Empirical Methods Natural Lang. Process., 2017, pp. 502–510.

[32] Q. Li and S. Shah, “Learning stock market sentiment lexicon and sentiment-oriented word vector from stocktwits,” in Proc. 21st Conf. Com- put. Natural Lang. Learn., Aug. 2017, pp. 301–310.

[33] M. Li, Q. Lu, Y. Long, and L. Gui, “Inferring affective meanings of words from word embedding,” IEEE Trans. Affective Comput., vol. 8, no. 4, pp. 443–456, Oct.–Dec. 2017.

[34] E. Cambria, S. Poria, D. Hazarika, and K. Kwok, “SenticNet 5: Discov- ering conceptual primitives for sentiment analysis by means of context embeddings,” in Proc. 32th Int. Conf. Assoc. Adv. Artif. Intell., 2018, pp. 1795–1802.

[35] Y. Wang, Y. Zhang, and B. Liu, “Sentiment lexicon expansion based on neural PU learning, double dictionary lookup, and polarity association,” in Proc. Conf. Empirical Methods Natural Lang. Process., 2017, pp. 553– 563.

[36] Y. Choi and C. Cardie, “Adapting a polarity lexicon using integer linear programming for domain-specific sentiment classification,” in Proc. Conf. Empirical Methods Natural Lang. Process., 2009, pp. 590–598.

[37] Y. Lu, M. Castellanos, U. Dayal, and C. Zhai, “Automatic construction of a context-aware sentiment lexicon: An optimization approach,” in Proc. 20th Int. Conf. World Wide Web., 2011, pp. 347–356.

[38] R. Xie and C. Li, “Lexicon construction: A topic model approach,” in Proc. Int. Conf. Syst. Inform., May 2012, pp. 2299–2303.

[39] D. M. Blei, A. Y. Ng, and M. I. Jordan, “Latent Dirichlet allocation,” J. Mach. Learn. Res., vol. 3, no. Jan, pp. 993–1022, 2003.

[40] M. Yang, D. Zhu, M. Rashed, and K.-P. Chow, “Learning domain-specific sentiment lexicon with supervised sentiment-aware LDA,” Frontiers Artif. Intelligence Appl., 2014, pp. 927–932.

[41] M. Taboada, J. Brooke, M. Tofiloski, K. Voll, and M. Stede, “Lexicon- based methods for sentiment analysis,” Comput. Linguistics, vol. 37, no. 2, pp. 267–307, 2011.

[42] D. Tang, F. Wei, N. Yang, M. Zhou, T. Liu, and B. Qin, “Learning sentiment-specific word embedding for twitter sentiment classification,” in Proc. Assoc. Comput. Linguistics, 2014, pp. 1555–1565.

[43] D. A. Shamma, L. Kennedy, and E. F. Churchill, “Tweet the debates: Understanding community annotation of uncollected sources,” in Proc. 1st SIGMM Workshop Social Media, 2009, pp. 3–10.

[44] P. Nakov et al., “SemEval-2013 task 2: Sentiment analysis in twitter,” in Proc. 7th Int. Workshop Semantic Eval., Jun. 2013, pp. 312–320.

[45] P. Nakov, A. Ritter, S. Rosenthal, F. Sebastiani, and V. Stoyanov, “SemEval-2016 task 4: Sentiment analysis in twitter,” in Proc. 10th Int. Workshop n Semantic Eval., 2016, pp. 1–18.

[46] M. Yang, B. Peng, Z. Chen, D. Zhu, and K.-P. Chow, “A topic model for building fine-grained domain-specific emotion lexicon,” in Proc. 52nd Annu. Meeting Assoc. Comput. Linguistics, 2014, pp. 421–426.

[47] J. Tang, Z. Meng, X. Nguyen, Q. Mei, and M. Zhang, “Understanding the limiting factors of topic modeling via posterior contraction analysis,” in Proc. Int. Conf. Mach. Learn., 2014, pp. 190–198.

[48] W. Wunnasri, T. Theeramunkong, and C. Haruechaiyasak, “Solving un- balanced data for Thai sentiment analysis,” in Proc. 10th Int. Joint Conf. Comput. Sci. Softw. Eng., May 2013, pp. 200–205.

Dong Deng received the B.S. degree in computer science, in 2014, from the School of Computer and Information Technology, Beijing Jiaotong University, Beijing, China, where he is currently working toward the Ph.D. degree. His research interests include sen- timent analysis, data mining, and machine learning.

Liping Jing received the Ph.D. degree in applied mathematics from the University of Hong Kong, Hong Kong, in 2007. From 2007 to 2008, she was a Research Associate with the Department of Com- puter Science, University of Texas at Dallas, Dallas, TX, USA, and from 2015 to 2016, a Visiting Scholar with ICSI and the AMPLab, University of California at Berkeley, Berkeley, CA, USA. She is currently a Professor with the Beijing Key Laboratory of Traffic Data Analysis and Mining, School of Computer and Information Technology, Beijing Jiaotong University,

Beijing, China. She has authored or coauthored more than 70 peer-reviewed re- search papers in various journals and conferences. Her research interests include high-dimensional subspace learning and machine learning. She was a Regular Reviewer and Program Committee Member for a number of international jour- nals and conferences.

Authorized licensed use limited to: University of the Cumberlands. Downloaded on July 24,2021 at 03:32:52 UTC from IEEE Xplore. Restrictions apply.

718 IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 27, NO. 4, APRIL 2019

Jian Yu received the B.S. and M.S. degrees in math- ematics and the Ph.D. degree in applied mathemat- ics from Peking University, Beijing, China, in 1991, 1994, and 2000, respectively. He is currently a Profes- sor with Beijing Jiaotong University, Beijing, China, where he is also the Director of the Beijing Key Lab- oratory of Traffic Data Analysis and Mining. His cur- rent research interests include machine learning, im- age processing, and pattern recognition.

Shaolong Sun received the Ph.D. degree majoring in management science and engineering from the Institute of Systems Science, Academy of Mathe- matics and Systems Sciences, Chinese Academy of Sciences, Beijing, China. He is currently a Research Assistant with the Department of Systems Engineer- ing and Engineering Management, City University of Hong Kong, Hong Kong. His research interests in- clude artificial intelligence, big data mining, machine learning, social networks analysis, knowledge man- agement, and economic and financial forecasting. He

has authored or coauthored more than ten papers in journals including Applied Energy and Journal of Environmental Management.

Michael K. Ng received the B.Sc. and M.Phil. de- grees from the University of Hong Kong, Hong Kong, in 1990 and 1992, respectively, and the Ph.D. degree from the Chinese University of Hong Kong, Hong Kong, in 1995. He is currently a Chair Professor with the Department of Mathematics, Hong Kong Baptist University, Hong Kong. His research interests include bioinformatics, image processing, scientific comput- ing, and data mining. He is on the editorial boards of international journals.

Authorized licensed use limited to: University of the Cumberlands. Downloaded on July 24,2021 at 03:32:52 UTC from IEEE Xplore. Restrictions apply.