Sentimental Analysis of Web Financial Reviews
Opportunities and Challenges
Changxuan Wan, Tengjiao Jiang, Dexi Liu and Guoqiong Liao
School of Information and Technology, Jiangxi University of Finance and Economics, Nanchang, China
Jiangxi Key Laboratory of Data and Knowledge Engineering, Jiangxi University of Finance and Economics,
Nanchang, China
Keywords: Web Financial Reviews, Web Financial Indexes, Opinion Targets, Sentiment Words, Sentiment Analysis.
Abstract: Web financial reviews are real-time, comprehensive and authentic. The construction and quantification of
Web financial indexes based on Web financial reviews is of great significance for the financial early
warning for enterprises. Comparing with product reviews and news commentaries, in Web financial reviews,
the opinion targets have more diverse compositions, the frequencies of opinion targets’ occurrence vary
greatly, and the sentiment words’ have more diverse parts of speech. These characteristics make the
extraction of opinion targets, the construction of Web financial indexes, and opinion targets-based
sentimental analysis all more complicated, posing new challenges to natural language processing.
1 INTRODUCTION
With the aggravation of the global financial
market’s instability, more attention is paid to the
financial crisis prediction for enterprises. Currently,
in most studies, financial crisis prediction is done by
a prediction model established based on the data in
financial statements.
However, there are several drawbacks in using
financial statements. First, financial statements are
easily manipulated; second, data in financial statements
are static, ignoring the characteristic of time series
of enterprise financial ratios; thirdly, financial
statements are released yearly, the data in them are
not real-time; and fourthly, the influence of historic
accumulation of financial ratios on present situation is
not considered. Therefore, determining risks merely
based on financial indicators in financial statements
is doomed to cause deviation of the prediction.
According to the theory of modern corporation
performance evaluation, in the era of knowledge
economy, comprehensive performance evaluation
equation must be introduced into the performance
evaluation of corporations; various non-financial
indicators should be added on the basis of conventional
financial indicators.
Comparing with financial indicators, the biggest
advantage of non-financial indicators is that they are
future-oriented, while financial indicators are past-
oriented. In The Choice of Performance Measures in
Annual Bonus Contracts published by Wharton in
1995, it is pointed out that non-financial indicators
are better indicators reflecting the management’s
performance and the company’s development
prospects. Unlike financial indicators, non-financial
indicators can not be obtained by calculating
financial data, so the non-financial index also has
some shortcomings, such as data collection and
quantify are difficult. Then, how should we
overcome the aforementioned drawbacks of
financial indicators and non-financial indicators and
introduce non-financial indicators on the basis of
financial indicators to establish a uniform indicator
evaluation system?
With the arrival of the era of big data, large
amount of financial reviews occur on the Internet
every minute. Those reviews cover all aspects of
enterprises’ past and current operation and prospects
of enterprises, and contain thorough analyses of
enterprises’ financial and non-financial indicators.
Therefore, not only are Web financial reviews real-time
and readily available, they are also comprehensive
and have wide coverage. In addition, Web financial
reviews contain expert interpretation, practical
experience of investors, as well as customer experience,
thus are in-depth and authentic. These characteristics
of Web financial reviews, on the one hand, overcome
the drawbacks of indicators from financial statements,
366
Wan C., Jiang T., Liu D. and Liao G..
Sentimental Analysis of Web Financial Reviews - Opportunities and Challenges.
DOI: 10.5220/0005137403660373
In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2014), pages 366-373
ISBN: 978-989-758-048-2
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
and on the other hand, reveal the influences of
financial and non-financial indicators on enterprises’
future development. By classifying the opinion
targets extracted from Web-financial reviews, the
financial and non-financial indicators in the reviews
can be obtained, and we call them the Web
financial indexes.
Being real-time, comprehensive, and in-depth, Web
financial reviews make possible comprehensiveness and
systematicness of financial and non-financial indicators
in enterprise financial early warning models. In addition,
the sentiment inclination of the reviews makes
possible quantification of Web financial indexes.
The extraction and sentimental quantification of Web
financial indexes based on Web financial reviews is a
meaningful, yet extraordinarily challenging project.
1) Extraction of opinion targets in Web financial
reviews and construction of Web financial indexes
The opinion targets refer to the objects modified
by the evaluative words. For example, the words
component, function and service of a product in
product reviews, and the words people, event and
subject of conversation in news commentaries are all
opinion targets. Groups of opinion targets constitute
topics, for example, the opinion targets in product
reviews can be classified into product group. In Web
financial reviews, opinion targets can be a national
policy, a sub-item in financial statements, or a
subject of conversation. The grouping of opinion
targets results in Web financial indexes. Now
available studies mostly focus on the extraction and
grouping of opinion targets in product reviews, and
by comparison, the extraction and classification of
opinion targets in Web financial reviews are much
more complicated because those reviews involve
wide range of areas.
(1) The opinion targets in product reviews are
generally nouns or noun phrases, such as 'Apple',
'screen' and 'keyboard layout' in a cellphone review.
In financial reviews, in addition to being nouns or
noun phrases such as 'raw material' and 'stock price',
opinion targets can also be subordinate clauses. For
example, in the sentence 'Share price rise quickly is
good.', the opinion target of the sentiment word
'good' is a verb phase ' share price rise quickly '.
Therefore, the extraction of opinion targets from
Web financial reviews is more complicated.
(2) In product reviews, opinion targets are more
evenly distributed. For instance, in a cellphone review,
the user would generally comment on the appearance,
audio, image display, etc. Financial reviews may
contain interpretation of financial statements,
deciphering of macro policies, and analysis of
personnel movement. The different numbers of
comments in different categories lead to very different
frequencies of opinion targets’ occurrence.
Consequently, the construction of Web financial
indexes based on opinion targets grouping is much
more complicated.
2) Quantification of Web financial indexes
In opinion target-based sentimental analysis, the
sentiment value of each opinion target is first
calculated based on the sentiment phrase, and then
the opinion target is classified into corresponding
topic/indicator based on the grouping of opinion
targets.
In product reviews, sentiment words are usually
adjectives, and available studies mostly perform
sentimental analysis based on adjective sentiment
words. Different from product reviews, Web financial
reviews contain sentiment words that have more
diverse parts of speech. Besides being adjective,
those sentiment words may be verb, adverb or noun,
especially verb. For example, in the previous example
‘Share price rise quickly is a good thing’, the word
‘rise’ is a verb sentiment word, and the phrase ‘good
thing’ is a noun sentiment word. The diversity of
sentiment words’ parts of speech in financial
reviews makes the identification of sentiment words
and the calculation of those words’ polarity and
intensity more difficult. In addition, this diversity
results in more flexible components the sentiment
words serve as in sentences, thus the sentiment
word-based extraction of opinion targets is also
more difficult.
The diversity of opinion targets’ composition,
differences in opinion targets’ frequencies, and
richness of sentiment words’ parts of speech in Web
financial reviews make the extraction of opinion
targets, the construction of Web financial indexes, as
well as the opinion target-based sentimental analysis
all more complicated, and bring new challenges to
natural language processing.
2 RELEVANT STUDIES
2.1 Relation between Web Financial
Reviews and Enterprises’ Financial
Statuses
The first research on Web financial reviews was
done by Wysocki (1999). After investigating the 50
listed companies that had the greatest amounts of
information during Jan 1998 to Aug 1998, Wysocki
noticed that with information from the notice board,
the trading volume and abnormal stock returns of
SentimentalAnalysisofWebFinancialReviews-OpportunitiesandChallenges
367
the next day can be predicted. Das and Chen (2007),
after exploring information from Yahoo, Amazon and
other forums, found that the information contain
contents significantly correlated with the return on
assets. Tetlock et al. (2008) conducted research on
the relationship between the negative words in news
reports about S&P500 companies from 1980 to 2004
and the companies’ profits and stock returns, and
discovered that the negative words in news reports
about listed companies can be used for the
prediction of those companies’ future lower profits.
Si et al. (2013) reported that the topic-based public
sentiments in Web financial reviews could help
improve the accuracy of stock price prediction. The
study of Bian et al. (2013) revealed that the result of
sentimental analysis of Web financial information can
be used as important indicator for the financial early
warning for listed companies.
The literature discussed above reveals the influence
of the amount, popularity and content of Web financial
reviews on investors as well as their reflection in
stock trading. However, available studies, when
performing mining of Web financial reviews, only
pay attention to sentimental polarity at document
level, or roughly count the numbers of positive and
negative sentiment words in a document to determine
the document’s polarity. In fact, Web financial
reviews contain far more information. In general,
every document contains multi-facet (financial
indicators and non-financial indicators) interpretation
of an enterprise. After determining the sentimental
polarity of each financial or non-financial indicator,
the sentiment value of the indicator can be applied to
financial early warning models in order to perform
more detailed, accurate, and refined analysis and
prediction. The construction and quantification of
Web financial indexes are detailed in the next few
sub-sections in three aspects, the extraction of
targets, extraction of target-sentiment word pairs,
and target-based sentimental analysis of texts.
2.2 Extraction of Targets
The opinion targets in product reviews are also called
aspects or attributes. For the extraction of opinion
targets, there are mainly three types of methods,
which are methods based on frequent nouns and
rules-based methods, machine learning-based methods,
and topic model-based methods.
1) Frequent nouns and rules-based methods
The frequent noun and rule-based methods
generally extract opinion targets based on the
following heuristic rule: the opinion targets in
product reviews are generally noun or noun phrase,
extract these words first, and then use the opinion
target-sentiment word relation to extract new opinion
targets and sentiment words. Hu et al. (2004)
proposed to, based on massive corpuses of a certain
field, first identify nouns based on part of speech
labeling, and then use Apriori algorithm to find
frequent nouns or noun phrases as the opinion
targets. On the basis of this study of Hu et al.,
Popescu et al. (2005) improved the accuracy of the
algorithm by further filtering the nouns or noun
phrases. They proposed to determine whether a noun
or noun phrase is an opinion target by calculating
the pointwise mutual information (PMI) between the
noun or noun phrase and the classification of the
opinion target to be extracted. This method has
improved opinion target extraction accuracy, yet
somewhat decreased recall rate. Moghaddam et al.
(2010) tried to determine the occurrence pattern of
product features based on standard product features
defined in fine-grained product reviews, and filter
high-frequency words based on the hit rates of
high-frequency nouns or noun phrases in the pattern.
2) Machine learning-based methods
Some other researchers applied machine learning
to the identification of opinion targets. Jakob et al.
(2010) used conditional random field (CRF) model to
extract opinion targets. Jin et al. (2009) treated the
extraction of characteristic words and sentiment
words as a sequence labeling task, where each word
in a review corresponds to a category label, and
proposed to adopt lexicalized hidden Markov model
(HMM) to search for the most possible label
sequence. Su et al. (2008) proposed a novel mutual
strengthening criterion for the mining of hidden
association between opinion targets and sentiment
words, and to identify hidden opinion targets based
on clustering.
3) Topic model-based methods
In recent years, with topic model gradually
becomes popular, scholars are applying it to the field
of sentimental analysis. It is used for identification
and classification of opinion targets.
Titov et al. (2008) found that standard latent
Dirichet allocation (LDA) model is not suitable for
extraction of fine-grained opinion targets. They
proposed a multi-grain latent Dirichlet allocation
(MG-LDA) model and a multi-aspect sentiment
(MAS) model, which are able to discover not only
general opinion targets, but also fine-grained
opinion targets. Andrzejewski et al. (2009) first
proposed DF-LDA model with constraints. They
introduced two types of constraints, must-link and
cannot link, as priori knowledge. However, as the
number of documents increases, the computational
KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
368
complexity of this model increases exponentially.
Zhai et al. (2011) proposed a constrained-LDA
model for the extraction and classification of
opinion targets from product reviews. They also set
two types of constraints, must-link and cannot-link;
the former classifies opinion targets with the same
composition to the same topic, while the latter
classifies opinion targets in the same sentence to
different topics. Moghaddam et al. (2013) presented
a factorized LDA (FLDA) model for cold start items,
which models opinion targets and reviewers at the
same time, and performs opinion target extraction
and rating on the basis of opinion target
classification.
The methods based on frequent nouns and rules
may cause omitting of some opinion targets that
occur at lower frequencies. In addition, not all
frequent nouns are opinion targets. With the
machine learning-based methods, training sets need
to be labeled manually, and data sets from different
fields have poor transferability. The topic
model-based methods cluster opinion targets with
similar semantics to the same topic, and are able to
explain the membership degrees of individual words
to a topic, thus well serve the purpose of opinion
target extraction and classification.
2.3 Extraction of Opinion
Target- Sentiment Word Pairs
There are two types of methods for the extraction of
opinion target-sentiment word pairs, machine learning
based methods and syntactic rule-based methods.
1) Machine Learning-based Methods
Jin et al. (2009) considered the extraction of
product features and sentiment words as a sequence
labeling task, and proposed to label the most
possible label sequence based on hidden Markov
model, and then use the sequence to identify opinion
targets and sentiment words. Lakkaraju et al. (2011)
proposed to use hidden Markov model to describe
the syntactic dependencies between opinion targets
and sentiment words, and extract product features and
corresponding sentiment words based on contextual
consistency.
2) Syntactic Rules-based Methods
Kobayashi et al. (2004) adopted templates to
express the modification relationship between
opinion targets and sentiment words; they designed
eight templates for this purpose. Bloom et al. (2007)
adopted Stanford Parser and manually constructed
31 syntactic rules to obtain appraisal expressions.
Bloom et al. (2009) used confidence rating method
for the automatic learning of rules. First, all possible
appraisal expressions in a sentence are identified,
then rules are extracted based on these appraisal
expressions, and the extracted rules are eventually
used to match the opinion targets and sentiment
words in the sentence in order to find rules with
higher confidence levels. Qiu et al. (2011) reported a
method called Double Propagation, which performs
identification and extraction of sentiment words and
opinion targets at the same time. Kamal et al. (2012)
designed rules for the extraction of appraisal
expressions on the basis of linguistic and syntactic
analyses of reviews.
The machine learning-based methods generally
treat the extraction of opinion targets and sentiment
words as a sequence labeling task; they often require
manual labeling of training sets. In contrast,
syntactic analysis-based methods extract the
syntactic relationships between opinion targets and
sentiment words at syntactic level instead of field
level, thus demonstrating better adaptability to
different fields.
2.4 Opinion Target-based Sentimental
Analysis of Texts
Methods for opinion Target-based sentimental analysis
could generally be divided into machine learning-
based methods and syntactic analysis-based methods.
1) Machine Learning-based Methods
Wilson et al. (2009) proposed to identify sentiment
words with supervised modification methods. Their
experiment showed that supervised learning classifier
that fuses multiple features can greatly improve the
extraction of opinion targets and sentiment words.
Fang et al. (2012) used latent structural model to
realize the opinion targets in product reviews, and their
method is capable of identifying the opinion targets’
level in the meantime. Liu et al. (2012) realized
fine-grained opinion mining based on word-based
translation model. Lu et al. (2011) proposed a
segmented topic model (STM), which performs joint
modeling of topic distribution at document and
sentence levels with two-parameter Poisson-Dirichlet
process, labels sentences with weak supervision to
strengthen the direct correlation between topic and
aspect/opinion targets, and obtains better multi-aspect
evaluative polarity by combining overall rating and
sentence labeling. Kontopoulos et al. (2013) proposed
an ontology-based dispatching method that effectively
improves Twitter sentimental analysis. This method
allocates sentiment scores to relevant attributes/opinion
targets of each topic, and is thus able to conduct more
detailed sentimental analysis of each specific topic.
These methods generally omit the syntactic
SentimentalAnalysisofWebFinancialReviews-OpportunitiesandChallenges
369
information of the sentences that the sentiment
words are in. This issue has been noticed by some
researchers, who have tried to determine the fine-
grained polarity of texts through syntactic analysis.
2) Syntactic Analysis-based Methods
Feng et al. (2012), based on dependency parsing,
obtained ADV dependency pair in adverbial-verb
structure with sentiment word as the center, and on
this basis, obtained the sentiment values of sentiment
sentences in micro-blogs. Wan et al. (2013) proposed
to determine the sentimental polarity of Web
financial reviews through sentimental analysis of
dependency pairs.
Machine learning-based methods treat the
association between opinion targets and sentiment
words as a sequence label, without taking the
syntactic association between opinion targets and
sentiment words into consideration. At present, the
opinion target-based polarity analysis of Web
financial reviews is not adequately thorough.
3 CHALLENGES IN
SENTIMENTAL ANALYSIS OF
WEB FINANCIAL REVIEWS
AND CORRESPONDING
STRATEGIES
3.1 Extraction of Opinion Targets
The available studies of opinion target extraction
mostly focus on the extraction of opinion targets in
product reviews. They generally limit opinion
targets as nouns or noun phrases and then perform
further identification. As the analysis in section 2.2
indicates, the frequent nouns and rules-based methods
essentially count the occurrence of nouns or noun
phrases. With supervised learning, better results can
be obtained when there is adequate training data.
However, with the rapidly growing information on
the Internet, newly occurred information may not be
labeled and become training corpuses before they
are outdated. Although the various emerging
semi-supervised learning methods are trying to
remedy this defect, iterative learning started from a
seed set will exhibit deviations after large amount of
training, and the consequent manual deviation
rectification and adjustment are massive. In recent
years, statistic topic model is becoming a popular
method for the topic discovery in massive documents.
The advantage of topic model is that besides opinion
targets discovery, it can also perform opinion targets
clustering.
The above analysis shows that in Web financial
reviews, the composition of opinion targets is more
diverse. Instead of being noun or noun phrase,
opinion targets can be verb phrase, verb-object phrase,
or even sentences. Therefore, we define opinion
targets in Web financial reviews as opinion target
expressions. Opinion target expressions can be nouns
or noun phrases, or composed of noun (phrase) and
verb. On the one hand, the nouns in opinion target
expressions are conducive to further classification of
opinion targets, and on the other hand, the verbs in
opinion target expressions help determine the oddity
of opinion targets. An odd opinion target reverses
the polarity of the sentiment words that modify it.
Meanwhile, since in Web financial reviews, the number
of comments on each Web financial indicator varies,
which result in very different frequencies of the
opinion targets. Therefore, direct application of topic
model to the extraction and classification of opinion
targets in Web financial reviews does not work well.
An opinion target refers to the object of
modification of a sentiment word. Dependencies
usually exist between sentiment words and opinion
targets. Therefore, it is possible to extract opinion targets
based on sentiment words and those dependencies.
(1) The diversity of sentiment words’ parts of
speech makes them able to be different structural
components of sentences. In Web financial reviews,
sentiment words may be adjective, verb, adverb or
noun, while opinion targets may be noun, noun
phrase, subject-verb phrase, verb-object phrase, or
sentence. Therefore, the rules of syntactic paths
between sentiment words and opinion targets in Web
financial reviews are far more complex.
(2) When using syntactic paths to extract opinion
targets, we noticed that even when sentiment words
serve as the same component of sentences, the
components their opinion targets serve as may be
different. For example, in the sentence '我看中这家
公司的发展前景(I prefer this company’s
development prospects ).' and the phrase '股价上涨
(Stock price rises).', the verb sentiment words '看中
(prefer)' and '上涨(rise)' both serve as predicates, yet
their opinion targets’ positions are different. The
opinion target of 'prefer' is the object of the sentence,
while the opinion target of 'rise' is the subject of the
sentence. Therefore, on the basis of syntactic rules,
the understanding of semantics should be added. For
example, psychological verbs modify the objects of
the sentences they are in, while non-psychological
verbs modify the objects. A natural question is then
whether it is possible to identify the components that
the adjective sentiment words’ opinion targets serve
KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
370
as based on syntax.
Therefore, it should be considered to extract
opinion targets based on both syntactic paths and
semantics. Such methods are detailed as follows.
1) Machine Learning and Semantic Analysis
Combined Methods
The first step is to label a sentiment word and its
candidate opinion targets. Since the opinion targets
of sentiment words may be subject or object of a
sentence, it is required to label the subject or object
the sentiment word modifies.
The second step is to extract all syntactic paths
(phrase syntax or dependency syntax) between the
sentiment word and all candidate opinion targets.
The third step is to generalize the syntactic paths
and form a syntactic path library by descending
order of the paths.
The last step is to match the sentiment word with
the semantics (synonym or co-occurrence) of subject
sentiment words (or object sentiment words). When
the threshold similarity reaches a certain value, the
subject (or object) is chosen as the sentiment word’s
opinion target. When the threshold similarity is
lower than a certain value, the strategy of ‘syntactic
path library’ plus ‘sentiment word’ is used to
identify the opinion target of the sentiment word.
2) Syntactic Rules and Semantic Analysis
Combined Methods
By combining syntactic rules and semantic analysis,
it can be directly determined whether a sentiment
word’s opinion target is the subject or the object.
A sentiment word can be verb, adjective, or
adverb. These different conditions are analyzed as
follows.
(1) When a sentiment word is a verb, it may be
psychological verb or non-psychological verb. It has
been proved in studies that when a sentiment word is
a psychological verb, it modifies the object. For
example, the opinion targets of psychological words
'喜欢(like)' and '看中(prefer)' are objects. When the
sentiment word is a non-psychological verb, its
opinion target is the subject. For example, in the
sentence '我看中这家公司的发展前景(I prefer this
company’s development prospects)', the opinion
target of sentiment word ' prefer' is the object of the
sentence 'prospects', and it can be further extended
to ' this company’s development prospects.'; And in
the sentence ' 股价上涨(Stock price rises)', the
opinion target of sentiment word 'rise' is the subject
of the sentence, i.e. 'stock price'.
(2) When a sentiment word is an adjective or a
verb, can a pattern about its opinion targets be
summarized? A sentiment word may serve as the
predicate of a sentence, the modifiers of the
predicate, the object, and the object’s modifiers. The
following assumptions are then made:
a) When the sentiment word is the predicate, its
opinion target is the subject;
b) When the sentiment word is a modifier of the
predicate, its opinion target is the subject-predicate
structure;
c) When the sentiment word is the object, its
opinion target is the subject-predicate structure;
d) When the sentiment word is a modifier of the
object, its opinion target is the subject-predicate-
object structure;
e) When the sentiment word is the attribute, its
opinion target is the word it modifies.
3.2 The Construction of Non-financial
Index System
In Web financial reviews, the number of comments
on each Web financial indicator varies, which results
in very different frequencies of opinion targets’
occurrence. For the construction and quantification
of Web financial indicators, low-frequency opinion
targets are often non-financial indicators, yet they
are of great importance in financial early warning
for enterprises. Therefore, direct use of topic model
for the extraction and grouping of the opinion targets
in Web financial reviews would lead to incompleteness
of the constructed Web financial indexes. The idea is
detailed as follows.
(1) Extract the opinion targets with the syntactic
path and semantic analysis combined method
introduced in section 3.1, use these opinion targets
as the aspects of topic model, and then group the
opinion targets with topic model in order to
construct the Web financial indexes.
(2) Classify opinion targets according to different
topics the Web financial reviews describe, for example
the deciphering of macro policies, interpretation of
financial statements, and analysis of corporate culture,
and then construct a hierarchical LDA model for the
hybrid classification of opinion targets of different
topics and frequencies.
3.3 Opinion Target-based Sentimental
Analysis of Web Financial Reviews
In opinion target-based sentimental analysis of Web
financial reviews, since one sentence may contain
multiple sentiment words and opinion targets, it is
needed to calculate the sentiment value of each
opinion target based on the sentiment phrase.
Sentiment phrase is a three-component group in the
SentimentalAnalysisofWebFinancialReviews-OpportunitiesandChallenges
371
form of < opinion target expression, sentiment word,
contextual modifiers of sentiment word>.
1) Influence of Opinion Target Expression on
Polarity of Sentiment Phrase
Opinion targets that are noun or noun phrase
may exhibit oddity, and an odd opinion target will
change the polarity of the sentiment word that
modifies it. For example, the word '减少(decrease)'
generally exhibits negative polarity, and the phrase
'营业收入减少(the operating income decreases)'
exhibits negative polarity, while the phrase '损失减
少(the loss decreases)' exhibits positive polarity.
This is because the word ‘loss’ is an odd target.
For a opinion target composed of noun and verb,
sometimes this verb may also be a sentiment word.
For example, in sentences '股价上涨得很快(the
stock price rises rapidly)' and '股价下跌得很快(the
stock price drops rapidly)', the sentiment word
'rapidly' modifies 'stock price rises' and 'stock price
drops', respectively. At this time, the sentimental
polarity and intensity of the entire opinion target
expression need to be determined first.
2) Influence of Sentiment Word’s Contextual
Modifiers on Polarity of Sentiment Phrase
The contextual modifiers of sentiment words are
mainly negative adverbs and adverbs of degree, and
their influences on the polarity of sentiment phrases
include:
(1) Influence of negative adverbs on polarity of
sentiment words;
(2) Influence of adverbs of degree on polarity of
sentiment words;
(3) The distance between negative adverbs or
adverbs of degree and sentiment words is called edit
distance. When a negative adverb and an adverb of
degree modify the same sentiment word
simultaneously, different combinations of their edit
distances from the sentiment word result in different
influences on the polarity and intensity of the
sentiment word.
4 CONCLUSIONS
The beginning of the era of big data brings us both
opportunities and challenges. Applying data mining
to Web financial reviews, which contain abundant
information, could help with investors’ investment
decision-making, enterprise operators’ management
decision-making, as well as credit rating in the
finance and insurance industry.
However, the mining of Web financial reviews
faces many challenges, for example the diversity of
sentiment words’ parts of speech, the diversity of the
opinion targets expressions, and the complexity of
the construction of Web financial indexes, as well as
the sentimental quantification of Web financial
indexes caused by these three features. In the
meantime, this challenging task is very meaningful.
REFERENCES
Andrzejewski, D., Zhu, X. , and Craven M.(2009).
Incorporating domain knowledge into topic modeling
via Dirichlet forest priors. In ICML’09: Proceedings of
the International Conference on Machine Learning,
Montreal, Quebec, Canade, 14-18 June 2009. ACM,
25-32.
Bian, H., Wan, C., Liu, D., Jiang, T. (2013). A study of
financial crisis prediction model for listed companies
taking into account Web financial information. Computer
Science, 40(11), 295-298, 315. (in Chinese)
Bloom, K., Argamon, S. (2009). Automated learning of
appraisal extraction patterns. Language and
Computers, 71(1), 249-260.
Bloom, K., Garg, N., Argamon, S. (2007). Extracting
appraisal expressions. In EMNLP’07: Proceedings of
the Human Language Technology Conference and the
Conference on Empirical Methods in Natural
Language Processing, Prague, Czech Republic, 28-30
June 2007. ACL, 308-315.
Das, S., Chen, M. (2007). Yahoo! For Amazon: Sentiment
Extraction from Small Talk on the Web. Management
Science, 53(9),1375-1388.
Fang, L., Huang, M. (2012). Fine granular aspect analysis
using latent structural models. In ACL 2012:
Proceedings of the 50th Annual Meeting of the
Association for Computational Linguistics, Jeju,
Korea,8-14 July 2012. ACL, 333-337.
Feng, S., Fu, Y., Yang, F., et al., 2012. Blog sentiment
orientation analysis based on dependency parsing.
Journal of Computer Research and Development,
49(11), 2395-2406. (in Chinese)
Hu, M., Liu, B. (2004). Mining and summarizing
customer reviews. In KDD’04: Proceedings of the 10th
ACM International Conference on Knowledge
Discovery and Data Mining, Seattle, Washington,
USA, 22-25 August 2004. New York: ACM, 168-177.
Jakob, N., Gurevych, I. (2010). Extracting opinion targets
in a single and cross-domain setting with conditional
random fields. In EMNLP’10: Proceedings of
Conference on Empirical Methods in Natural
Language Processing. Cambridge, MA, 12-14
October 2010. ACL,1035-1045.
Jin, W., Ho, H., Srihari, R. (2009). OpinionMiner: a novel
machine learning system for Web opinion mining and
extraction. In KDD’09: Proceedings of ACM
International Conference on Knowledge Discovery and
Data Mining (SIGKDD), Paris, 19-21 September 2009.
New York: ACM, 1195-1204.
Kamal, A., Abulaish, M., Anwar, T. (2012). Mining
feature-opinion pairs and their reliability scores from
KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
372
Web opinion sources. In WIMS’12: Proceedings of the
2nd International Conference on Web Intelligence,
Mining and Semantics，Craiova, Romania, 13-15 June
2012. ACM,15.
Kobayashi, N., Inui, K., Matsumoto, Y., Tateishi K., and
Fukushima T. (2004). Collecting evaluative
expressions for opinion extraction. In IJCNLP’04:
Proceedings of the International Joint Conference on
Natural Language Processing, Hainan Island, China,
22-24 March 2004. Vol.3248, Springer, 596-605.
Kontopoulos, E., Berberidis, C., Dergiades, T., Bassiliades,
N. (2013). Ontology-based sentiment analysis of
Twitter posts. Expert Systems with Applications,
40(10), 4065-4074.
Lakkaraju, H., Bhattacharyya, C., Bhattacharya, I.,
Merugu, S.(2011). Exploiting coherence for the
simultaneous discovery of latent facets and associated
sentiments. In SDM’11: Proceedings of 2011 SIAM
International Conference on Data Mining, Mesa,
Arizona USA, 28-30 April 2011. SIAM, 498-509.
Liu, K., Xu, L., Zhao, J. (2012). Opinion target extraction
using word-based translation model. In EMNLP’12:
Proceedings of the Conference on Empirical Methods
in Natural Language Processing and Computational
Natural Language Learning, Jeju, Korea,8-14 July
2012. ACL, 1346-1356.
Lu, B., Ott, M., Cardie, C., Tsou, B. (2011). Multi-aspect
sentiment analysis with topic models. In ICDM’11:
Proceedings of the 11th IEEE International
Conference on Data Mining Workshops, Vancouver,
Canada, 11-14 December 2011. IEEE,81-88.
Moghaddam, S., Ester, M.(2013). The FLDA model for
aspect-based opinion mining: addressing the cold start
problem. In WWW’13: Proceedings of International
Conference on World Wide Web, Rio, 13-17 May 2013.
ACM, 909-918.
Moghaddam, S., Ester, M. (2010). Opinion digger: an
unsupervised opinion miner from unstructured product
reviews. In CIKM’10: Proceedings of the 19th ACM
International Conference on Information and Knowledge
Management, Toronto, Canada, 26-30 October 2010
ACM, 1825-1828.
Popescu, A., Etzioni, O. (2005). Extracting product
features and opinions from reviews. In
HLT/EMNLP’05: Proceedings of the Human
Language Technology Conference and the Conference
on Empirical Methods in Natural Language Processing,
Vancouver, Canada, 6-8 October 2005. ACL, 339-346.
Qiu, G., Liu, B., Bu, J., Chen, C. (2011). Opinion word
expansion and target extraction through double
propagation. Computational Linguistics, 37(1), 9-27.
Si J., Mukherjee A., Liu, B., Li Q., Li H., Deng X. (2013).
Exploiting topic based Twitter sentiment for stock
prediction. In: Proceedings of the 51st Annual
Meeting of the Association for Computational
Linguistics, Sofia, Bulgaria, 4-9 August 2013. ACL,
24-29.
Su, Q., Xu, X., Guo, H., Guo Z., Wu X., Zhang X., Swen
B., and Su Z. (2008). Hidden sentiment association in
Chinese Web opinion mining. In WWW’08:
Proceedings of the 17th International Conference on
World Wide Web, Beijing, 21-25 April 2008. ACM,
959-968.
Tetloek, P., Saar-Tsechansky, M., Macskassy, S. (2008).
More Than Words: Quantifying Language to Measure
Firms’ Fundamentals. The Journal of Finance,
63(3),1437-1467.
Titov, I., McDonald, R. (2008). Modeling online reviews
with multi-grain topic models. In WWW’08:
Proceedings of the 17th International Conference on
World Wide Web, Beijing, 21-25 April 2008. ACM,
p111-120.
Wan, C., Jiang, T., Zhong, M., Bian, H., 2013. Sentiment
computing of Web financial information based on the
part-of-speech tagging and dependency parsing.
Journal of Computer Research and Development,
50(12), 2554-2569. (in Chinese)
Wilson, T., Wiebe, J., Hoffmann, P. (2009). Recognizing
contextual polarity: an exploration of features for
phrase-level sentiment analysis. Computational
Linguistics, 35(1), 399-433.
Wysocki, P. (1999). Cheap talk on the Web: the
determinants of postings on stock message boards.
Working Paper, University of Michigan Business
School. No.98025. November 1999.
Zhai, Z., Liu, B., Xu, H., Jia, P.(2011). Constrained LDA
for grouping product features in opinion mining. In
PAKDD’11: Proceedings of Pacific-Asia Conference on
Knowledge Discovery and Data Mining, Shenzhen,
China, 24-27 May 2011. Springer Berlin Heidelberg,
448-459.
SentimentalAnalysisofWebFinancialReviews-OpportunitiesandChallenges
373
