A Prospect on How to Find the Polarity of a Financial News by Keeping

an Objective Standpoint

Position Paper

Roxana Bersan, Dimitrios Kampas and Christoph Schommer

Dept. of Computer Science and Communication, University of Luxembourg, Luxembourg City, Luxembourg

Keywords:

Artiﬁcial Agents, Financial News, Document Polarity, Wisdom of Crowd, Self-organisation, Social Systems.

Abstract:

This position paper raises the question on how we can keep an independent standpoint regarding the ﬁnding of

a polarity in a news document. As we know, an usefulness and relevance of a text news may be seen differently

by a group of evaluators. The differences are depending on their interests, their knowledge, and/or their ability

to understand. Recent research in literature mostly follow a top-down approach, which is either a context-

based solution or a dictionary-based approach. With respect to the perspective (standpoint) of an evaluator, we

therefore come up with an alternative approach, which is bottom-up and which tends to overcome the power of

a single evaluator. The idea is to introduce a collection of theme-related artiﬁcial agents (ﬁnancial, economic,

or political, ...), which are able to vote. A decision regarding the polarity of a ﬁnancial news bases on the

interplay of a social collection of agents (a swarm), which serve and assist the artiﬁcial agents while fulﬁlling

simple (linguistic, statistical) tasks.

1 MOTIVATION

The European Financial Crisis has emerged within

the last years, with many ups and downs, with many

consequences and decisions for politics and economy.

For example, Eurobonds have been suggested, attract-

ing a great deal of attention while ﬁnancial news ap-

peared in a Tsunami-style of eruptively ﬂowing pace.

Besides, ﬁnancial and political activities have taken

place, political communities have emerged, and coali-

tions established. Also, a certain number of states

have been down-rated, Greece (and potentially other

European states) are close to insolvency. All these in-

formation has been well-noted in ﬁnancial news.

We concern ourselves with such ﬁnancial text

news (Thomson Reuters), which represent a reﬂec-

tion of momentary political, economical, and ﬁnan-

cial incidents. Financial text news can inﬂuence

decisions, expose realistically and unaltered current

events, and/or contribute to the formation of an opin-

ion. Our concern therefore is: assuming that we

have ﬁnancial texts with an exclusive concentration

on facts and objectivity, can we then ﬁnd indica-

tions regarding ﬁnancial, political, or economic de-

cisions, for example with respect to the European cri-

sis? Can we identify a relationship to the ﬁnancial

market (stock market and others)? Can we observe a

composition of similar thinking people? Can we be

proactive and illustrate the emergence of the crisis as

well as a future recurrence?

In Computer Science research, several directions

regarding the analysis of texts have evolved. One of

them is sentiment analysis of texts and with it the

ﬁnding of an inherent polarity of the document. A

sentiment classiﬁcation refers to identifying and ex-

tracting subjective information that appears in source

materials and to determining the attitude of a person

concerning an overall contextual polarity of a docu-

ment. The sentiment may be a judgement or an eval-

uation, an affective state, or the intended emotional

communication. Following this, a ﬁnding of answers

to the questions above might be rather simple in a way

that a certain number of existing techniques may be

applied. More easily, we could argue that we only

have to analyse the documents linguistically, statisti-

cally, and from a Machine Learning point of view, and

that we then may come up with a sentiment decision.

However, this is not as easy as it seems.

A crucial argument is that we must guarantee a

neutral perspective (or standpoint), with almost no

apriori expectations. The reason is that a ﬁnancial

news may be interpreted in a different way, depending

on what a evaluator thinks, believes, and/or knows.

As an example, let us consider the following ﬁnancial

172

Bersan R., Kampas D. and Schommer C..

A Prospect on How to Find the Polarity of a Financial News by Keeping an Objective Standpoint - Position Paper.

DOI: 10.5220/0004191601720177

In Proceedings of the 5th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2013), pages 172-177

ISBN: 978-989-8565-38-9

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

text news:

“Juncker suggested to introduce Eurobonds.

This is a good sign for the new Europe.”

The interpretation of the above sentences is am-

biguous and may cause - because of the evaluator’s

position - different conclusions as well as a misunder-

standing. If the apriori perspective of the evaluator is

somehow positive with respect to Eurobonds, then the

content is very appreciated and the document classi-

ﬁed as to be positive. If it is not, then the document

becomes a negative polarity and with it, possibly a

negative signal for ﬁnancial markets. Therefore, the

argument of having a neutral perspective is needed

with respect to a classiﬁcation of the polarity.

In the following, we will concern the polar-

ity/sentiment of a document and present a collection

of research works that has been made recently. We

will target the problem of having independent per-

spectives (standpoints) and claim that a fair, stable,

and reliable decision can only be made by a voting of

emancipated artiﬁcial agents.

2 SELECTED RESEARCH

2.1 Feature Spaces

A ﬁrst idea on how to discover a polarity of a news

texts are the geometric models. Given a set of (prede-

ﬁned) features F={ f

, . . . , f

}, where for example the

features represent ﬁnancial terms, locations, interests,

et cetera, then each ﬁnancial news can be represented

itself as a vector in the space. We take the frequency

of a feature f

in a ﬁnancial news document as a coor-

dinate of the normalized vector F.

Regarding the polarity of ﬁnancial news docu-

ments, we may start with a set of training docu-

ments, whose polarity is already known (supervised

approach). Their position in the space then gives a

ﬁrst hint on whether new documents, which are close

enough, are becoming even more positive-polar or

negative-polar. However, the assignment of a new ﬁ-

nancial news to the polarity feature space is problem-

atic. The features (dimensions) may be too weak or

less appropriate or their relevance has changed over

time. There is also a big uncertainty regarding ﬁnan-

cial news with regards to their contents: two such doc-

uments can be similar, but the presence of a negation

or an antonym may force a different polarity. More-

over, the perspective (standpoint) of an individual is

not sufﬁciently respected, since the ﬁxing of the di-

mension and/or the supervised polarity assignment of

the training documents are subjective.

The reason why we mention this is caused (among

others) by a work of (C. Scheible, 2012), who present

a novel graph-theoretic method for the initial anno-

tation of high-conﬁdence training data for bootstrap-

ping sentiment classiﬁcation. Here, the polarity is

estimated here by a theme-speciﬁc ‘PageRank’ algo-

rithm. The authors argue that basically sentiment in-

formation is propagated from an initial seed lexicon

througha joint graph representation of words and doc-

uments. They show that their approach outperforms

a baseline classiﬁer and that its performance can be

further improved by a bootstrapping method that can

take advantage of the entire feature space available.

2.2 Polarity in Text Documents

In literature, a conscious discussion on perspectives

(standpoints) is rarely made. Almost any research

work concerns a concrete application or a technical

how-to-do, accomplished with arguments describing

its need. Some applications use a dictionary-based

solution, others a context-basedsolution. As some ex-

amples, (Hassan and Radev, 2010) propose a method

to automatically identify the polarity of words by tak-

ing advantage of a Markov random walk model to

a large word relatedness graph and producing a po-

larity estimate for any given word. The authors say

that a key advantage of their model is its ability to ac-

curately and quickly assign a polarity sign and mag-

nitude to any word. (Richter et al., 2010) describe

a new method for extracting negative polarity item

candidates (called NPI candidates) from dependency-

parsed German text corpora focusing on target multi-

word expressions. (Schumaker et al., 2012) raise the

question whether the choice of words and tone used

by the authors of ﬁnancial news articles can correlate

to measurable stock price movements. If yes, so the

authors, can then the magnitude of price movement

be predicted using these same variables? The authors

answer these questions by using the Arizona ﬁnancial

Text (AZﬁnText) system, a ﬁnancial news article pre-

diction system, and pair it with a sentiment analysis

tool.

(Devitt and Ahmad, 2007) aim to explore a com-

putable metric of positive or negative polarity in ﬁ-

nancial news text, which is consistent with human

judgments. The authors say that this can be used

in a quantitative analysis of news sentiment impact

on ﬁnancial markets. (Sakai and Masuyama, 2009)

propose a method of assigning polarity to causal in-

formation extracted from Japanese ﬁnancial articles

concerning business performance of companies. The

authors assign a polarity (positive or negative) to

causal information in accordance with business per-

AProspectonHowtoFindthePolarityofaFinancialNewsbyKeepinganObjectiveStandpoint-PositionPaper

173

formance. (Drury et al., 2011) propose a strategy

to segment quotations inside a text by an inferred

“opinion maker” role and then apply individual polar-

ity classiﬁcation strategies to each group of the seg-

mented quotations. They have modelled a contextual

informationwith Random Forests based on a vector of

unigrams. (Heerschop et al., 2011) propose a system

called Pathos, which is a framework to perform doc-

ument sentiment analysis. Pathos is partially based

on a discourse structure of the document. The au-

thors hypothesize that - by splitting a text into impor-

tant and less important text spans and by subsequently

making use of this information by weighting the sen-

timent conveyed by distinct text spans in accordance

with their importance - they improve the performance

of a sentiment classiﬁer. A document’s discourse

structure is obtained by applying Rhetorical Struc-

ture Theory on sentence level. (Kaji and Kitsure-

gawa, 2006) propose a novel fully-automated method

of building polarity-tagged corpus from HTML docu-

ments to utilize certain layout structures and linguis-

tic pattern. In general, Polarity Dictionaries are not

less prone to being subject to subjectiveness. (Paulo-

Santos et al., 2011), for example, argue that most ap-

proaches in ﬁnding polarity dictionaries rely on lin-

guistic works concerning part-of-speech tagging or

rich lexical resources such as WordNet. The authors

show and examine the viability to create a polarity

lexicon using only a common online dictionary with

ﬁve positive and ﬁve negative words, a set of highly

accurate extraction rules, and a simple yet effective

polarity propagation algorithm. The algorithm evalu-

ation results show an accuracy of 84.86% for a lexi-

con of 3034 words.

3 A PROSPECTIVE APPROACH

The term rational agent is described by (Russell and

Norvig, 2010) as an entity that given the built-in

knowledge representation, the actual state of the envi-

ronment and the set of possible actions he can take, he

selects an action so as to maximize its utility measure.

The most important agent categories

• Model-based Reﬂex Agents. These take a decision

about its next action based on a set of conditional

action rules about the effect of the actions on the

environment state

• Goal-based Agents. These types of agents are

dedicated to ﬁnd action sequences for achieving

a higher goal

• Utility-based Agents. These agents worke more

on representation, modelling and learning.

(Michael Brenner, 2009) consider the problem-

atic of intermingling planing with acting in dynamic

and partially observable multi-agent environments. In

(Trevor Bench-Capon, 2012), the authors utilize arti-

ﬁcial agents in order to foster on a decision regard-

ing economic experiments with games implementa-

tions. In (J´erˆome Lang, 2012), binary tree rules are

used for deciding the winner as elected by a majority

vote. (Davide Grossi, 2012) look into the possibility

of incorporating a form of dependence relation in the

ﬁeld of game theory for an agent interaction. Coop-

erative games are analyzed, where coalitions under-

take agreements based on dependency relations. (An-

dreasWitzel, 2012) analyze the epistemic reasoning of

communicating rational agents with focus on the dis-

tributed form of iterative elimination of strictly domi-

nated strategies.

3.1 Demands

Regarding the polarity ﬁnding, we demand for adap-

tive artiﬁcial agents (Russell and Norvig, 2010),

which are able to represent and acquire knowledge, to

learn from internal and external information ((Clark,

2001)), and to take advandage of the Wisdom of

Crowds ((Surowiecki, 2004)). An artiﬁcial agent

is able to classify documents based on its internal

knowledge base. With this, it owns an aptitude, which

is a competency regarding a ﬁeld of application, for

example ﬁnance, politics, economics, and others. An

artiﬁcial agent can either represent a natural person

(Merkel, Sarkozy, ...), a country (Greece, Germany,

...), or another individual. Each agent owns a stand-

point and is able to vote, classifying a ﬁnancial news

individually to a polarity, which is either positive,

negative, or neutral. Using a set-based operation like

the intersection can be applied to prove a stable, re-

liable, and plausible polarity decision, where a ﬁnan-

cial news is then positive (negative, neutral), if the

majority of the agents vote for positive (negative,neu-

tral), respectively. We assume that the fundamental

characteristics of an agent are the following:

• Contradicting opinions: since there exist much

more than one perspective (there are often many

truths), an agent must have have an adaptive

knowledge base.

• Able to take a decision: an agent must be able to

operate on text, to ﬁnd associative structures, co-

occurrences, or other forms of patterns.

• Independence of voting: An agent must state

his/her polarity decision independently and has

only then the right to vote.

• Presence of a theme: An agent can only decide a

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

174

document polarity, if an evaluation theme exists.

Otherwise, a voting may become directionless.

Then,

polarity(d, r

) =







+1, ∆(r

, r

j−

) > ε

−1, ∆(r

, r

j−

) < −ε

0, −ε ≤ ∆(r

, r

j−

) ≤ ε

(1)

with |d| = n = i + j and where r

j−

) refer to a

positive (negative) vote regarding the document. ε

is a threshold, which only classiﬁes a ﬁnancial news

to -1 or +1 if ∆ (x,y) is the absolute difference of

(x,y). Only in case that the voting is equal, the doc-

ument is seen to be neutral. This follows the concept

of (Surowiecki, 2004), who argues that decisions are

taken by a large group, even if the individuals within

the group are not smart; but these decisions are al-

ways better than decisions made by small numbers of

’experts’.

3.2 Artiﬁcial Agents

An alternative idea is to understand the polarity as a

decision, which is taken by a majority (or a weighted

sum of) of slave agents, which serve a given master

agent. These slave agents share a small capacity, are

assigned a simple task, and collaborate as a part of a

social system or swarm. One consequence of such an

architectural framework is a small amount of apriori

knowledge, because the participating entities have to

do a little task requiring less of it. Also, a plausibility

of the polarity decision will be inherently given. All

entities’ decision, being either ’positive’, ’negative’,

or ’neutral’ can be identiﬁed and arguments for the

ﬁnal decision retrieved. As the decision is made by

many collaborating entities, we assume the decision

to be more fault-tolerant, more resistant against tem-

poral changes, and less vulnerable to a wrong doc-

ument classiﬁcation. A single change of the knowl-

edge landscape (for example ‘Sarkozy’ is no longer

president but ‘Hollande’ is now) will not have such a

big effect. Moreover, the social system might work

autonomously and organises itself, reducing the num-

ber of investigated efforts. And ﬁnally, an indepen-

dent perspective is maintained. With that, we may un-

derstand an agent as an artiﬁcial entity, which knows

its user (reviewer of the ﬁnancial news), but which is

served by even small entities, i.e., slave agents.

Assume that we have a certain number of Euro-

pean key players, country names, locations, and other

facts. In an intelligent environment, the social system

could detect such facts by itself and neglect such facts

in case of inactivity over a certain period of time, but

we keep it more simple here and assume that a cer-

tain number of fact slave agents (for example, focus-

ing on the key players in Europe, countries, locations,

etc.), whose task is to serve the artiﬁcial agent and to

check a document for occurrences and frequencies of

assigned terms. If the frequency is sufﬁcient, possibly

above a given threshold, then each fact slave agent

contributes to the polarity decision.

We also may consider k-ary operations like

agrees(X), brings(X,Y), has(X,Y), or gives(X,Y,Z),

which are addressed by action slave agents aiming

at instantiating the arguments or even word polarities

like war (-), Eurobond (+/-), or good (+), where we

assign an individual polarity agent, whose task is to

control the presence of predeﬁned words. Of course,

many other types of agents may be used, for exam-

ple a negation slave agent, whose task could be to

convert an action slave agent’s decision; or an un-

certainty agent, whose task is to reduce a certainty

of the agent’s decision, for example by a multiplica-

tive compensation. Warehouse slave agents may have

the task to put all these information together, bringing

the whole information landscape to a consistent and

reasonable decision. Finally, statistical slave agents

and linguistic slave agents may be taken as well,

for example to deliver statistical and linguistic num-

bers/values.

But which role do the warehouse slave agents

play? Do they just compute weights and relation be-

tween action slave agents and polarity slave agents

or should they perform more than that? Moreover,

which are the role of the linguistic agents? Is a lin-

guistic analysis not already incorporated by the action

slave agents? Which is the role of the statistical slave

agents, especially, are they not yet incorporated by the

polarity slave agents? To give a more precise answer,

we suggest the following:

• A Fact slave agent can be either a subject slave

agent or an object slave agent. Each of these

agents can have sub-hierarchies of their own, for

example a subject slave agent may have subcat-

egories like ‘politician’, ‘company’, et cetera, an

object slave agent categories ‘location’, ‘event’,

et cetera. As an example, ‘Merkel is the chancel-

lor’ is a subcategory of ‘politician’, and with that,

a subcategory of the subject agent. ‘Madrid is a

city’ is a subcategory of a country, and with it,

a subcategory of a ‘location’. ‘Summit G-20’ is

subcategory of ‘event’, which is a subcategory of

an object agent.

• Action slave agents can be, as mentioned above,

k-ary. But we think that using action slave agents

with ≥ 3 parameters burdens the relation extrac-

tion too intensively. Assuming to determine one

term would be as good as 90%, then we probably

get an accuracy of almost 72% for three terms.

AProspectonHowtoFindthePolarityofaFinancialNewsbyKeepinganObjectiveStandpoint-PositionPaper

175

• Polarity slave agents can be applied for the sub-

categories ‘verb’, ‘noun’, and ‘adjective’. Some

verbs and adjective may have a given standard po-

larity (Example ‘good’ polarity is positive, ‘hates’

polarity is negative), whereas a polarity of nouns

may differ over time (Example: ‘war’ is con-

stantly, but the polarity of ‘Eurobond’ is probably

not).

• Linguistic slave agents would perform some lin-

guistic analysis deciding on the polarity of the

sentence based on some predeﬁned policy. E.g.,

when the verb has a negative or positive polarity

the sentence takes the polarity of the verb. If the

verb has an objective polarity, then the polarity of

the sentence is the polarity of the nouns, the adjec-

tives, or the adverbs (Example: “Merkel supports

Eurobonds.”). Thus, the polarity of the sentence

is taken by the polarity of Eurobonds. Linguistic

agents capture negations as well reversing the po-

larity of a sentence as well. E.g.“Merkel does not

agree on Eurobonds”.

• Uncertainty slave agents are responsible for de-

creasing the polarity volume of the sentence by

capturing the uncertainty word.

• Warehouse slave agents are basically responsible

for the decision, because they integrate the infor-

mation coming from the other agents. However,

they are not allowed to vote.

3.3 How to Find a Decision?

An agent may be composed of many slave agents,

which perform a simple task. The working together

of these agents will then become the fundament with

respect to the polarity. In the simplest case, a voting

of all agents with equal rights can be taken into ac-

count. But before the decision on the polarity of a

document is taken, it must be considered whether al-

ternative types of voting can be applied (or not), espe-

cially plurality voting systems, single-winner voting

systems, or multiple-winner voting systems. For ex-

ample, whether there exist the word ‘Eurobond’or not

can be subject to a plurality voting system. But, which

countries are pro ‘Eurobond’ or against ‘Eurobond’ it

is a multiple-winner one.

Regarding the voting decision, there are numerous

paths in theory and application, which can be applied.

These theories have their roots in the ﬁelds of Game

Theory, Auction theory, and multi-agent systems. Ex-

amples for the ﬁeld of Game theory are Nash equilib-

rium and the revelation principle of economics; ex-

amples for the ﬁeld of Auction Theory are English

auction, Dutch auction, Vickrey auction, and sealed

ﬁrst-prive auction. It is important to keep this in

mind, because such a system takes into consideration

an application, such as the increase of utility func-

tions, the prediction of some economical phenomena,

et cetera. Without this, a system would sound noth-

ing more than a data collection system with no direct

application.

4 CONCLUSIONS

The given idea is a visionary and prototypically try

to overcome the problem of having a subjective per-

spective (standpoint) regarding the polarity ﬁnding of

a document. As mentioned, the interpretation of a ﬁ-

nancial news may depend on a reviewer’s knowledge,

interest, and much more. Having designed artiﬁcial

agents of different thematic directions, being assisted

by many self-organising and self-evaluating types of

agents, then this may overcome the given problem.

Regarding the voting, which is only allowed for the

agents, an independent voting is recommended. With

that, we believe to fulﬁll the given demands of hold-

ing a perspective, an ability to take a decision, an

independence of voting, and a presence of a theme.

Our aim is now to follow this idea and to come up

with a more detailed concept. An implementation is

planned.

ACKNOWLEDGEMENTS

This work is supported by the Fonds National de la

Recherche (FNR) Luxembourg within the research

project ESCAPE, which is an interdisciplinary project

between the Department of Computer Science &

Communication and the Luxembourg School of Fi-

nance (LSF). The authors thank Mihail Minev as well

as Prof. Grammatikos and Andreas Chouliaris for in-

ternal discussions.

REFERENCES

AndreasWitzel, Krzysztof R. Apt, J. A. Z. (2012). Dis-

tributed iterated elimination of strictly dominated

strategies. Autonomous Agent Multi-Agent Systems.

C. Scheible, H. S. (2012). Bootstrapping sentiment labels

for unannotated documents with polarity pagerank. In

LREC 2012.

Clark, A. (2001). Reasons, robots, and the extended mind.

In Mind and Language, volume 16:2, pages 121–145.

Davide Grossi, P. T. (2012). Dependence in games and de-

pendence games. Autonomous Agent Multi-Agent Sys-

tems.

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

176

Devitt, A. and Ahmad, K. (2007). Sentiment polarity iden-

tiﬁcation in ﬁnancial news: A cohesion-based ap-

proach. In ACL 2007.

Drury, B., Dias, G., and Torgo, L. (2011). A contextual clas-

siﬁcation strategy for polarity analysis of direct quo-

tations from ﬁnancial news. In RANLP 2011, pages

434–440.

Hassan, A. and Radev, D. (2010). Identifying text polarity

using random walks. In ACL 2010, pages 395–403.

Heerschop, B., Goossen, F., Hogenboom, A., Frasincar, F.,

Kaymak, U., and de Jong, F. (2011). Polarity analy-

sis of texts using discourse structure. In CIKM 2011,

pages 1061–1070.

J´erˆome Lang, Maria Silvia Pini, F. R. D. S. K. B. V.

T. W. (2012). Winner determination in voting trees

with incomplete preferences and weighted votes. Au-

tonomous Agent Multi-Agent Systems Agent Multi-

Agent Systems.

Kaji, N. and Kitsuregawa, M. (2006). Automatic construc-

tion of polarity-tagged corpus from html documents.

In ACL 2006.

Michael Brenner, B. N. (2009). Continual planning and act-

ing in dynamic multiagent environments. Autonomous

Agent Multi-Agent Systems.

Paulo-Santos, A., Ramos, C., and Marques, N. (2011). De-

termining the polarity of words through a common on-

line dictionary. In EPIA, pages 649–663.

Richter, F., Fritzinger, F., and Weller, M. (2010). Who

can see the forest for the trees? extracting multiword

negative polarity items from dependency-parsed text.

JLCL 2010, 25(1):83–110.

Russell, S. and Norvig, P. (2010). Artiﬁcial Intelligence - A

Modern Approach (3. internat. Ed.). Pearson Educa-

tion.

Sakai, H. and Masuyama, S. (2009). Assigning polarity

to causal information in ﬁnancial articles on business

performance of companies. IEICE Transactions, 92-

D(12):2341–2350.

Schumaker, R., Zhang, Y., Huang, C., and Chen, H. (2012).

Evaluating sentiment in ﬁnancial news articles. Deci-

sion Support Systems, 53(3):458–464.

Surowiecki, J. (2004). The Wisdom of Crowds: Why the

Many Are Smarter Than the Few and How Collective

Wisdom Shapes Business, Economies, Societies and

Nations. Little Brown & Co.

Trevor Bench-Capon, Katie Atkinson, P. M. (2012). Using

argumentation to model agent decision making in eco-

nomic experiments. Autonomous Agent Multi-Agent

Systems.

AProspectonHowtoFindthePolarityofaFinancialNewsbyKeepinganObjectiveStandpoint-PositionPaper

177