Jason Agents for Knowledge-aware Information Retrieval Filters

Dima El Zein

and C

elia da Costa Pereira

Laboratoire i3s, Universit

e C

ote d’Azur, CNRS, UMR 7271, Sophia Antipolis, France

Keywords:

Agent Based Framework, Knowledge Aware Filter, User Knowledge, Information Retrieval, BDI Agents.

Abstract:

This paper proposes a novel use of Belief-Desire-Intention agents in Information Retrieval. We present a

cognitive agent that builds its beliefs about the user’s knowledge during his/her interaction with the search

system. The agent reasons about those beliefs and derives new ones using contextual rules. Those beliefs

serve to personalise the search results accordingly. The framework is developed using an extended version

of the Jason agent programming language; the choice of Jason’s extension to model the agents is justiﬁed by

some of its advantageous features, in particular, the possibility to represent gradual beliefs. A running example

will illustrate the presented work and highlight its added value.

1 INTRODUCTION,

MOTIVATION AND RELATED

WORK

It is increasingly common for search and recommen-

dation systems to personalise the items proposed ac-

cording to the user’s preferences, location, proﬁle,

etc. However, most of these systems build the user’s

proﬁle based on his/her search history and do not con-

sider the evolution of the user’s information needs

from a “cognitive point of view” (Culpepper et al.,

2018). For example, some existing contributions in

personalisation applied content-based techniques by

using the content of the documents read by the user to

construct the user’s proﬁle (Ricci et al., 2011; Garcin

et al., 2012). The related user proﬁles are often

static or not frequently updated, hence they cannot

help represent the user’s knowledge, which is con-

stantly evolving. The constant evolution of the user’s

knowledge is an important aspect to be considered

when proposing information that is supposed to be

novel and/or helpful for the user to achieve a goal.

In our opinion, this gap should be essentially ﬁlled:

search results must be adapted to the users’ beliefs,

the knowledge they acquire, or even the goals they

want to achieve. The difﬁculty of extracting, rep-

resenting, and measuring the cognitive aspects may

explain this gap. Indeed, to the best of our knowl-

edge, there are only a few implemented work in In-

https://orcid.org/0000-0003-4156-1237

https://orcid.org/0000-0001-6278-7740

formation Retrieval (IR) systems that considers “un-

derstanding” the user’s mental attitudes (belief and

knowledge changes, goals, . . . ) and personalising the

results accordingly. We believe that to be promis-

ing, an approach should consider users as cognitive

agents (da Costa M

ora et al., 1998; Rao and Georgeff,

2001) with their own beliefs and knowledge of the

world.

The ﬁrst contributions using an agent-based archi-

tecture in IR have been proposed twenty years ago;

the main goal was to track the user’s activities to

personalise Web search (Guttman and Maes, 1998;

Bakos, 1997). Then, in order to better understand the

user’s behaviour during the search, some user-related

characteristics, such as location and type of device,

were considered (Carrillo-Ramos et al., 2005; Kuru-

matani, 2004). In (Yu et al., 2021), the authors ex-

plored the correlation between the content of a doc-

ument read and the search behaviour from the one

hand, and a user’s knowledge state and knowledge

gain from the other hand. The results showed that

the knowledge gain can be predicted from the users’

search behaviour and from the content features of the

documents they read (e.g., number of money words,

number of religion words, number of words in each

page, etc.).

A theoretical proposal for an IR framework con-

sidering the user’s knowledge to personalise the

search result was recently proposed in (El Zein

and da Costa Pereira, 2020a). The scope of the

framework was textual document retrieval where the

user’s knowledge was assumed to be the informa-

466

El Zein, D. and Pereira, C.

Jason Agents for Knowledge-aware Information Retrieval Filters.

DOI: 10.5220/0010884900003116

In Proceedings of the 14th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2022) - Volume 2, pages 466-476

ISBN: 978-989-758-547-0; ISSN: 2184-433X

tion contained in the document(s) he/she was read-

ing. The framework considered an agent, designed in

a Belief-Desire-Intention (BDI) architecture (Rao and

Georgeff, 2001), and was made “aware” of the user’s

knowledge. The agent’s beliefs represent the user’s

knowledge that is extracted from the documents the

user has read. Beliefs are considered to be gradual

and are associated to some degrees that deﬁne the ex-

tent to which a piece of knowledge is “entrenched”.

These degrees represent the preferences between be-

liefs, and are used in case of divergence or contradic-

tion between new and old beliefs. In particular, they

help deciding which belief to maintain and which to

discard in case of contradiction. The agent also has

some contextual rules (or knowledge rules) that are

used to derive new beliefs. The main task is to “ﬁlter”

the search results returned by an IR system to ﬁt the

user’s knowledge.

In (El Zein and da Costa Pereira, 2020b) the au-

thors extended the Jason platform (Bordini et al.,

2007) in order to account for the concept of graded

beliefs, implement belief reasoning and belief chang-

ing capabilities. The resulting approach allows agents

to reason about the degree of certainty of beliefs, track

the dependency between them and revise the belief set

accordingly.

In this paper, we propose an Information Fil-

ter agent which puts together the advantages of

the theoretical framework proposed in (El Zein and

da Costa Pereira, 2020a) with the extended Jason

platform proposed in (El Zein and da Costa Pereira,

2020b). We will develop and describe the implemen-

tation of such a framework, and discuss the advan-

tages and challenges associated with it.

For example, if the user reads a document about

human evolution, the agent will be aware of the user’s

knowledge of that subject. Then, suppose the user

submits the query “Charles Darwin”, the search sys-

tem returns a list of documents on several topics about

Darwin, such as his bibliography, his theory of natural

selection and social Darwinism. Since the agent be-

lieves that its user already has knowledge about natu-

ral selection, it should decide which, among the docu-

ments relevant to the user’s query, to return and which

not to. For example, if the agent’s purpose is to re-

turn new information, then documents dealing with

the subject of natural selection will be discarded.

The language required to develop our ﬁltering

agent must allow the following: (1) representing the

user’s knowledge in the form of agent beliefs (2) asso-

ciating beliefs to a degree of entrenchment (3) revis-

ing the beliefs in case of contradiction (4) updating

the degree of a belief if it already exists (5) deriving

new beliefs by reasoning with the knowledge rules.

All the mentioned conditions are satisﬁed in the ex-

tended version of Jason (El Zein and da Costa Pereira,

2020b).

The paper is organised as follows. Section 2 dis-

cusses the preliminaries to understand the features of

the framework. Section 3 describes Jason, its current

limitations, and the features of its extended version.

Section 4 discusses the cognitive IR framework. Sec-

tion 5 describes the proposed implemented prototype

as a proof-of-concept in the news articles (BBC) do-

main and, ﬁnally, some conclusions and perspectives

of future work are presented in Section 6.

2 PRELIMINARIES

2.1 Rule-based Agents

A Rule-based agent (Jensen and Villadsen, 2015) has

a belief base consisting of facts (ground literals) and

rules (Horn clauses). The facts can originate from

communication with other agents, observations of

the environment, downloaded information from web

sources or other ressources. Facts might change over

time as a result of the inference process or of the

addition and deletion of other facts from the agent’s

belief base. A literal α is a predicate symbol that

is possibly preceded by a negation symbol ¬. We

consider a ﬁnite set R of rules, which have of the

form α

&α

...&α

→ β where α

,α

,...,α

(n ≥

1) and β are literals. β is called the derived belief,

and each α

is a premise of the rule. The & represents

the logical and operator. We consider the agent’s be-

liefs when the agent’s rules have run to quiescence,

i.e., after all the agent’s rules have been applied to all

the literals in the agent’s memory. Note that this set is

ﬁnite if the original set of rules and ground literals is

ﬁnite.

2.2 Belief Revision and Partial

Entrenchment Ranking

Belief revision is, by deﬁnition, the process of modi-

fying the belief base to maintain its consistency when-

ever new information becomes available. The AGM

belief revision theory (Alchourr

on et al., 1985) de-

ﬁnes postulates that a rational agent should satisfy

when performing belief revision. In such a theory,

a belief base is closed under logical consequence. We

consider a belief base K and a new piece of informa-

tion α. K is inconsistent, when both α and ¬α are

in Cn(K), or Cn(K) = ⊥, or both α and ¬α are log-

ical consequences of K. Three operators are consid-

ered: Revision K ∗ α : adds a belief α as long as it

Jason Agents for Knowledge-aware Information Retrieval Filters

467

does not cause a contradiction in K. If the addition

will cause inconsistencies in K, the revision opera-

tion starts by minimal changes in K to make it con-

sistent with α, then adds α. Expansion K + α: adds

a new belief α that does not contradict with the ex-

isting beliefs. Contraction K ÷ α: removes a belief

α and all other beliefs that logically imply/entail it.

The sentences in a belief set may not be considered

equally important as it is assumed in the AGM pos-

tulates: belief is gradual and an agent might have be-

liefs more entrenched/accepted than others. Williams

(Williams, 1995) have proposed a quantitative ap-

proach for the AGM framework, by developing ﬁnite

partial entrenchment rankings to represent epistemic

entrenchment – a piece of information is labelled by

a degree of conﬁdence denoting how strongly we be-

lieve it. The epistemic entrenchment (G

ardenfors and

Makinson, 1988) captures the notions of signiﬁcance,

ﬁrmness, or defeasibility of beliefs.

Intuitively, epistemic entrenchment relations in-

duce preference orderings of beliefs according to their

importance in the face of change. If inconsistency

arises during belief revision, the least signiﬁcant be-

liefs (i.e., beliefs with the lowest entrenchment de-

gree) are given up until consistency is restored. The

belief revision operator(s) must then take into con-

sideration the degree i of the belief to be added and

decide whether to add it or not. We discuss later in

Section 5.1 the belief revision algorithm K ∗ (α, i) we

followed.

2.3 Alechina’s Belief Revision and

Tracking

Alechina et al. (Alechina et al., 2005) proposed be-

lief revision and contraction algorithms for resource-

bounded agents. They consider a ﬁnite state and a

ﬁnite program having a ﬁxed number of rules used to

derive new beliefs from the agent’s existing beliefs.

The approach associated a preference order (similar

to Williams’ approach (Williams, 1995) ) for each be-

lief and tracked dependencies between them.

For every ﬁred rule instance, a Justiﬁcation J will

record: (i) the derived belief and (ii) a support list,

s, which contains the premises of the rules. The de-

pendency information of a belief had the form of two

lists: dependencies and justiﬁcations. A dependen-

cies list records the justiﬁcations of a belief, and a

justiﬁcations list contains all the Justiﬁcations where

the belief is a member of support. The approach rep-

resents the agent’s belief base as a directed graph with

two types of nodes: Beliefs and Justiﬁcations. A Jus-

tiﬁcation has one outgoing edge to the belief it is a

justiﬁcation for, and an incoming edge from each be-

lief in its support list.

Preference on Beliefs and Quality of justiﬁcations

As beliefs are associated with preferences, justiﬁca-

tions are associated with qualities. A quality of a

justiﬁcation is represented by non-negative integers in

the range [0, . . . , m], where m is the maximum size of

working memory. The lower the value, the least the

quality.

Deﬁnition 1. The preference value of a belief α,

p(α), is equal to that of its highest quality justiﬁca-

tion.

p(α) = max{qual(J

),...,qual(J

)} (1)

Deﬁnition 2. The quality of justiﬁcation J, qual(J),

is equal to the preference of the least preferred belief

in its support list.

qual(J) = min{p(α) : α ∈ support of J} (2)

Independent beliefs have at least one justiﬁcation

with an empty support list (non-inferential justiﬁca-

tion). They are usually those in the initial belief base

or those perceived from the environment. It was as-

sumed that non-inferential justiﬁcation is associated

with an a priori quality.

3 JASON: PROPERTIES,

LIMITATIONS & EXTENSION

This section presents an overview of the Jason lan-

guage’s architecture (Bordini et al., 2007) and its fea-

tures (v.2.4) (Jas, 2021). We also discuss the features

and the motivation of the extended version proposed

in (El Zein and da Costa Pereira, 2020b) that we will

use later in our framework.

3.1 Architecture

A Jason agent, similarly to other agents modeled in

BDI, is deﬁned by sets of beliefs, plans, and goals or

intentions. Jason’s beliefs are represented by predi-

cates. Their existence in the agent’s belief base means

the agent currently believes that to be true. The ∼ op-

erator refers to the negation, explicitly representing

that the agent believes a literal to be false. Annota-

tions distinguish the Jason syntax: an annotation is a

list of terms placed after a belief, enclosed in square

brackets, revealing details about it.

A plan is composed of three parts: the triggering

event, the context, and the body. It is expressed as

follows:

+triggering event : context ← body. (3)

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

468

Table 1: Comparison between Jason and its extension’s fea-

tures.

Feature Original Extended

Beliefs

Dependencies Not tracked Tracked

Inconsistency Accepted

Not

accepted

Graded No

Yes

(degOfCert)

Preference

preference

High  Low

New  Old

Plans

Knowledge-rule with

n conditions

n plans 1 plan

Order of conditions Dep—endent Independent

(with +tei)

Execution of applica-

ble plans with same

triggering event

One plan

only

All plans

triggering event represents one condition that might

initiate the plan’s execution; it can be the addition (+)

or deletion (-) of a belief or a goal. context is a con-

junction of literals that need to be satisﬁed to make the

plan applicable – and possibly executed. A plan is ap-

plicable if: (i) ﬁrst, its triggering event occurred, and

(ii) its context (one or several conditions) is a logical

consequence of the agent’s current beliefs. The body

is a sequence of actions or goals to be achieved upon

the plan execution. Together the triggering event and

the context constitute the plan’s head.

3.2 Extension of Jason: Graded Belief

Revision

The subject Jason extension (El Zein and

da Costa Pereira, 2020b) models knowledge-rule

agents that are capable of reasoning with uncertain

beliefs, tracking the dependency between beliefs

as done in (Alechina et al., 2005) and explained in

Section 2.3 and revising the belief set accordingly.

In the following, we highlight the limitations of the

original version and how the extension overcame

them. Table 1 compares the features of the two

versions.

3.2.1 Representation and Execution of

Knowledge Rules

To represent a rule having n conditions in the form of

&α

...&α

→ β using Jason plans, the premises

of rules are supposed to be in the plan’s head. That

 refers to “more preferred”.

means that one of the conditions of the premises must

be the triggering event and the others in the context

(for example, +α

: α

&...α

→ β ). However,

the plan execution in the original Jason version is re-

liant on the occurrence of the triggering event: a plan

is executed only if the context conditions were sat-

isﬁed before the triggering event takes place. The

order of the triggering event and the context condi-

tions matters for the execution of a plan. If the con-

dition a

was satisﬁed before the others, the plan will

not be executed. One alternative could be to write n

plans. The extended version of Jason allows the ex-

pression of knowledge-rules by the so-called Trigger-

Independent plans. Those plans will be executed

whenever the combination of several conditions is sat-

isﬁed, no matter which condition was satisﬁed ﬁrst. In

other terms, they do not wait for one speciﬁc trigger

condition to occur to execute the plan. The syntax of

Trigger-Independent plans should have the reserved

word “+tei” that stands for trigger event independent

in the trigger part and all the other conditions in the

context. The plan’s new syntax to represent knowl-

edge rules is proposed:

+tei : context

← body. (4)

context

has all the premises α

& α

&...&α

and

the body contains β the derived belief.

Using the original Jason in the case where two or

more plans had the same trigger and all had a satisfy-

ing context ﬁeld, would return only one plan for exe-

cution. The returned plan would be by default the ﬁrst

plan according to the order in which plans were writ-

ten in the code. Contrarily, using the extension in the

same case would return/execute all the plans having

satisfying conditions.

3.2.2 Degree of Certainty

The notion of “believing” in Jason is Boolean: An

agent either believes something is true or false or is ig-

norant about it. The extension allowed the representa-

tion of gradual beliefs by expressing “degOfCert(X)”

in the annotation part of a belief - X represents the

belief certainty deﬁned as follows: We deﬁne the cer-

tainty of a belief α as representing the degree to which

the agent believes the belief is true.

The degree of certainty associated with initial be-

liefs, beliefs communicated by other agents and be-

liefs perceived by the agent must be explicitly deﬁned

by the source. As for derived beliefs their related de-

gree will be discussed in 3.2.3.

3.2.3 Deriving and Tracking Beliefs

Derived beliefs are dependent on the premises that

derived them; therefore to calculate their related de-

Jason Agents for Knowledge-aware Information Retrieval Filters

469

Figure 1: Graph over the beliefs and justiﬁcations.

gOfCert, the dependency between beliefs must be

tracked. The extension tracked the beliefs following

the approach discussed in 2.3: a justiﬁcation is repre-

sented by a derived belief, a support list, and a qual-

ity; a belief is represented by a dependencies list, a

justiﬁcations list and a degree of certainty. Whenever

a knowledge-rule, named trigger-independent plan, is

ﬁred and a new belief is added, a justiﬁcation node is

created. This node links the rule’s premises with the

derived belief. The degree of a derived belief is au-

tomatically calculated by the interpreter using Equa-

tion 1. When any of the beliefs is contracted, the re-

lated justiﬁcations are removed as well. Justiﬁcations

with an empty support lists are created upon the addi-

tion of initial, communicated, and perceived beliefs.

Unlike in (Alechina et al., 2005), no a priori quali-

ties are assigned for the justiﬁcation of independent

beliefs, as the degrees are explicitly stated.

Example 3.1. Figure 1 illustrates the belief tracking,

considering four beliefs α, β, γ, and µ, and a rule α

& β → γ. The rule means that if the agent believes in

α and β, it believes in γ. For example, Justiﬁcation

is denoted as (γ,[α,β]); γ is the derived belief and

[α,β] is the support list. J

is in the dependencies list

of γ and in the justiﬁcations list both α and β. If γ

were also derived from µ, i.e. µ → γ, then its depen-

dencies list would also include another justiﬁcation

denoted as (γ,[µ]). If the belief α was the result of

an observation, its dependencies list would include a

justiﬁcation J

= (α,[]) having an empty support list.

3.2.4 Belief Revision

Contradictory beliefs were accepted in the Jason’s be-

lief base and no belief revision was performed; no

preference on beliefs neither. The agent could believe

in α and its opposite ∼ α at the same time. The ex-

tended version integrated the notion of belief’s cer-

tainty into the belief revision decisions and did not al-

low belief inconsistency. In case of contradiction, the

preference is given for the belief with the higher de-

gree: belief with the smaller certainty degree in the in-

consistency pair will be contracted/discarded, and the

other belief will be added/kept. In the case of equal

certainties, the new belief is given the preference.

A contraction algorithm was proposed: A belief α

is not contracted unless a more preferred belief ∼ α

was added. When contracting a belief α, there is no

need to contract beliefs that derived it: when the rule

deriving α will attempt to add it again, the addition

will be discarded because it will be faced by ∼ α that

is more preferred. In other terms, the belief in ques-

tion is contracted with its related justiﬁcations without

contracting neither the rule’s premises nor the rule it-

self. Beliefs with no justiﬁcations will also be con-

tracted.

4 THE KNOWLEDGE-AWARE IR

FRAMEWORK

The IR framework proposed in (El Zein and

da Costa Pereira, 2020a) consists of a search agent

that personalises results to the user’s knowledge. The

framework considers a client-side agent that uses the

content of the documents read by the user to under-

stand his/her knowledge. For every submitted query,

the ﬂow is as follows: (i) the agent sends the query to

the search system and receives a list of candidate doc-

uments (ii) the agent examines the content of the doc-

uments in the list and measures the similarity of each

document with the set of beliefs (iii) the agent returns

a ﬁltered list of documents according to the similar-

ity results (iv) the user reads a document (v) the agent

considers the content of the document is an acquired

knowledge by the user. The keywords representing

the examined documents are added as agent beliefs

(vi) a reasoning cycle is performed to run the applica-

ble rules and revise the beliefs whenever needed.

The IR agent is modeled as a Rule-based entity.

When the IR agent has α in its belief base, it believes

that the user knows that α is true. If the belief base

contains ¬α, then the agent believes the user knows

that α is not true. When neither α nor ¬α is in the

belief base, the agent believes neither the user knows

α is true nor the user knows that α is false. The agent

also has some knowledge rules that will help it reason

and derive news beliefs. During an agent’s reasoning

cycle, the validity of the rules is checked. A rule is

considered valid if all the conditions in the premises

are satisﬁed (the premise exists in the belief base), the

rule is ﬁred and the belief in the body is added as a

belief. The rules were considered static, and their ex-

traction/origin was not discussed.

The agent acquires its beliefs about the user’s

knowledge from the documents the user has read.

When the user reads a document d, the agent extracts

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

470

the document’s content and considers it an acquired

knowledge by the user. The authors in (El Zein

and da Costa Pereira, 2020a) applied the Rapid Auto-

matic Keyword Extraction RAKE (Rose et al., 2010)

as an easy and understandable method, to extract the

set of keywords representing the document and cal-

culate their related score. The agent’s beliefs will be

then represented by the set of keywords extracted with

RAKE, called extracted beliefs.

Knowledge is gradual: an agent might have be-

liefs more entrenched than others. The “degree” mea-

suring this entrenchment is deﬁned as below:

Deﬁnition 3. The degree of a belief α is the degree

to which the agent believes the user is knowledgeable

about α. It is represented by a decimal [0; 1], where 0

means the lowest degree –the agent believes the user

has absolutely no knowledge about α, and 1 means

the highest degree –the agent believes the user has

the maximum knowledge about α.

A document d = {(k

),...,(k

)} is a set of

tuples where k

is the keyword extracted by RAKE and

is its related score. k

is associated with an ex-

tracted belief b

whose degree is calculated as fol-

lows:

degree(b

) = λ ·

max(s

)

∈d

(5)

In Equation 5 the RAKE score of an extracted key-

word is normalized then multiplied by an adjustment

factor λ ∈ [0,1] that weakens the scores’ magnitude.

The factor’s value may vary based on different factors

like the source of the document, for example.

This equation allows the calculation of the degree

for extracted beliefs only. As for derived beliefs, their

degrees will depend on the degree of premises that de-

rived them. For that reason, the beliefs’ dependency

is tracked using Justiﬁcations and nodes as discussed

in Section 2.3.

The ﬁltering process is based on the similarity

Sim(B,d) between the agent’s set of beliefs B =

{(b

;degree(b

)),...,(b

;degree(b

))} and the con-

tent of a document d = {(k

),...,(k

)} to be

proposed to the user. A similarity measure was pro-

posed considering the degrees of the intersected be-

liefs and the document knowledge.

Sim(B,d) =

(

max{

∑

∈d

[e(B,k

))−e(B,¬k

)],0}

|S|

if |S| 6= 0

0 otherwise.

(6)

The formula is inspired by the similarity function

proposed by Lau et al. in (Lau et al., 2004). S is the

set of keywords appearing both in d and in B, deﬁned

by S = {k

∈ d : e(B,k

) > 0 ∨ e(B,¬k

) > 0}.

The extent e(B,k

) = degree(k

) if k

∈ B; and 0

otherwise. In simple terms, the similarity formula

calculates the average belief degree of the intersected

keywords between the document and the beliefs. The

result is a value between 0 and 1; where 1 means that

all the intersected keywords are strong entrenched be-

liefs. A similarity 0 means either that the document

has no keywords in common with the beliefs or that it

contains more contradictory information than similar

ones when compared to the beliefs.

The similarity formula “rewards” the documents

containing common keywords with the set B and pe-

nalizes those containing keywords whose correspond-

ing negated beliefs are in B.

Finally, a cutoff value γ is set for Sim(B, d) that

allows deciding whether the knowledge inside a doc-

ument is similar to a set of beliefs or not. We can cite

at least two main applications for this framework: (1)

Reinforcing the user’s knowledge: returning the doc-

uments that are “close” to the agent’s beliefs (those

having a similarity score higher than the cutoff) will

be returned. (2) Novelty: returning documents with

new content with respect to what the user already

knows, the documents having similarity below the

cutoff will be returned.

5 PROPOSED INFORMATION

FILTER AGENT

5.1 System Design

The primary concern of this paper is the development

of a proof of concept of an agent-based system in

the cognitive IR domain. We develop the IR ﬁlter

discussed in Section 4 using the extended version of

Jason discussed in Section 3.2. The work presented

here is, to our knowledge, the ﬁrst implemented work

of information ﬁlter agents that considers the user’s

knowledge. We justify the choice of the extended Ja-

son language to model agents by its ability to model

rule-based agents.

Jason’s extension considers beliefs as facts, as-

signs entrenchment degree to them represented by de-

gOfCert and deals with belief inconsistency. It also

allows the representation of knowledge-rules that will

derive new beliefs thanks to the trigger-independent

plans. In addition, Jason implements the dependency

approach proposed by Alechina et al. in (Alechina

et al., 2005), and used in (Alechina et al., 2006), to

track the dependencies between beliefs by associat-

ing dependency and justiﬁcations lists for each be-

lief. Another particular advantage of Jason is that it

is an open-source interpreter written in Java, which

makes the development easy and customisable. A fur-

Jason Agents for Knowledge-aware Information Retrieval Filters

471

ther advantage of choosing the Jason extension pro-

posed in (El Zein and da Costa Pereira, 2020b) to im-

plement the IR framework proposed in (El Zein and

da Costa Pereira, 2020a) is that both, the framework

and the extension, track the beliefs by following the

method discussed in Section 2.3.

To perform the proposed integration, we associate

the notions of the IR framework to the Jason’s lan-

guage as follows:

• Jason’s beliefs are the agent’s beliefs about the

user’s knowledge.

• The entrenchment degree of a belief will be rep-

resented by the degOfCert in Jason’s belief anno-

tation. It will represent the agent’s estimation of

the user’s knowledge regarding a concept.

• Initial beliefs: are beliefs that represent initial in-

formation about the user. Their degrees are ex-

plicitly expressed in the beliefs’ annotation. Dur-

ing the ﬁrst reasoning cycle, for every initial be-

lief, one justiﬁcation with an empty support list is

created.

• Extracted beliefs: are keywords extracted from

the content of the documents. They will be con-

sidered as perceived beliefs and their belief degree

will be calculated as per Equation 5.

• Derived beliefs: are beliefs that result from ﬁring

the applicable rules and from the reasoning pro-

cess. Their degree is calculated as per Equation 1.

• Contextual rules: are considered static and cannot

be revised.

The facts represent information that the agent has

currently obtained about its user’s knowledge. The

user’s knowledge is represented as Jason beliefs, their

related degree is expressed as degOfCert tracked in

the belief annotation. Those beliefs might change

over time as a result of the addition/deletion of other

beliefs due to: (i) reading new documents that might

contain new information, contradictory information,

or redundant information with different degrees, and

to (ii) the rule’s reasoning process itself, that will de-

rive new beliefs from the agent existing beliefs or

delete beliefs in case of inconsistency. To repre-

sent a rule α

&α

→ β in Jason, the syntax is +tei :

& α

← β.

To maintain the belief base consistency, the en-

trenchment degree of beliefs must be raised or low-

ered via a belief revision operation K ∗ (α,i) where α

is a new belief and i is its new entrenchment degree.

We propose the following to revise belief:

K ∗ (α, i) =











If α /∈ K : K + (α,i)

If α ∈ K :

K + (α,i), if i > j

Nothing, if i < j

If ¬α ∈ K :

K ÷ (¬α, j) then K + (α,i), if i > j

Nothing, if i < j

(7)

The revision operator checks ﬁrst if α already exists in

the belief base. If it is not in the belief base, it is added

with the degree i. If α already exists, the two degrees i

and j are compared. When the new degree i is smaller

than the existing degree j, the degree of α in the belief

base is not changed. When i is higher than the existing

degree j, an expansion operation K + (α,i) will be

initiated and will increase the degree of α from j to i.

The revision operator ﬁnally checks if ¬α is already

in the belief base. If it already exists with a degree

j, the preference will be given to the belief with the

higher degree. When i is higher than j, α will have

the preference to stay, ¬α must be ﬁrst contracted (or

assigned the lowest entrenchment degree equal to zero

for example). Then, α is added with degree i. Finally,

when i is smaller than j, the addition of α is discarded.

5.2 Use Case: Application for Novelty

We built the information system framework discussed

in Section 4 in Java and modeled the ﬁlter agent us-

ing the extended version of Jason discussed in Sec-

tion 3.2. We present in this section a use case of

an interaction between a user and the proposed sys-

tem. The system will be employed to return novel

documents with respect to the user’s knowledge. We

examine the returned results in response to the sub-

mitted queries, investigate the knowledge extraction

process and discuss the ﬁltering decision.

The IR part is built on top of the open-source li-

brary Apache Lucene (luc, 2020) conﬁgured with the

standard analyser for indexing and for searching. It is

set to return 10 documents ranked by relevance to the

query. We use a public dataset of short BBC articles

(Greene and Cunningham, 2006), select the 400 doc-

uments of the technology topic and remove duplicate

documents. We rename the text ﬁles to include the

article’s title and ﬁnally build the search index.

We choose to set the adjustment factor λ of Equa-

tion 5 to be equal to 0.9. The advantage of this fac-

tor is that it prevents having the entrenchment degree

of the most representative keyword of a document d

to be equal to 1. The mentioned keyword is the one

having the maximum RAKE score: k

where k

∈

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

472

Figure 2: Initial state of Jason ﬁltering agent.

Figure 3: The list of documents returned by Lucene for q

d and s

= max(s

)

∈d

; when its related entrench-

ment degree is calculated it is “normalised” by divid-

ing it by itself which results to 1. That would mean

if the user has read a document having “Vaccine efﬁ-

ciency” as the most representative keyword, the agent

would consider the user already reached the maxi-

mum knowledge about vaccine’s efﬁciency, which is

not realistic. We set the similarity’s cutoff value, in-

troduced in Section 4 to γ = 0.25. This value is the

threshold to the similarity Sim(B, d) between the be-

lief set and the candidate document d. In a novelty

context, only documents having a similarity smaller

than γ are returned. The closer the cutoff value is to

0 the more conservative the approach and the more

novel documents are returned.

We assume at time zero, i.e. before the user-

system interaction starts, the agent has no informa-

tion about the user: the belief set B is empty. The

agent has a plan representing a contextual rule that

derives the user knows about the Sony corporation if

he/she knows about both the Gizmondo store and the

PlayStation portable psp. This plan could originate

for example from mining contextual knowledge from

textual corpus, like the information ﬂow discussed

(Lau et al., 2004). The initial state of the agent is

expressed in Figure 2.

The interaction starts by submitting a query

“Gaming device”. The agent relays the query to

Lucene and receives a ranked list of 10 documents re-

sponding to the query -displayed in Figure 3. Since

the agent has no beliefs about the user’s knowledge

yet, there is nothing to compare with the content of

the documents. In this case, any document’s content

will be considered novel to the user’s knowledge. In

consequence, the agent proposes to the user the list of

10 documents without any ﬁlter applied.

Table 2: Sample of the keywords extracted from d

363

Belief

RAKE

Score

Belief

Degree

the gizmondo combined media player 28.1 0.65

multi player gaming 9.3 0.21

gaming gadget 6.3 0.14

gizmondo store 5.7 0.13

the psp 5.3 0.12

ds handheld 4.7 0.1

Knowledge Extraction and Belief Representation.

The user selects for reading the document d

359

en-

titled “Gizmondo gadget hits the shelves”. The in-

formation inside it will be acquired by the user as

new knowledge. To represent this knowledge, the

agent uses RAKE to extract the keywords represent-

ing the document and associates to each of them an

entrenched degree. 39 keywords are extracted from

the document, some of which are illustrated in Table

2. We have replaced the spaces between words with

underscores in order to respect the syntax of Jason’s

belief. The table shows the RAKE score of some key-

words as well as their associated entrenchment degree

calculated using Equation 5.

In d

359

, the keyword “the british-backed gadget

faces stiff competition” has the highest RAKE score

max

∈d

) = 38.6 . To normalize the scores of other

keywords, their RAKE score will be divided by 38.6

then multiplied by the adjustment factor λ. If we con-

sider for example the belief “gaming gadget” having

a rake score of 6.3, its entrenchment degree is cal-

culated as follows deg(gaming gadget) = 0.9 ·(6.3 ÷

38.6) = 0.14. This means that the agent believes the

user has some knowledge about gaming gadget with

the degree 0.14 on a scale of 0 to 1. In total, 39 expan-

sion operations B + (α,Belie f Degree) are performed

to add the new beliefs.

Reasoning and Deriving Beliefs. Once the knowl-

edge is extracted and the beliefs were added, a

reasoning cycle is run to ﬁre applicable rules and

derive new beliefs if needed. After reading the docu-

ment d

359

, the agent plan becomes applicable since its

premises conditions are satisﬁed: the gizmondo store

and the psp beliefs are in the belief base. Therefore,

the plan is ﬁred and its body gets executed: the belief

sony is added. To calculate the entrenchment degree

of the derived belief, a Justiﬁcation J

gets cre-

ated with a quality degree equal to qual(J

) =

min{deg(gizmondo store),deg(the psp)} =

min{0.13;0.12} = 0.12. The belief sony is then

added with degree(sony) = max{qual(J

)} = 0.12.

B + (sony, 0.12) is performed.

Filtering the Results. The user submits another

query q

“PSP”, Lucene returns to the agent 10 can-

didate documents displayed in Figure 4. Now that the

Jason Agents for Knowledge-aware Information Retrieval Filters

473

agent is aware about the user’s knowledge, it is capa-

ble of ﬁltering the documents according to what the

user knows. The similarity Sim(B,d) between the set

of beliefs and the content of each of the 10 documents

will be measured.

We take for example d

025

and d

363

, calculate their

similarity with the belief set and examine the re-

lated ﬁltering decision. The document d

025

enti-

tled “Sony PSP console hits US in March” is rep-

resented by 19 keywords, out of which 1 is com-

mon with the belief set. The set of keywords ap-

pearing both in d and in B is S = {ds handheld}.

Sim(B,d

025

) = max{degree(ds handheld, 0} ÷ 1 =

max{0.1,0} ÷ 1 = 0.1. On the other side, d

363

entitled “sony psp handheld console hits us” has

4 out of 41 common keywords with the belief

set. S = {the gizmondo combined media player,

gaming gadget, the psp, multi player gaming} , the

Sim(B,d

363

) = max{0.65 + 0.21 + 0.14 + 0.12,0} ÷

4 = 0.28. By interpreting these similarities, we con-

clude that d

363

is more similar to the belief set (rep-

resenting the user’s knowledge) than d

025

. In other

terms, if the user reads d

025

, he/she will acquire less

novel information compared to d

363

Knowing that the similarity cutoff value is 0.25,

8 documents are returned to the user including d

025

and excluding d

363

. The list of returned documents

is displayed in Figure 4. Recalling that the aim of

this example is to return the documents having novel

content: documents having a similar content with the

agent’s beliefs should be excluded. Notice that d

359

was returned by Lucene in response to query q

but

ﬁltered out by the agent because it was already read

by the user. The similarity with the belief set is 1.

Belief Expansion. In response to q

the user se-

lects to read d

025

. The 19 keywords representing

the document are then added as the user’s knowl-

edge: the belief base is revised by 19 keywords B ∗

(α,Belie f Degree) as discussed in Section 2.2. The

associated entrenchment degree for ds handheld (the

only common keyword with the belief base) is 0.58.

When the agent is adding this belief, it notices that the

user already has some knowledge about it of 0.1. The

agent then increases the related belief degree to 0.58.

A reasoning cycle runs, they are no contradictions

to resolve and no plans to ﬁre.

6 CONCLUSION AND FUTURE

WORK

This work presented a new use of BDI agents in IR. In

the proposed framework, the beliefs of a BDI (Jason)

agent have been used to represent the user’s knowl-

edge. Besides, the beliefs can be gradual; their related

degrees reﬂect how entrenched is an agent’s belief

about the user’s knowledge regarding a speciﬁc topic.

The agent also can reason about the user’s knowledge,

derive new facts, and decide which belief(s) to hold in

case of inconsistency. The proposed gives the oppor-

tunity of usage in applications requiring personaliza-

tion and understanding of certain cognitive aspects of

the user.

Two of the possible applications of the framework

are (1) novelty, where the returned documents must

contain new information with respect to the user’s

knowledge and, (2) knowledge reinforcement, where

documents must contain information that is similar to

what the user already knows.

We have also presented an example as a proof of

concept of our proposal. We developed the cogni-

tive framework in Java, built the search index with

the Lucene library (luc, 2020), integrated it with Ja-

son’s extension (El Zein and da Costa Pereira, 2020b)

and ﬁnally tested it on a dataset of BBC short news

(Greene and Cunningham, 2006). The example de-

scribed a series of real interactions between the user

and the proposed framework: the user submits queries

to the system, the system responds with personalised

documents, the user selects one document to read and

possibly would issue another query, and so on. We

have also explained how the agent extracted the user’s

knowledge, represented it as beliefs, and ﬁnally used

it to “ﬁlter” the returned documents. Besides, by in-

terpreting the interaction’s result, we have shown how

the IR agent can acquire numerous beliefs about con-

cepts that the user is aware of, i.e., for a single doc-

ument of 500 words, 70 beliefs could be extracted.

Notice that those concepts are not equally known by

the user; this justiﬁes the allocation of entrenchment

degree to each belief. The use case showed how the

agent’s belief set can “expand” when the user acquires

more knowledge (by reading more documents). Be-

sides, it showed how the degree of knowledge about

a concept/keyword can increase when the user reads

some information that he/she already knows.

Although the developed framework considered

the possibility of representing negated beliefs and re-

vising the belief set accordingly, the actual extraction

of negated knowledge is a challenging task. The prac-

tical research advancement to extract such informa-

tion from non-structured text remains an unresolved

issue in the literature (Blanco and Moldovan, 2011).

Therefore, the notion of belief revision could not be

tested in a real case scenario.

Another challenge was the calculation of the sim-

ilarity between the documents’ keywords and the set

of beliefs. The proposed framework assumes that a

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

474

Figure 4: The list of documents returned for q

before and after ﬁltering.

keyword and a belief are “similar” only if they are lit-

erally the same. For example, “multiplayer gaming”

and “multi player gaming” are considered two differ-

ent keywords. We believe that this could be over-

come by applying some normalisation or standardisa-

tion techniques that are publicly available like NLTK

(Bird et al., 2009) and Stanford Core NLP (Manning

et al., 2014). Furthermore, the method used to ex-

tract the keywords does not take into account their se-

mantics, nor does the similarity formula which com-

pares all the keywords. This task might be more chal-

lenging as it requires the integration of some Natural

Language Processing techniques. For future work, we

plan to consider enhancing the developed framework

by extracting normalised keywords with the possibil-

ity of semantically comparing them.

REFERENCES

(2020). Apache lucene. https://lucene.apache.org/.

(2021). Jason agent programming. http://jason.sourceforge.

net/wp/.

Alchourr

on, C. E., G

ardenfors, P., and Makinson, D.

(1985). On the logic of theory change: Partial meet

contraction and revision functions. The journal of

symbolic logic, 50(2):510–530.

Alechina, N., Bordini, R. H., H

ubner, J. F., Jago, M., and

Logan, B. (2006). Belief revision for agentspeak

agents. In AAMAS, pages 1288–1290. ACM.

Alechina, N., Jago, M., and Logan, B. (2005). Resource-

bounded belief revision and contraction. In Interna-

tional Workshop on Declarative Agent Languages and

Technologies, pages 141–154. Springer.

Bakos, J. Y. (1997). Reducing buyer search costs: Impli-

cations for electronic marketplaces. Management sci-

ence, 43(12):1676–1692.

Bird, S., Klein, E., and Loper, E. (2009). Natural language

processing with Python: analyzing text with the natu-

ral language toolkit. ” O’Reilly Media, Inc.”.

Blanco, E. and Moldovan, D. (2011). Some issues on de-

tecting negation from text. In Twenty-Fourth Interna-

tional FLAIRS Conference.

Bordini, R. H., H

ubner, J. F., and Wooldridge, M. (2007).

Programming Multi-Agent Systems in AgentSpeak Us-

ing Jason (Wiley Series in Agent Technology). John

Wiley & Sons, Inc., Hoboken, NJ, USA.

Carrillo-Ramos, A., Gensel, J., Villanova-Oliver, M., and

Martin, H. (2005). Pumas: a framework based on

ubiquitous agents for accessing web information sys-

tems through mobile devices. In Proceedings of the

2005 ACM symposium on Applied computing, pages

1003–1008.

Culpepper, J. S., Diaz, F., and Smucker, M. D. (2018). Re-

search frontiers in information retrieval: Report from

the third strategic workshop on information retrieval

in lorne (swirl 2018). In ACM SIGIR Forum, vol-

ume 52, pages 34–90. ACM New York, NY, USA.

da Costa M

ora, M., Lopes, J. G. P., Vicari, R. M., and

Coelho, H. (1998). Bdi models and systems: Bridging

the gap. In ATAL, pages 11–27.

El Zein, D. and da Costa Pereira, C. (2020a). A cognitive

agent framework in information retrieval: Using user

beliefs to customize results. In The 23rd International

Conference on Principles and Practice of Multi-Agent

Systems.

El Zein, D. and da Costa Pereira, C. (2020b). Graded be-

lief revision for jason: A rule-based approach. In In-

ternational Joint Conference on Web Intelligence and

Intelligent Agent Technology (WI-IAT’20).

Garcin, F., Zhou, K., Faltings, B., and Schickel, V. (2012).

Personalized news recommendation based on collabo-

rative ﬁltering. In 2012 IEEE/WIC/ACM International

Conferences on Web Intelligence and Intelligent Agent

Technology, volume 1, pages 437–441. IEEE.

ardenfors, P. and Makinson, D. (1988). Revisions of

knowledge systems using epistemic entrenchment. In

Proceedings of the 2nd conference on Theoretical as-

pects of reasoning about knowledge, pages 83–95.

Greene, D. and Cunningham, P. (2006). Practical solutions

to the problem of diagonal dominance in kernel docu-

ment clustering. volume 148, pages 377–384.

Guttman, R. H. and Maes, P. (1998). Agent-mediated in-

tegrative negotiation for retail electronic commerce.

In International Workshop on Agent-Mediated Elec-

tronic Trading, pages 70–90. Springer.

Jensen, A. S. and Villadsen, J. (2015). Plan-belief revision

in jason. In ICAART (1), pages 182–189. SciTePress.

Kurumatani, K. (2004). Multi-agent for mass user support.

Lecture Notes in Artiﬁcial Intelligence (LNAI), 3012.

Lau, R. Y., Bruza, P. D., and Song, D. (2004). Belief revi-

sion for adaptive information retrieval. In Proceedings

Jason Agents for Knowledge-aware Information Retrieval Filters

475

of the 27th annual international ACM SIGIR confer-

ence on Research and development in information re-

trieval, pages 130–137.

Manning, C. D., Surdeanu, M., Bauer, J., Finkel, J. R.,

Bethard, S., and McClosky, D. (2014). The stanford

corenlp natural language processing toolkit. In Pro-

ceedings of 52nd annual meeting of the association

for computational linguistics: system demonstrations,

pages 55–60.

Rao, A. and Georgeff, M. (2001). Modeling rational agents

within a bdi-architecture.

Ricci, F., Rokach, L., and Shapira, B. (2011). Introduction

to recommender systems handbook. In Recommender

systems handbook, pages 1–35. Springer.

Rose, S., Engel, D., Cramer, N., and Cowley, W. (2010).

Automatic keyword extraction from individual docu-

ments. Text mining: applications and theory, 1:1–20.

Williams, M.-A. (1995). Iterated theory base change: A

computational model. pages 1541–1549.

Yu, R., Tang, R., Rokicki, M., Gadiraju, U., and Dietze, S.

(2021). Topic-independent modeling of user knowl-

edge in informational search sessions. Information

Retrieval Journal, 24(3):240–268.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

476