SENTENCE SIMILARITY MEASURES TO SUPPORT

WORKFLOW EXCEPTION HANDLING

A. Aldeeb, D. M. Pearce, K. Crockett and M. J. Stanton

Department of Computing and Mathmatics, Manchester Metropolitan University

Chester Street, Manchester, M1 5GD, U.K.

Keywords: Workflow Management System, Exception Handling, Case-based Reasoning, Sentence Similarity

Measures.

Abstract: Exceptions occurrence in workflow systems is common. Searching in the past exceptions handlers’ records,

looking for any similar exception serves as good sources in designing the solution to resolve the exception

at hand. In the literature, there are three approaches to retrieve similar workflow exception records from the

knowledge base. These approaches are keyword-based approach, concept hierarchies approach and pattern

matching retrieval system. However, in a workflow domain, exceptions are often described by workflow

participants as a short text using natural language rather than a set of user-defined keywords. Therefore, the

above mentioned approaches are not effective in retrieval of relevant information. The proposed approach

considers the semantic similarity between the workflow exceptions rather than term-matching schemes,

taking account of semantic information and word order information implied in the sentence. Our findings

show that sentence similarity measures are capable of supporting the retrieval of relevant information in

workflow exception handling knowledge. This paper presents a novel approach to apply sentence similarity

measures within the case-based reasoning methodology in workflow exception handling. A data set,

comprising of 76 sentence pairs representing instance level workflow exceptions are tested and the results

show significant correlation between the automated similarity measures and the human domain expert

intuition.

1 INTRODUCTION

A workflow management system (WFMS) is

essentially a set of tools for modelling, enactment,

and monitoring of business processes (Jablonski and

Bussler, 1996). Workflow process definition

(workflow schema) is the formal representation of a

business process (Casati et al., 2000). The workflow

schema is composed of activities (tasks) that

collectively achieve the business goal. Workflow

tasks are performed by workflow participants

(Human or automated agent) according to their roles

and the structure of the organization.

It is not guaranteed that designers always do a

perfect job in defining a workflow type that totally

represents all properties of the underlying business

process (Hwang et al., 1999). In addition, the IT

infrastructure of the WFMS and external factors can

raise problems. Therefore, the occurrence of

workflow exceptions is unavoidable and there is a

need to handle those exceptions efficiently. Rule-

based reasoning (RBR), Model-based reasoning

(MBR) and case-based reasoning (CBR) are

approaches being used to handle exceptions in

workflow systems (Luo et al., 2003; Hwang et al.,

2005). Workflow exceptions may require human

intervention to establish proper handlers. Those

handlers can be stored in a knowledge base to be

used to handle similar exceptions in future in case of

no available rules to handle them. Searching the

exceptions handlers’ records in the knowledge base,

looking for any similar exception serves as good

sources in designing the solution to resolve the

exception at hand. (Luo et al., 2003; Montani, 2009;

Hwang et al., 1999; Schmidt and Vorobieva, 2008;

Grigori et al., 2001; Aldeeb et al., 2008). This can

be achieved by applying CBR methodology to

support the management of exceptions in business

process execution. The main challenge in applying

CBR to support exceptions handling in workflow

systems is how to represent exceptions as cases,

finding an effective retrieval mechanism of similar

256

Aldeeb A., Pearce D., Crockett K. and Stanton M. (2010).

SENTENCE SIMILARITY MEASURES TO SUPPORT WORKFLOW EXCEPTION HANDLING.

In Proceedings of the 12th International Conference on Enterprise Information Systems - Artiﬁcial Intelligence and Decision Support Systems, pages

256-263

 SciTePress

cases and the calculation of the similarity. Current

applications of CBR in workflow exception handling

use keyword-based retrieval system, concept

hierarchies, and pattern matching and the use of

decision trees to retrieve similar cases from the case

database (Luo et al., 2003; Montani, 2009; Hwang et

al., 1999; Schmidt and Vorobieva, 2008; Grigori et

al., 2001). Some of these approaches depend on

matching individual words in the current exception

description with individual words in the textual

material in the case base. On the other hand, the

concept hierarchies approach is based on defining a

concept of similarity and incorporates the notion of

concept hierarchies. A concept hierarchy is a partial

order of concepts, which indicates general-to-

specific ordering where each case attribute has its

own concept hierarchy. However, in a workflow

domain, exceptions are often described by workflow

participants as a short text using natural language

rather than structural patterns of sentences. The

workflow participants may express the same

exception using quite different sentences in terms of

structure and word content because of the diversity

in human word usage. In addition, the same word

can have different meanings. Therefore, irrelevant

information may be retrieved and the relevant

information may be missed. Some approaches

restricting the allowable vocabulary, use

intermediaries to generate indexing and search keys,

or constructing explicit models of relevant domain

knowledge. However, these approaches lack the

flexibility to support the diversity in word usage of

human, require expert users to generate indexing and

search keys. These shortcomings and the limitations

require an alternative approach which needs to

consider the semantic similarity between the

workflow exceptions rather than term-matching

schemes. This motivates us to investigate the area

of semantic sentence similarity measures and their

potential application in workflow exception handling

(Li et al., 2006; Feng et al., 2008; Aminul and

Inkpen, 2008; Aliguliyev, 2009; Landauer et al.,

1998).

This paper presents a novel approach to the

application of sentence similarity measures within

the CBR methodology to handle instance level

workflow exceptions. A case study of the motor

insurance process is used to prove the concept of our

approach. The initial findings are encouraging and

show that sentence similarity measures can be

applied in the retrieval of relevant information in

workflow exception handling in the knowledge base.

The rest of this paper is organized as follow:

Section 2 discusses exception handling in workflow

systems. Section 3 introduces some sentence

similarity measures used in this research. In section

4, CBR as a methodology in the proposed approach

is presented while section 5 illustrates a proof of

concept prototype and case study. Finally, section 6

concludes and mentions some enhancements

foreseen as a future work.

2 EXCEPTIONS HANDLING

IN WFMS

WFMSs are designed to follow standard business

processes and routine. However, these processes

face the need to handle exceptions that fall outside

the normal control flow (Casati et a., 2000).

Exceptions occur commonly in workflows (Kumar

and Wainer, 2005; Sadiq et al., 2005; Hwang and

Lee, 2005). Workflow exception is any deviation

from an ideal collaborative process that uses the

available resources to achieve the task requirements

in an optimal way (Klein and Dellarocas, 2000).

There are four main causes of business process

exceptions: system errors, data issues, external

factors, and process design (Kelly, 2005). System

errors can be independent of the transaction data and

business logic and can be caused by underlying

system problems, such as servers being down or

services that are not available. Data issues can be

missing, invalid or inconsistent data. External

factors can trigger a process exception, such as that a

specific item is out-of-stock or unavailable. Lastly,

there can be process design issue that raise

exceptions when specific cases need non-standard

treatment for business reasons. Handling those

exceptions depends on their type, severity and at

what level they occur. Possible approaches to handle

exceptions include ignore, retry, partial roll-back

followed by forward execution, add some extra

activities, delete some planned activities, or any

change to the part of the workflow definition that not

executed yet (Hwang et al., 1999).

Some workflow exceptions can be anticipated by

the workflow designer, therefore they are called

expected exceptions. However, others can not be

anticipated and they are called unexpected

exceptions. The expected exceptions are handled by

rule-based reasoning. Those rules are characterised

by the following components (Casati et a., 2000;

Luo, et al., 2003):

 The Event part represents the symptoms of an

exception

SENTENCE SIMILARITY MEASURES TO SUPPORT WORKFLOW EXCEPTION HANDLING

257

 The Condition is a boolean statement that checks

if the symptoms is really an exception

 The Action describes the procedures that must be

invoked to deal with the exception.

However, relying on predefined rules sometimes

is not enough to deal with the unexpected workflow

exceptions caused by ad-hoc changes. In this case,

human intervention may be required to establish an

appropriate handler. The successful exception

handler can be stored for the future to deal with

similar exceptions. Therefore, case-based reasoning

can be applied to handle workflow exceptions by

retrieving the similar exceptions handlers in the

knowledge base. As we mentioned in the previous

section, workflow exceptions are often described by

workflow participants as a short text (sentence)

using natural language rather than a set of user-

defined keywords. This makes the process of

building concept hierarchies and generating index

keys of instance level workflow exceptions

complicated. Table 1, shows an example of instance

level workflow exceptions. Comparing those

exceptions using sentence similarity measures

directly is more practical, save time and effort that

will be spend in building concept hierarchies and

indexing keys. In the next section, sentence

similarity measures are discussed.

3 SENTENCE SIMILARITY

MEASURES

Sentence similarity measures have many

applications, for example, Web page retrieval, text

mining to discover unseen knowledge from textual

database (Atkinson et al, 2004), text summarization

(Erkan, and Radev, 2004), text categorization (Ko et

al., 2004) and machine translation (Liu and Zong,

2004). Similarity computation techniques designed

to detect the similarity between long texts are

centred in shared words because similar long texts

usually have a degree of co-occurring words.

However, in short texts of sentence length, word co-

occurrence may be rare or even null (Li et al., 2006)

because people express similar meaning using quite

different sentences.

The Latent Semantic Analysis (LSA) is one of

the active researches in sentence similarity

computation and information retrieval is (Landauer

et al., 1998; http://lsa.colorado.edu/). LSA is based

on statistical information of words in huge corpus.

In LSA approach, a semantic space is automatically

constructed for retrieval. The basic postulate is that

there is an underlying latent semantic structure in

word usage data that is partially hidden or obscured

by the variability of word choice. A statistical

approach is utilized to estimate this latent structure

and uncover the latent meaning. Words, the text

objects and, later, user queries are processed to

extract this underlying meaning and the new, latent

semantic structure domain is then used to represent

and retrieve information. A set of representative

words needs to be identified from a large number of

contexts (each described by a corpus). A word by

context matrix is formed based on the presence of

words in context. The matrix is decomposed by

singular value decomposition (SVD) into the

product of three other matrixes, including the

diagonal matrix of singular values. The diagonal

singular matrix is truncated by deleting small

singular values. In this way, the dimensionality is

reduced. The original word by context matrix is then

reconstructed from the reduced dimensional space.

Through the process of decomposition and

reconstruction, LSA acquires word knowledge that

spreads in context. When LSA is used to compute

sentence similarity, a vector for each sentence is

formed in the reduced dimension space; similarity is

then measured by computing the similarity of these

two vectors. Because of the computational limit of

SVD, the dimension size of the word by context

matrix is limited to several hundred. As the input

sentences may be from an unconstrained domain

(and thus not represented in the contexts), some

important words from the input sentences may not

be included in LSA dimension space

Li et al. (2006) proposed a method named

(STASIS) that can be used generally in applications

requiring sentence similarity computation. This

method is fully automatic and adaptable across a

range of potential application domains. The

proposed method dynamically forms a joint word set

only using all the distinct words in the pair of

sentences. Then, for each sentence:

 A raw semantic vector is derived with the using a

lexical database.

 A word order vector is formed for each sentence

using information from lexical database.

 The significance of the words and their

contribution to the sentence meaning is weighted

using information derived from corpus.

 By combining the raw semantic vector with

information content from the corpus, a semantic

vector is created for each of the two sentences.

 Semantic similarity is computed based on the

two semantic vectors.

 An order similarity is calculated using the two

order vectors.

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

258

 Finally, the sentence similarity is derived by

combining semantic similarity and order

similarity.

In the evaluation phase of the STASIS, a set of

sentence pairs are collected from a variety of articles

and books in computational linguistics. An initial

experiment on this data illustrates that the proposed

method provides similarity measures that are fairly

consistent with human knowledge (Li et al., 2006).

Both LSA and STASIS measures are used in

implementation of the proposed approach. The next

section illustrates the case-based reasoning

methodology in handling workflow exceptions and

applying sentence similarity measures in the

retrieval phase.

4 CASE-BASED REASONING

METHODOLOGY

CBR is a reasoning paradigm that exploits the

specific knowledge of previously experienced

situations, called cases, to learn and generate

hypotheses about new situations (Montani, 2009;

Shiu and Pal, 2004a). The use of CBR can reduce

the amount of effort needed to formalize the

knowledge, since representing a real world situation

as a case is often simple. The use of CBR facilitates

an automatic acquisition and increases the operative

knowledge, without requiring a hard and time

consuming formalization of knowledge itself, as it

required by other methodologies, such as rule-based

or model-based reasoning (Montani, 2009).

However, rule-based and model-based reasoning are

more effective for applications where theory, not

experience, is the primary guide to problem solving

and the solutions are designed for a specific problem

and are difficult to be adapted (Limam et al., 2003).

A case consists of problem description and case

solution. CBR can therefore be described by the

CBR-cycle which comprises four activities (Watson,

1999):

1- Retrieve similar cases to the problem

description

2- Reuse a solution suggested by a similar

case

3- Revise or adapt that solution to better fit

the new problem if required

4- Retain the new solution once it has been

confirmed or validated

For complicated real world applications there are

some degree of fuzziness and uncertainty that almost

always encountered (Shiu and Pal, 2004b). AI

techniques such as fuzzy logic, neural networks and

genetic algorithms are helpful in areas where

uncertainty, learning and knowledge inference are

part of the system application (Shiu and Pal, 2004a;

Pal et al., 2004; Jeng and Liang, 1995; Pal and Shiu,

2004). In this research we suggest the sentence

similarity measures be part of the CBR to support

cases similarities calculation and retrieval phase of

the CBR-cycle.

Our approach is to maintain records about the

past experience of handling exceptions. Those

records form cases in the knowledge base to be used

to handle exceptions which need to be managed in

similar way, but may occur in different workflow

instances. The structures of the cases that represent

exceptions in the knowledge base are described in

the next section.

4.1 Exception Representation as a Case

The structure of the workflow exception cases is

inspired by and adapted from the work of Hwang et

al. (1999). Each workflow exception case consists of

three main components; Exception Information

Block (EIB), Exception Handler Block (EHB), and

Workflow Instance Block (WIB).

The EIB represents the problem description part

of the case and contains the following main

exception features or attributes:

 Exception Description: Is the semantic

information that describes the exception by the

workflow user. This description takes the form of

short text (sentence length).

 Status: Is the status of the workflow instance that

raises the exception e. g., initiated, active,

suspended, terminated or complete.

 Workflow Participant: is the performer of the

activity that raises the exception. This can be

automated agent or human.

 Workflow version: represents to which version

of the workflow schema the affected workflow

instance belongs.

 Time: indicate when the exception occurs

 Frequency: is a number indicates how many

times this exception case is successfully used.

When this number reaches a certain threshold

value, it will trigger the evolution of the process

model.

The EHB represents the solution part of the case

and indicates the action to be taken to handle the

exception. This can be an automated action or

manual action requiring user intervention. Generally,

the action can be under one of the following

categories:

SENTENCE SIMILARITY MEASURES TO SUPPORT WORKFLOW EXCEPTION HANDLING

259

 Maintaining the workflow normal behaviour e.

g.: Ignore record, notify or add resource.

 Modifying the workflow behaviour e. g.: retry,

suspend, modify, remove, change sequence,

terminate, re-assign and delay.

 Modify the workflow schema e. g. making a new

version and the affected workflow instance will

follow the new version.

The third block, WIB, contains the workflow

instance itself which consists of the data and control

attributes of the workflow. The control attributes are

used by the exception handler.

4.2 The Exception Handling

Mechanism

The exception handling procedure is as follows:

 The workflow participant propagates the

exception raising case to the Exception

Handling Coordinator (EHC) in the WFMS

accompanied by brief event description (short

text of sentence length, 15 to 25 words). This

sentence is the semantic information that

describes the type of the exception.

 Upon receiving the exception case, the EHC

creates a temporary case template (TCT). The

TCT contains two blocks; workflow instance

information block (WIB) and exception

information block (EIB).

 The event description in the TCT will be

compared with the exception description

attributes in the case database using sentence

similarity measures. Similar cases will be

retrieved based on a certain similarity rang

established by the human domain expert.

 The most similar case with the highest overall

matching mark (based on an established

threshold value defined by the domain expert) is

chosen and its solution will be applied with or

without adaptation.

 If no adaptation the case usage frequency field

is incremented and the TCT will be deleted.

 In case of adaptation, a new case will be created

by adding the modified exception handler block

(EHB) to the TCT and storing it as a new case

in the case base

 If there is no similar case found, the TCT will

be stored as a new case in the case base and

EHB will be added. The new EHB is based on

the judgement of the process engineer and the

domain expert

 Case database maintenance will be performed

regularly merging highly similar cases or

removing unused cases.

As we mentioned above, the retrieval of similar

cases depends on the sentence similarity measure

between the current exception description and the

exceptions descriptions in the case database. In

addition, some CBR approaches e. g. (Weber and

Wild, 2005) use conversational scheme

(question/answer) with workflow participant to find

the best match between the case at hand and a

number of similar cases in case database. As these

questions and answers are short texts (sentence

length), again sentence similarity measures will be

very useful in retrieval of the best match. In next

section, a prototyping of the system and a case study

are presented.

5 A CASE STUDY

Car insurance claims handling is a process which

needs to be automated and managed efficiently and

be adaptable to the changing circumstances.

However, a fully automated solution for claim

processing cannot handle exception claims and does

not have real-time situational awareness capabilities.

Therefore, an effective exception handling

mechanism is needed. This is achieved by a

dedicated exception handling coordinator which can

be part of the workflow server or a separate node

connected with the workflow administrator.

To better illustrate how the proposed exception

handling mechanism works, a prototype of workflow

management system is developed to automate a

motor insurance claim process. The process consists

of 11 tasks involving 4 workflow participants

representing four roles in the insurance company.

These roles are: claim team, legal team, finance

team and technical team. Figure 1 shows the motor

insurance claim process at the modelling phase in

Petri notation. The workflow management system is

implemented in Java, and relies on Microsoft Access

database.

To prove the concept presented in this paper the

WFMS prototype carries out its run-time function

including the exception handling routine which

includes:

a) Instantiating a number of workflow instances

(claims) and coordination of tasks between the

workflow participants is practised

b) Exceptions are generated by the workflow

participants in the running instances and

exception handling procedures mentioned early

in the paper are examined.

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

260

0.2

0.4

0.6

0.8

1.2

1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73

workflow exceptions pairs

similarity

Figure 2: Exceptions pairs’ similarity measurement using

LSA.

0.2

0.4

0.6

0.8

1.2

1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73

workflow exceptions pair

similarity

Figure 3: Exceptions pairs’ similarity measured by human

domain expert

6 CONCLUSIONS

Workflow exceptions are often described by

workflow participants as a short text (sentence

length) in natural language. Therefore, sentence

similarity measures can be used to compute the

similarity between the exception at hand with the

exceptions stored in case database in order to find

the proper exception handler. This paper presented a

framework to apply semantic sentence similarity

measures within case-base reasoning paradigm in

workflow exception handling. We believe that

sentence similarity measures are capable techniques

in helping retrieve appropriate cases. This research

provides the starting point in the semantic sentence

similarity application in WFMSs. As the sentence

similarity measurement techniques are still active

research area, we expect new enhancements to those

techniques which will provide more accurate results.

We concentrated in this paper on presenting the

ideas, the concepts, the architecture and the initial

experimental results of our approach. Further work

and publication will include the following:

 Statistical and mathematical modelling of the

proposed approach and the obtained results to

compute the accuracy and the optimization of

our measurements

 Study the effect of the noise (noisy sentence

which includes missed words, grammatical error

and spelling error) and how the similarity

measurement techniques cope with this

 Apply the proposed approach to different

business processes from different domains.

REFERENCES

Aldeeb, A., Crockett, K. and Stanton, M., 2008. An Inter-

organizational Peer-to-Peer Workflow Management

System. In Proceeding of the 10

International

Conference on Enterprise Information Systems

(ICEIS-2008), pp. 85-92, Barcelona, Spain

Aliguliyev, R., 2009. A New Sentence Similarity Measure

and Sentence based Extractive Technique for

Automatic Text Summarization. Expert Systems with

Applications, 36, pp. 7764-7772.

Aminul, I. and Inkpen, D., 2008. Semantic Text Similarity

Using Corpus-based World Similarity and String

Similarity. ACM Transactions on Knowledge

Discovery from Data, Vol 2, No. 2, Article 10.

Atkinson, J., Mellish, C. and Aitken, S., 2004. Combining

Information Extraction with Genetic Algorithms for

Text Mining. IEEE Intelligent Systems, Vol. 19, no. 3,

Casati, F., Ceri, S., Paraboschi, S. and Pozzi, G., 2000.

Specification and Implementation of Exceptions in

Workflow Management Systems. ACM Transactions

on Database Systems, Vol. 24, No. 3, pp. 405-451

Erkan, G. and Radev, D., 2004. LexRank: Graph-Based

Lexical Centrality As Salience in Text Summarization.

J. Artificial Intelligence Research, vol. 22, pp. 457-

479

Feng, J., Zhou, Y. and Trevor, M., 2008. Sentence

Similarity based on Relevance. Proceedings of

IPMU'08, pp. 832-839

Grigori, D. et al., 2001. Improving Business Process

Quality through Exception Understanding, Prediction,

and Prevention. Proceedings of the 27

VLDB

Conference, Roma, Italy.

Hwang, G. and Lee, Y., 2005. A Flexible Failure-

Recovery Model For Workflow Management Systems.

International Journal of Cooperative Information

Systems. Vol. 14, No. 1 pp. 1-24

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

262

Hwang, S., Ho, S. and Tang, J., 1999. Mining Exception

Instances to Facilitate Workflow Exception Handling.

Proceedings, 6

International Conference on

Database Systems for Advanced Applications, pp. 45-

52.

Jablonski, S. and Bussler, C., 1996. Workflow

Management: Modelling Concepts, Architecture and

Implementation. International Thomson Publishing.

Jeng, B. and Liang, T. (1995) Fuzzy Indexing and retrieval

in case-based Reasoning Systems. Expert Systems with

Applications, 8, no. 1, pp. 135-142

Kelly, D., 2005. Where Do Business Process Exceptions

Come From? EbizQ,

http://www.ebizq.net/topics/bpm/features/6382.html?

&pp1, viewed on 10/06/2009 12:16

Klein, M. and Dellarocas, C., 2000. A Knowledge-based

Approach to Handling Exceptions in Workflow

Systems. Computer Supported Cooperative Work, 9:

pp. 399-412

Ko, Y. Park, J. and Seo, J., 2004. Improving Text

Categorization Using the Importance of Sentences.

Information Processing and Management, Vol. 40, pp.

65-79

Kumar, A. and Wainer, J., 2005. Meta Workflow as a

Control and Coordination mechanism for Exception

Handling in Workflow Systems. Decision Support

Systems 40, pp. 89-105.

Landauer, T., Foltz, P. and Laham, D., 1998. Introduction

to Latent Semantic Analysis. Discourse Processes,

vol. 25, nos. 2-3, pp. 259-284

Li, Y. et al., 2006. Sentence Similarity Based on Semantic

Nets and Corpus Statistics. IEEE Transactions on

Knowledge and Data Engineering, 18, pp. 1138-1150.

Limam, S., Marir, E. and Reijers, H., 2003. Electronic

Journal on Knowledge Management, Vol. 1, 2, pp.

113-124

Liu, Y. and Zong, C., 2004. Example-Based Chinese-

English MT. Proceedings of 2004 IEEE Conference,

Systems, Man, and Cybernetics, vols. 1-7, pp. 6093-

6096

Luo Z. et al., 2003. Exception Handling for Conflict

Resolution in Cross-Organizational Workflows.

Distributed and Parallel Databases, 13, 271-306.

Montani, S., 2009. Prototype-based Management of

Business Process Exception Cases. Applied

Intelligence,(online version), Springer Netherlands

O'Shea J., Bandar, Z., Crockett, K. and McLean, D., 2008.

A Comparative Study of Two Short Text Semantic

Similarity Measures. LNAI 4953, pp. 172-181.

O'Shea, K., Bandar, Z. and Crockett, K., 2008. A Novel

Approach for Constructing Conversational Agents

using Sentence Similarity Measures. Proceedings of

World Congress on Engineering, International

Conference on Data Mining and Knowledge

Engineering, London, pp. 321-326.

Pal, S. and Shiu, 2004. Foundations of Soft Case-based

Reasoning. John Wiley, NY

Pal, S., Polkowski, L. and Skowron, A., 2004. Rough-

Neural Computing: Techniques for Computing with

Words. Springer Verlag, Berlin

Sadiq, S. Orlowska, M. and Sadiq, W., 2005. Specification

and Validation of Process Constraints for Flexible

Workflows. Information Systems, 30, pp. 349-378.

Schmidt, R. and Vorobieva, O., 2008. Using Case-based

Reasoning To Explain Exceptional Cases. Proceedings

of 10

International Conference on Enterprise

Information Systems, pp. 119-124.

Shiu, S. and Pal, S., 2004. Case-Based Reasoning:

Concepts, Features and Soft Computing. Applied

Intelligence 21, pp. 233-238

Shiu, S. and Pal, S., 2004. Case-based Reasoning:

Concepts, Features and Soft Computing. Applied

Intelligence21, pp. 233-238

Watson, I., 1997. Applying Case-Based Reasoning:

Techniques for Enterprise Systems. Morgan Kaufmann

Publishers, San Francisco, USA.

Weber, B. and Wild, W., 2005. Conversational Case-based

Reasoning Support for Business Process Management.

American Association for Artificial Intelligence

(www.aaai.org).

SENTENCE SIMILARITY MEASURES TO SUPPORT WORKFLOW EXCEPTION HANDLING

263