A CONVERSATIONAL EXPERT SYSTEM SUPPORTING

BULLYING AND HARASSMENT POLICIES

Annabel Latham, Keeley Crockett and Zuhair Bandar

Intelligent Systems Group, The Manchester Metropolitan University, Chester Street, Manchester M1 5GD, U.K.

Keywords: Expert system, Conversational agent, Knowledge engineering, Knowledge tree, Bullying, Harassment.

Abstract: In the UK, several laws and regulations exist to protect employees from harassment. Organisations

operating in the UK create comprehensive and carefully-worded bullying and harassment policies and

procedures to cover the key aspects of each regulation. In large organisations, such policies often result in a

high support cost, including specialist training for the management team and human resources (HR)

advisors. This paper presents a novel conversational expert system which supports bullying and harassment

policies in large organisations. Information about the bullying and harassment policies and their application

within organisations was acquired using knowledge engineering techniques. A knowledge tree is used to

represent the knowledge intuitively, and a dynamic graphical user interface (GUI) is proposed to enable the

knowledge to be traversed graphically. Adam is a conversational agent allowing users to type in questions

in natural language at any point and receive a simple and direct answer. An independent evaluation of the

system has given promising results.

1 INTRODUCTION

Organisations in the UK have policies to enforce

laws which protect employees from bullying and

harassment. Designed to support several laws and

regulations, such policies and procedures must be

comprehensive and carefully-worded. These

policies often result in a high support cost, requiring

specialist training across the organisation. Victims

of bullying and harassment may be reluctant to

discuss such sensitive issues within the organisation,

fearing lack of confidentiality, bias by support staff

and repercussions. The cost to organisations of not

dealing with bullying can be high – bullying at work

is estimated to be costing the UK National Health

Service more than £300m in sick days and

recruitment costs (BBC.co.uk 2008). The rather

inert nature and high support cost of bullying and

harassment policies means they lend themselves to

automation. An expert system would offer

anonymous, 24-hour access to advice which was

consistent, appropriate and valid (avoiding human

bias), and would be generic, offering cross-sector

application across the UK.

An ability to enter into a natural language

dialogue with an expert system would be the

ultimate development in a user interface, giving

users direct and non-linear access to expert

knowledge. This is especially the case in heavily

legislated areas where the language used may be

difficult to understand. Human computer interface

designers have used the metaphor of face-to-face

conversation for some time, but have only recently

attempted “to design a computer that could hold up

its end of the conversation” (Cassell 2000). A

conversational agent (CA) is a computer program

which can interact using natural language (Cassell

2000). There are three main development

approaches for CAs – natural language processing

(NLP), pattern matching or artificial intelligence

(AI). NLP studies the constructs and meaning of

natural language, applying rules to process

information contained within sentences (Khoury et

al 2008). Pattern matching uses an algorithm to find

the best match for key words and phrases within an

utterance (Pudner et al 2007). The AI approach

compares the semantic similarity of phrases to

decide on the meaning of the input (O’Shea et al

2008; Li et al 2006), enabling a CA to cope with

input which is not grammatically correct or

complete. This paper presents a novel conversational

expert system (CES) which models bullying and

harassment policies and procedures in large

163

Latham A., Crockett K. and Bandar Z. (2010).

A CONVERSATIONAL EXPERT SYSTEM SUPPORTING BULLYING AND HARASSMENT POLICIES.

In Proceedings of the 2nd International Conference on Agents and Artiﬁcial Intelligence - Artiﬁcial Intelligence, pages 163-168

DOI: 10.5220/0002697801630168

 SciTePress

organisations. Knowledge engineering techniques

are used to acquire information about the policies

and how they are applied. A dynamic GUI is

proposed for the CES, which changes according to

the user’s need, enabling the knowledge to be

traversed graphically and which allows natural

language questions to be asked at any point.

This paper is organised as follows: Section 2

outlines bullying and harassment, Section 3

describes CAs and their applications, Section 4

explains the methodology for constructing the CES,

Section 5 summarises the evaluation of the CES and

Sections 6 and 7 include the results, discussions and

conclusions.

2 BULLYING & HARASSMENT

Bullying and harassment is widespread within the

workplace (Burnes & Pope 2007, Lee 2000). In

2005, a national survey of the UK National Health

Service (NHS) staff found that 15% of staff had

experienced bullying, harassment or abuse from

colleagues and 26% experienced these problems

from patients (Healthcare Commission 2006). In

addition to the health implications for victims,

workplace bullying and harassment costs

organisations lost productivity, sick leave, high staff

turnover and the cost of replacing staff, as well as

the possibility of legal actions by employees (Unite

the Union 2007).

In the UK, no single piece of legislation covers

bullying and harassment. Organisations must adhere

to several different laws protecting employees from

harassment due to diverse causes (ACAS 2007).

Implementing this legislation has led organisations

to develop comprehensive policies, often resulting in

large and complex policy documents. These policies

often require additional training and support for

employees understanding and following the process

of reporting bullying and harassment. The high

support cost and fairly static nature of these policies

suggests that automation would be advantageous. A

CES would reduce costs and add benefits such as

allowing anonymous, 24 hour access to information

ensuring that the advice given is consistent,

appropriate and valid.

3 CONVERSATIONAL AGENTS

Communicating with a computer using natural

language has been a goal in artificial intelligence for

many decades, stimulated by the ‘Turing Test’

(Turing 1950). Conversational agents designed with

the sole aim of holding a conversation are known as

chatbots (Carpenter 2007). In the context of this

paper, CAs are tools for addressing specific

problems, such as ConvAgent’s InfoChat, a goal-

oriented CA (Michie & Sammut 2001).

CAs are ideally suited to simple question-

answering systems (Sadek 1999, Convagent 2005,

Owda et al 2007) as they are intuitive to use and

allow users direct, non-linear access to information.

A natural language questioning facility in an expert

system would additionally widen access to the

expertise. The role of a CA in the context of this

paper is to accept natural language questions and

produce an appropriate response. A pattern

matching CA is suited to systems where

conversations are restricted to specific knowledge

areas, such as bullying and harassment policies. The

pattern matching approach requires the development

of conversation scripts, a similar idea to call centre

scripts, which match key input words and phrases to

suitable responses. CA scripts are often organised

into contexts and may be linked in a tree or network

structure, offering powerful pattern matching at

various levels (e.g. a script level to respond to

abusive language). The CA receives an input and

searches scripts to find the best matching response.

Different contexts and conversation histories are

used to help find appropriate matches, for example,

the meaning of a user utterance “Yes, please show

me” can only be understood in relation to the current

context and previous utterances of the conversation.

4 CONSTRUCTING THE CES

This section describes the methodology and key

stages in constructing the CES. Figure 1 shows the

overall structure of the CES. A central controller

communicates with the knowledge base, dynamic

GUI and CA to manage user interactions. The

knowledge base contains expert knowledge and was

constructed in two phases. First expert knowledge

was gathered (knowledge engineering, described in

Section 4.1) and then that knowledge was structured

(explained in section 4.2). The CES user interface

has two main components – the dynamic GUI

displays a graphical representation of the knowledge

which changes according to user choices (outlined in

section 4.3), and the CA allows natural language

questions to be asked (described in section 4.4).

ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence

164

Figure 1: CES system structure.

4.1 Knowledge Engineering

Knowledge engineering is the stage of expert system

development concerned with the acquisition and

elicitation of knowledge from experts, and was done

in parallel with requirements gathering (Tuthill

1990). A domain review of current UK harassment

legislation, policies and guidance was undertaken

and a real-life large organisation’s Bullying and

Harassment policy was adopted. The documentation

study was followed up by meetings with the

organisation’s management team to elicit further

knowledge and clarification on points arising from

the review. An analysis of knowledge acquired

about the policy and its use by employees identified

four key categories of user problem, forming the

basis of the knowledge represented by the CES.

Each category leads to a series of additional

questions to further identify the problem. The four

main questions to be addressed by the CES are:

 What counts as bullying and harassment?

 How do I report bullying or harassment?

 What can I do if I am not being taken

seriously?

 What can I do if I am being victimised?

4.2 Structuring Knowledge

The knowledge base underpins a CES and therefore

must accurately reflect the expert knowledge and its

use in practice. A knowledge tree was the most

suitable representation of the knowledge for this

system as it follows a path of questions and feels

natural to users. Knowledge trees are biological

representations which consist of nodes, where values

are assessed, branches to be followed according to

node values, and leaves which terminate the branch

and represent a conclusion. Information is gathered

by the CES based on user choices, which determine

the value of each node, and therefore the branch to

follow and the next question to ask (the next node).

The knowledge tree was designed iteratively and

modelled using Convagent’s InfoRule TreeTool

(Convagent 2005) which is an interactive system

that graphically displays the tree. The TreeTool

facility to simulate the running of the tree was used

so that experts could verify its accuracy. Figure 2

shows a small section of the knowledge tree within

the TreeTool. The entire tree has more than 1000

nodes over 11 levels, which illustrates the

complexity of the problem. Within the Decision

Tree folder, the main tree is called ‘bullying’, and

the start node of the tree is att_query_type. The start

node has four branches, one for each of the main

questions modelled. The not_taken_seriously

branch leads to a leaf node, a CES response which

advises the user to discuss their case with a specialist

advisor. Also seen in Figure 2 is a subtree,

follow_up, which is used to avoid repetition of nodes

within the tree (similar to a macro).

Figure 2: CES knowledge tree.

4.3 Graphical User Interface

The CES users will range in experience from novice

to expert so the user interface must enable two-way

communication for all users (McGraw 1992); if

users cannot use the CES as an effective consultant

it will not be successful. Graphical interfaces are

more intuitive to use for novice users. The CES is

destined to be run on various web browsers over the

Internet or within company intranets. The two main

user interface elements will now be described.

4.3.1 Dynamic GUI

The GUI allows users to view a representation of the

knowledge and select appropriate options.

Depending on the user’s position in the knowledge

tree, the graphical representation changes to reflect

the available options. This dynamic interaction was

A CONVERSATIONAL EXPERT SYSTEM SUPPORTING BULLYING AND HARASSMENT POLICIES

165

intended to feel natural like a conversation rather

than a linear procedure.

Figure 3: Dynamic user interaction.

Figure 4: CES response.

The screen was divided horizontally into three

main areas (see Figure 3): the top area displays

Adam’s instructions and advice, the centre shows a

graphical representation of the knowledge tree and

the bottom area allows natural language questions to

be entered. Adam asks questions to gather

information about the user’s query, and the user

responds by clicking one of the highlighted answer

boxes. The CES dynamically accesses the

knowledge base and responds with a new question

and set of answer boxes. Figure 3 shows an

interaction where the user is exploring how to report

bullying, having taken no action so far (previous

answers are shown graphically using red triangles

and lines). The user wants to take informal action,

so clicks the ‘Informal’ answer box (Figure 4), and

is given guidance by Adam, shown graphically using

an ellipse. At any point the user can restart the

session or go back and explore the knowledge

further, using buttons at the foot of the screen. The

dynamic GUI for the CES was designed with system

reuse in mind as its display draws its information

directly from the knowledge base. Amendments

made to the knowledge tree are automatically shown

graphically on screen. This offers easy system

maintenance and the option of applying new

knowledge bases, e.g. finance policies.

4.3.2 Conversational Agent

The CA component of the CES integrates a natural

language question-answering function. The ability

to ask questions widens access to the system by

allowing users to employ their own vocabulary. The

CA also offers users direct access to knowledge

without tracing through the CES questions.

Convagent’s InfoChat CA (Michie & Sammut

2001), a goal-oriented CA employing natural

language pattern matching, was chosen. A list of

frequently asked questions (FAQs) about bullying

and harassment was compiled during the domain

review. A number of scripts were developed which

contained rules with patterns of keywords and

phrases to match each FAQ to an appropriate

response. A lower level of script was designed to

respond to unknown or abusive language. An

example rule from one of the InfoChat scripts is

shown below. In the rule, a is the activation level

used for conflict resolution (Sammut 2001); p is the

pattern strength followed by the pattern and r is the

response. Also seen in the example is the wildcard

(*) and macros (<explain-0>) containing a number

of standard patterns, each matched separately.

<Rule-01>

a:0.01

p:50 *<explain-0> *bullying*

p:50 *bullying* <explain-0>*

p:50 *<explain-0> *a bully*

p:50 *a bully* <explain-0>*

r: Bullying is persistent, threatening, abusive,

malicious, intimidating or insulting behaviour,

directed against an individual or series of

individuals, or a group of people.

Figure 5: Asking a question.

ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence

166

At any stage within the CES, the user could type

natural dialogue into an on-screen box to “chat”

directly with Adam. Figure 5 shows an example

interaction with Adam, continuing the previous

interaction (Figures 3 and 4) by clarifying Adam’s

response. The user types their question, ‘Which

senior manager should I meet’, into the question box

and presses ‘Ask Adam’. The CA has been scripted

to understand this question, and Adam provides an

appropriate answer at the top of the screen.

5 USABILITY EVALUATION

The knowledge tree was evaluated by the

organisation’s management team, who simulated its

execution using the TreeTool. The CES was

evaluated by the HR policy experts within the

organisation and an independent graphic designer

specialising in websites. During the evaluation, a

questionnaire was used to record feedback on system

usability, appearance and functionality.

Additionally a representative group of users were

selected to evaluate the CES, all employees in

different roles for different companies, aged between

25 and 45 and experienced in interacting with HR

departments. The group included both men and

women from administrative, academic and

managerial backgrounds. A scenario of possible

sexual harassment was developed along with a

questionnaire to record feedback on the appearance,

usability and functionality of the CES. The group

were asked to read the scenario and then complete

two tasks within the system – find out whether the

behaviour described is sexual harassment and how to

stop the behaviour without jeopardising your job.

Users then completed the questionnaire

anonymously and the results were analysed. Some

of the users were unobtrusively observed during

evaluation, giving qualitative feedback about facial

expressions and speed of use. The results of the

evaluation will now be discussed.

6 RESULTS AND DISCUSSION

Results of the independent evaluation show that the

CES was generally well received, being

understandable and intuitive to use. Table 1 shows

the average results for each section of the

questionnaire. The lowest score (of 5) was given to

the typeface used within the GUI, although the

average score was 7.8. The highest average score of

9.33 indicated that users had found the information

they sought without difficulty (Table 1, Section 2,

Question 6). One user commented that they “did not

need to ‘learn’ to use the advisor”. The next highest

average score of 9 was given to the clarity of

instructions, ease of navigation and use of Restart

and End buttons. 50% of users suggested

differentiating Adam’s instructions from the advice

given would improve the usability of the system, for

example using a different coloured font to make it

clear when users are expected to respond.

Table 1: User evaluation results.

Section 1 – Screen Design Avg Score

(out of 10)

1. Look and feel 8.5

a. Colour 8.17

b. Size 8.5

c. Typeface 7.83

2. Placing of buttons and boxes 8.5

3. Clarity of instructions 9

4. Central option buttons

a. Size of buttons (width) 8.67

b. Positioning of buttons 8.67

c. Connecting lines (width) 7.83

d. Indication of current selection 8

Section 2 – Functionality

1. Adam’s text for instructions

a. Readable? 8.5

b. Noticeable? 8

2. Ease of navigation 9

3. Restart & End buttons 9

4. Clicking of option buttons 8.83

5. Typing of text and Ask Adam button 8.33

6. Did you find the information and

advice you wanted?

9.33

7. How useful did you find the advice

given to you?

7.83

Observation of the tests revealed two interesting

issues. 33% of users attempted to click the response

icon, and were expecting to be taken to an email or a

form to arrange to see an advisor. Also, 33% of

users pressed the browser’s Back button, which

caused the tree buttons to behave unpredictably

(causing an error in one case). Ease of use is critical

to the success of a CES to support a Bullying and

Harassment policy. Users felt that the system was

intuitive and easy to use, giving high scores to ease

of navigation and the ability to find information.

The results have shown that the system allows users

to access information about sensitive issues like

bullying anonymously and at any time, aiding

organisations in positively implementing harassment

policies and improving workforce cohesion.

A CONVERSATIONAL EXPERT SYSTEM SUPPORTING BULLYING AND HARASSMENT POLICIES

167

7 CONCLUSIONS AND

FURTHER WORK

This paper has presented a novel conversational

expert system which has been applied to a real world

application to advise employees on bullying and

harassment policies. The CES features a graphical

representation of knowledge which may be explored

dynamically by the user, and a facility to ask natural

language questions and receive a response. During

development, the CES was generalised, meaning it

can adapt to a different, unknown knowledge base.

As well as improving the ease of maintenance,

similar applications can be developed more rapidly

by reusing the GUI component and limiting new

development to the knowledge base and

conversational scripts. The system was evaluated by

a representative group of users, who found the

system intuitive to use and the information easy to

access. Implementation of the system could benefit

both employees and the organisation in providing

anonymous access to information about the bullying

and harassment policy at any time of day, leading to

a positive implementation of the policy and a more

cohesive group of employees.

In future a pilot evaluation study of the CES is

required before conclusions may be drawn about the

usability of the system compared to telephone advice

or meeting advisors. It is then planned to develop

new CES for different HR procedures. Future

improvements include expanding the CA scripting to

include further definitions and explanations,

extending the system for multiple users, and

investigating the use of different CAs which employ

less labour-intensive techniques for developing

dialogue scripts.

ACKNOWLEDGEMENTS

The authors thank Convagent Ltd for the use of their

InfoChat and TreeTool software and Mr John Shiel

of Krann Ltd for his evaluation of the GUI.

REFERENCES

ACAS 2007, Advice leaflet - Bullying and harassment at

work: guidance for employees, ACAS, viewed 24 May

2007,

http://www.acas.org.uk/index.aspx?articleid=797

BBC.co.uk 2008, ‘NHS Bullying costing millions’, BBC,

viewed 10 March 2008, http://search.bbc.co.uk/cgi-

bin/search/results.pl?scope=all&tab=all&recipe=all&q

=Bullying+NHS+2008&x=0&y=0

Burnes, B & Pope, R 2007, ‘Negative behaviours in the

workplace: A study of two Primary Care Trusts in the

NHS’, International Journal of Public Sector

Management, vol. 20 no. 4, pp. 285-303.

Carpenter, R 2007, jabberwacky – live chat bot, viewed 20

June 2007, http://www.jabberwacky.com/

Cassell, J 2000, Embodied Conversational Agents, MIT

Press, Cambridge MA.

Convagent Ltd 2005, Convagent, viewed: 23 May 2007,

http://www.convagent.com/demo.htm

Healthcare Commission 2006, National survey of NHS

staff 2005 Summary of key findings, Commission for

Healthcare Audit and Inspection, London.

Khoury, R, Karray, F & Kamel, MS 2008, ‘Keyword

extraction rules based on a part-of-speech hierarchy’,

International Journal of Advanced Media and

Communication, vol. 2 no. 2, pp. 138-153.

Lee, D 2000, ‘An analysis of workplace bullying in the

UK’, Personnel Review, vol. 29 no. 5, pp. 593-612.

Li, Y, McLean, D, Bandar, Z, O’Shea, J & Crockett, K

2006, ‘Sentence Similarity Based on Semantic Nets

and Corpus Statistics’, IEEE Transactions on

Knowledge and Data Engineering, vol. 18, no. 8, pp.

1138-1150.

Michie, D & Sammut, C 2001, Infochat Scripter’s

Manual, Convagent Ltd, Manchester, UK.

O’Shea, K, Bandar, Z & Crockett, K 2008, ‘A Novel

Approach for Constructing Conversational Agents

using Sentence Similarity Measures’, World Congress

on Engineering, International Conference on Data

Mining and Knowledge Engineering, pp. 321-326.

Owda, M, Bandar, Z & Crockett, K 2007, ‘Conversation-

Based Natural Language Interface to Relational

Databases’, IEEE/WIC/ACM International Conference

on Web Intelligence and Intelligent Agent

Technology, pp. 363-367.

Pudner, K, Crockett, K & Bandar, Z 2007, ‘An Intelligent

Conversational Agent Approach to Extracting Queries

from Natural Language’, World Congress on

Engineering, International Conference of Data Mining

and Knowledge Engineering, pp 305-310.

Sadek, D 1999, ‘Design Considerations on Dialogue

Systems: from theory to technology – the case of

Artimis’, in ESCA Workshop: Interactive Dialogue in

Multi-modal Systems, p. 173, Kloster, Irsee.

Sammut, C 2001, ‘Managing Context in a Conversational

Agent’, Electronic Transactions on Artificial

Intelligence, Vol. 3, No. 7, pp. 1-7.

Turing, A 1950, ‘Computing machinery and intelligence’,

Mind, vol. LIX, no. 236, p. 433.

Tuthill, GS 1990, Knowledge Engineering: Concepts &

Practices for Knowledge-Based Systems, TAB Books

Inc, Blue Ridge Summit, PA.

Unite the Union 2007, Dignity at Work – tackling bullying

in the workplace, Unite the Union, viewed: 12 August

2008, http://www.dignityatwork.org/

ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence

168