Analyzing Multi-Source Social Data for Extracting
and Mining Social Networks
I-Hsien Ting Hui-Ju Wu
Department of Information Management Institute of Human Resource Management Department of Information Management
National University of Kaohsiung National Changhua University of
Kaohsiung City, Taiwan Education
iting@nuk.edu.tw Changhua, Taiwan
d94311001@mail.ncue.edu.tw
AbstractIn recent years, social computing has become a very
popular application in the Internet, and therefore large amount
of social (communication) data has been collected in different
social computing application. This paper will introduce a
methodology to collect and analyze multi-source social, and by
this for extracting social networks from the data. A system
architecture will also be presented in this paper to show how the
data can be collected, pre-processed, analyzed. Furthermore, the
system will allow the users to use the data as a resource for
personal decision support.
Keywords-Social Networking; Instant Messenger; E-mail; Data
Mining; Social Network Analysis
I. INTRODUCTION
With the rapid growth of Internet and communication
technologies, there are many communication and social
activities of people have been transferred to Internet-based
platform, e.g. e-mail communication, instant messaging
software and social networking websites (such as Blog and
web albums), etc. [8] Under this background, large amount of
personal communication and social data has been aggregated
and stored in different locations [15]. However, these valuable
data have not been well organized, treated and used. Thus, it is
an interesting research issue about how to use current
information techniques to process and analyze these data, such
as artificial intelligence, data mining or visualization technique
[18].
Social network analysis and construction are originally in
the research fields of Sociology. In recent years, many research
issues of information science and social networking have been
concerned due to the development of information techniques
and the requirements of data processing ability [13]. The target
of social network analysis and construction is relationship data
and it is therefore suitable to process and analyze
communication and social data that discussed previously [7].
Since the communication data, such as e-mails and the logs
of instant messenger, are very common data in our daily life.
However, there is less work focusing on how to organize,
Pei-Shan Chang
National University of Kaohsiung
Kaohsiung City, Taiwan
purplemio@gmail.com
process and analyze these data [31]. In this paper, we therefore
will propose method to analyze the common social data and
discuss how to extract social networks from multi-sources
social data dynamically. Furthermore, this paper will propose a
system architecture to use the three techniques of social
network analysis, social network construction and visualization
to process and analyze those valuable data. The system will
allow user to input tasks for dynamic social network analysis
and construction and the final results will be presented by
visualized mean and interface for decision support.
The structure of this paper is organized as below: In section
1, the background and introduction will be introduced. Some
related literatures of social network extraction, social network
and data mining and social network analysis will be reviewed
in section 2. A system architecture about how to extract
dynamic social networks from multi-sources data will be
proposed in section 3 as well as the introduction of the
components in the system. In section 4, we will focus on how
to extract social networks from social data and how to use data
mining and AI techniques for decision support. In section 5,
this paper will be concluded with the suggestions for future
research.
II. LITERATURE REVIEW
In this section, related literatures will be reviewed,
including social networks analysis, social networks extractoin
and social networking for decision support.
A. Social Networks Analysis
The research methodology of social network analysis is
developed to understand the relationship between actors, and
the term actor can be a person, an organization, an event or an
object [4]. In a social network, each actor is presented as a
node and each pair of nodes can be connected by lines to show
the relationships. The social network structure graph is a graph
that formed by those lines and nodes, and social network
analysis is therefore a methodology that used to understand the
graph and the relationships and actors in the social network
[11][34][27].
 There are three important elements that included in a social
network: actors, ties, and relationships. Actors are the essential
elements in the social network to define the people, events or
objects. Ties are used to construct the relationship between
actors by using a mean of path to establish the relationship
directly or indirectly. Ties can also be divided into strong tie
and weak tie according to the strength of the relationships; they
are also useful for discovering the subgroups of the social
network. Relationships are used to illustrate the interactions
and relationship between two actors. Furthermore, different
relationships may cause the network to reflect different
characteristics [32][33].
The most important measurements of SNA include network
size, diameter, density, centrality and structure holes [5]. Size
is a measurement to measure the amount of nodes or links in a
network, and the measurement of diameter is to measure the
amount of nodes between two nodes in a network. Density is
used to calculate the closeness of a network [23][28]. These
measurements are common used in many social network
related researches and will be used in this paper as well.
Traditionally, researches about SNA are mainly focus on
small group of actors and are process manually in most cases.
[6] However, with the rapid growth of Internet and web
techniques, more and more data have been collected and it has
become a hard task to process these data by only the mean of
manually [9]. Therefore, the scholars of information
technology and computer science are starting to devote related
researches to deal with these research issues [12][26].
Currently, the researches of computer science in SNA can be
divided into four main topics, including social networks
construction, social networks extraction, social networks
analysis and visualization.[24]
B. Social Networks Extraction
In the research field of information technology and
computer science in social networking, social networks
extraction is a subfield focusing on extract social networks
from large amount of communication data. With the rapid
growth of Internet and WWW, there are various kinds of data
have been generated due to communication purpose. The
common used communication data such as email
communication data, web usage logs, event logs, instant
messenger logs, logs of telecommunication, etc[22][29].
Currently, there are some researches which are focusing on
the extraction of these social data. For example, Bird et al.
propose a method to extract social networks from e-mail
communications [3]. Agrawal et. al using web mining
techniques to understand the behavior of users in newsgroup
[2]. Web is considered as the biggest database in the world, so
that various social networks can be extracted from this resource,
such as Furukawa et al. were trying to identify social networks
from blogspace [14][19] Jin et al. and Matsuo et al. developed
systems and tried to extract social networks from the web [17]
[21]. Adamic and Adar developed a method to discover the
relationship of friends and neighbours in the web [1].
Most of the researches that discussed above are focusing on
a single source for social network extraction. However, the
issue of how to extract social networks from different sources
has not been discussed well in related literatures. It is also a
hard task about how to integrate multi-source data for social
networking extraction. In addition to the problem of multi-
source data, instant messenger is a very popular and hot
software for people to send message and communication
recently. However, it has not been seen in recent research about
how to extract social networks from the data. These research
issues will be discussed in this paper.
C. Web Mining Techniques for Social Networking
According to different analysis targets and resources, the
web mining techniques can be divided into three different types,
which are Web Content Mining, Web Structure Mining and
Web Usage Mining [30].
Web content mining is a web mining technique to analyze
the contents in the web, such as texts, graphs, graphics, etc [2].
Recently, most of web content mining researches are focused
on the text data processing and few are focused on other
multimedia data. Natural language process is therefore the
main technology that used in this area. The concept and
techniques of Semantic Web and Ontology also have to be
studied [16][ 20].
Web structure mining is a technique that can be used to
analyze the links and structure of websites [10]. Graph theory
is usually the main concept and theory for web structure
mining to analyze and explain the structure of websites. In
addition, the extraction of the structure of websites is always
essential in this research area [12]. Therefore, its usually the
concern about how to design and implement a crawler (or
spider, bots) to extract and construct the structure of websites,
such as the research topic of Deep-web.
Web usage mining is a web mining technique that can be
used to analyze how the websites have been used, such as the
navigation behavior of users. The server-side Clickstream data
(logs file) is the main sources that used for web usage mining.
Client-side data (such as client-side logs file, cookies) is
sometimes to be used due to some research concerns, such as in
order to record more complete behavior of users. Different web
usage mining analyses include basic statistical analysis of the
navigation behavior of users in a website, such as how many
times the website has been browsed, where the users comes
from, etc. Furthermore, advanced web usage mining analyses
can also be provided, such as more complex analysis for
understand the navigation history of users in a website or cross-
website analysis [25].
 Figure 1. The architecture of the social network extraction system
III. SYSTEM ARCHITECTURE
According to the research background and motivation of
this paper, we have designed a system architecture to addressed
the raised issues. The system will allowed us to collect social
data from different sources, such as e-mail, instant messenger
and blog. The multi-source social data will then be pre-process
and analyze. The processed data can be used to extract social
networks automatically and dynamically. The system then can
be further developed to a decision support system. However,
this paper will not focus on the decision support system and
only the methodology to collect multi-source social data, pre-
processing and how to integrate the data for generating social
networks, which is the main contribution of this paper. The
architecture of the social network extraction system is
presented in figure 1.
As shown in figure 1, the system can be divided to two
major phases according to the characteristic of processing. The
two phases are offline data collection and processing and on-
line process. The elements and process of the two phases will
be introduced in detail below.
A. Offline data collection and processing
The first phase of the system is mainly working offline, and
there are three elements in this phase including multi-source
social data collection, data extraction engine and a database.
1) Multi-source social data collection
This is the first step of the system. In this paper, we intend
to collect social data which are most related to personal daily
communication. Thus, three types of data will be collected
including instant messenger data, email data and blog browsing
data. The messaging history of MSN or other messengers will
be recorded in a structural format, such as xml. The detail of
the message contents and communication target and time will
be stored in the file. In the paper, the history file of MSN
messenger is used as the social data of instant messenger.
About data collection, a data collection system will be
introduced in section 4 of the paper. The system is developed
by web-based concept and it allows the users to upload email
and the history file of MSN messenger by either automatically
or manually mean. About the collection of blog browsing data,
we use a client side agent to collect the navigation history of a
use when using particular blog. With the ability of client side
logging, the complete browsing data will be recorded without
missing, such as the behavior of browsing, posting a message
and responding to a message.
2) Data extraction engine
The second step of the system is data extraction engine.
The engine will firstly used to process the data that collected
from previous step. Then, useful data will be extracted and
filtered out from the raw data. The detail of how the data will
be processed and extracted will be discussed in section 4 of the
paper.
3) Database
After data collection and extraction, the output of the data
extraction enginer will then be stored in a database. The
database is designed according to the characteristics of
different sources of social data.
B. Online processing
The second phase of the system is a possible application of
the paper in the future. The works in this phase are mainly
processed online according to the data that collected offline.
The elements of this phase will be introduced as follow even
they are not the main focus of the paper.
4) Ontology-base
The second phase of the system is an possible application
of the paper. In the system, the user can use the collected social
data for personal decision support. It will allow the user to
input a keyword and other parameters to ask for decision
support. A social network for dealing with the problem will
then be generated, which provides possible solutions for the
user.
 The ontology base in the system is used to illustrate the
semantic of the keyword that use input. It is also a very
important element of the system for social network extraction.
5) User input:
There are two essential user input of the system, including a
keyword and other parameters. The parameters are limitation
and condition for the system to scale down the extracted social
network. For example, if the input keyword is BBQ, it means
the user may want to create a network about BBQ from the
social data to find out previous communications about the issue
of BBQ in email, MSN or blog browsing. Then, the other
parameters could be red wine and relationship >2. It means
that the user want to find friends to attend the BBQ party.
However, the participatns must attened to BBQ party before
more then twice and bring red wine with them.
6) SNA engine:
After the inputting of keywords and parameters, the system
will use SNA methods to get information for social networks
construction, such as the nodes, relationships, closeness, etc.
The detail of how to calculate the relationship value from the
multi-source social data will be discussed in next section.
7) SN Construction engine:
The SN construction will use the results of SNA to prepare
the visualization of the SN. For example, how much nodes will
be used in the SN, the characteristics of the links, the detail of
the network, etc. The gathered information will be used for SN
construction and visualization.
8) Visualization engine:
In this research,the library of OpenGL will be used to
visualize the network based on the information that provided
by SNA engine and SN construction engine.
9) Dynamic and task-oriented social network:
Finally, the system will generate a social network based on
the input of user, which is considered as a dynamic and task-
oriented social network.
IV. DATA COLLECTION SYSTEM AND EXTRACTION
EXTRACTION ENGINE
In this section, the paper will focus on three main tasks of
the system architecture that introduced previously. The three
tasks include the data collection system, data extraction
methodology and relationship calculation methods for social
networks construction.
A. The data collection system
Data collection is the first step of the system, and therefore
we design a sub-system for uploading related data. The system
is a web-based system, and it allows the user to upload email
file (in *.eml or text format), MSN history data (in *.xml
format) and client side logging data (in *.log or text format).
The format and file sample of email and MSN history data will
be introduced below.
About email file, the system allows users to upload email
file manually or automatically. Although different email agents
or servers may produce various email file, the system will
accept any email file and there are some common fields in
different email file format. Fields extraction of the uploaded
email file will be discussed later in this paper to show which
fields are useful for the research. Figure 2 shows a sample
email file which is saved by an email agent and figure 3 shows
the collected mail in the system.
Return-path: <eri@xx.xx.xx.xx>
Envelope-to: RSs@xx.xxxx.xx.xx
Received: from funnelweb.cs.york.ac.uk ([144.32.161.232]
Message-ID: <47552CF4.70806@xx.xxx.xx.xx>
Date: Tue, 04 Dec 2007 10:33:24 +0000
From: E Rid <eri@xx.xxx.xx.xx>
Reply-To: eri@xx.xx.xx.xx
User-Agent: Thunderbird 2.0.0.9 (Windows/20071031)
MIME-Version: 1.0
To: RSs@xx.xxx.xx.xx
Subject: java versus C benchmarks
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
Status: RO
Figure 2. A sample email file
Figure 3. The email collection system
In this research, only the instant messenger history file of
MSN is accepted. The MSN history file is stored in a .xml file
and based on the format of xml. Each contactor in the MSN
contactor list has an independent history file, and the
information that stored in the file include session ID, Date and
Time, from, to and message content. Figure 4 shows a sample
MSN history file that used in the paper.
<?xml version="1.0"?>
<?xml-stylesheet type='text/xsl' href='MessageLog.xsl'?>
<Log FirstSessionID="1" LastSessionID="12">
<Message Date="2009/7/4" Time=" 12:50:55" DateTime="2009-
07-03T16:50:55.390Z" SessionID="31">
<From><User FriendlyName="Want SAP, Netweaver, J2EE
consultant"/></From>
<To><User FriendlyName="Derrick-"/></To><Text Style="font-
family:; color:#000000; ">
ok...I got it...
</Text>
</Message>
</Log>
Figure 4. A sample MSN history file
B. Data extraction methodology
In order to filter-out unnecessary data from the collected
email and MSN history file, we developed a data extraction
methodology to extract useful data for constructing social
networks. In section 4.B, we have introduced a sample email
file format. However, different email agents and servers may
have different email format. Thus, we have selected some
important fields in the email file which are useful for us to
calculate the social relationship between the communicators of
the collected emails and MSN history.
From the email file, some necessary fields will be extracted,
including deliver-to, receive-id, date, to, from, W5 are the weight values for each email relationship.
subject, msg-id, priority, reply-to, mailer (agent),
encode. content-type, content, cc. These fields will be E i W1 E send W 2 E receive W 3 E forward W 4 E cc W 5 E coreceiver
extracted from the original. Some of the extracted fields are
used to identify emails and some are important for relationship
measurement.
In addition to the extraction of email fields, we also extract
useful fields from the MSN history file. The fields will be
extracted from the file, including msn-from, msn-to, msn-
content, msn-datetime, msn-id, msn-sessionid, msn-
totage. Among all of the extracted fields, the msn-sessionid
field is used to record the session number and msn-totage
field is used to identify a communication with multi-users.
C. From social data to social networking
After collecting and extracting useful information from the
muti-source social data. The relationship can then be calculated.
When user inputting keywords and parameters, the system will
match the keyword and the ontology based to find related email
records, MSN messages and blog content.
Figure 5. Communication frequency table
The relationship (communication frequency) is the most
important element to form a social network. Thus, a series of
formulations for measure the relationship for different kinds of
social communication will be discussed in this section. The
data collection system that introduced in section 4.A also has
the ability to measure the frequency for the input keyword as
presented in figure 5.
The relationship of two nodes can be measured by
formulation (1). Ri means the relationship from a specific node
(a keyword with parameters) to a node i in a social network. Ei
mean the email relationship for node i, Mi means the MSN
relationship for node i, and Bi means the Blog relationship for
node i. W1, W2, W3 are three difference weight value to
measure the importance of each relationship.
Ri W1 E i W 2 M i W 3 Bi (1)
The details of relationship E, M and B can be measured by
formulation 2, 3, 4, and 5. In formulation 2, Esend means the
frequency about how many how many mail send to node i,
Ereceive means how many mail receive from i, Eforward is how
many main forward from i, Ecc denotes how many mail
received as cc from i, and Eco-receiver means how many mail
received as co-receiver from i. Furthermore, W1, W2, W3, W4,
(2)
Formulation 4 is the complete formulation for MSN
relationship measurement, Msend means a message send to node
i, Mreceive is a message received from i, Mmulti denotes a
message send for multi-communicators at the same time and
Minteraction is an advanced interaction between node i such as
video conference, file sharing, etc. However, only the
measurements Msend , Mreceive , Mmulti are used in the paper as
the formulation 3. In formulation 3 and 4, W1, W2, W3, W4 are
weight values for each MSN relationship.
M i W1 M send W 2 M receive W 3 M multi (3)
M i W1 M send W 2 M receive W 3 M multi W 4 M interaction (4)
For Blog relationship, the measurement is shown in
formulation 5. Bbrowsing means browsing frequency to a Blog of
node i, Bbookmarking is the frequency of adding a Blog of node i to
bookmark and Binteractoin is the frequency of interaction in the
Blog of node i, such as response to a Blog entry.
Bi W1 Bbrowsing W 2 Bbookmarking W 3 Binteraction (5)
The formulations above are helpful for most cases to
measure the relationships of the nodes to a specific node in a
social network. Then the system can use the analyze results to
generate and visualize the social network.
V. CONCLUSION AND FUTURE RESEARCH
Social and communication data are very common data in
our daily life; however these data have not been used well for
use to make decision. In this paper, we firstly provide an
overview about the characteristics of these data and to illustrate
how to use the concept of social networking and web mining to
analyze the data. A system architecture is then provide to give
the reader a picture about how to use the multi-source social
data to generate dynamic and task-oriented social networks and
by this to assist the decision making. More detail process of
data collection, data extraction and relationship measurement
in the system are also provided.
In the future, we will try to move our research focus to the
second part (online processing) of the system and to implement
the decision support system. Furthermore, we will try to study
how to use the techniques of web mining to get better analysis
results and to enhance the accuracy of the decision support
system. In addition, we will also try to understand more social
data sources which are useful for including in the system and
our future research.
ACKNOWLEDGMENT
This work is partially supported by a NSC research grant,
TAIWAN (NSC 97-2410-H-390-022).
REFERENCES
[1] Adamic, L. A., and Adar, E. (2007) Friends and Neighbors on the
Web Social Networks, Vol. 25, 2007, pp. 211-230
[2] Agrawal, R., Rajagopalan, S., Srikant, R., and Xu, Y. (2003) Mining
Newsgroup Using Networks Arising From Social Behavior In
Proceedings of World Wide Web 2003 Conference, Budapest, Hungary,
pp. 529-535
[3] Bird, C., Gourley, A., Devanbu, P., Gertz, M. and Swaminathan, A.
(2006) Mining Email Social Networks In Proceedings of MSR 2006,
May 22-23, 2006, Shanghai, China.
[4] Borgatti, S. P., and Everett, M.G. (2002) Ucinet for Windows: Software
for Social Network Analysis, Harvard: Analytic Technologies.
[5] Burt, R.S., (1992). Structural Holes, Harvard University Press,
Cambridge,MA.
[6] Cooley, R. Mobasher, B. and Srivastave, J. (1997) Web Mining:
Information and Pattern Discovery on the World Wide Web In
Proceedings of the 9th IEEE International Conference on Tool with
Artificial Intelligence, 1997, pp. 558-567, Newport Beach, CA, USA
[7] Cross, R. and Parker, A (2004). The Hidden Power of Social
Networks, Harvard University Press
[8] Chin, A. and Chignell, M. (2006) Finding Evidence of Community
from Blogging Co-Citations: A Social Network Analytic Approach In
Proceedings of the IADIS International Conference on Web Based
Communities 2006, San Sebastian, Spain, February 26-28, 2006
[9] Churchill, E. F., and Halverson, C. A. (2005) Social Networks and
Social Networking IEEE Internet Computing, September/October 2005,
pp.14-19
[10] Dingt C. H. Q., Zha, H., Husbands, P., and Simont, H. D. (2004) Link
Analysis: Hubs and Authorities on the World Wide Web SIAM
Review, Vol. 46, No. 2, pp. 256-268.
[11] Freeman, L., Centrality in Social Networks: Conceptual Clarification,
Social Networks, 1979.
[12] Fu, F., Chen, X., Liu, L., and Wang, L. (2007) Social Dilemmas in An
Online Social Network: The Structure and Evolution of Cooperation
Physics Letters A, Vol 371, 2007, pp. 58-64
[13] Fu, F., Liu, L., Wang, L. (2008) Empirical Analysis of Online Social
Networks in the age of Web 2.0 Physics Letters A , Vol. 387, 2008, pp.
675-684
[14] Furukawa, T., Matsuo, Y., Ohmukai, I., Uchiyama, K., Ishizuka, M.
(2007) Social Networks and Reading Behavior in the Blogosphere In
Proceedings of ICWSM 2007, Boulder, Colorado, USA, pp. 51-58
[15] Garton, L., Haythornthwaite, C., and Wellman, B.( 1997) Studying
Online Social Networks, Journal of Computer Mediated
Communication (3:1).
[16] Godbole, N., Srinivasaiah, M., Skiena, S.: Large-Scale Sentiment
Analysis for News and Blogs. In: Proceedings of ICWSM 2007, Boulder,
Colorado, USA (2007)
[17] Jin, Y. Z., Matsuo, Y., and Ishizuka, M. (2007) Extracting Social
Networks among Various Entities on the Web In Proceedings of the
Fourth European Semantic Web Conference, 2007
[18] Kumar, R., Novak, J., and Tomkins, A. (2006) Structure and Evolution
of Online Social Networks In Proceedings of KDD 2006 Conference,
August 20-23 2006, Philadelphia, Pennsylvania, USA, pp. 611-617
[19] Lento, T., Welser, H. T., Gu, L., and Smith M. (2006) The Ties that
Blog: Examining the Relationship Between Social Ties and Continued
Participation in the Wallop Weblogging System In Proceedings of the
15th International World Wide Web Conference, May 23-26 2006,
Edinburgh, Scotland
[20] Mika, P.: Flink: Semantic Web Technology for the Extraction and
Analysis of Social Networks. Web Semantics 3(2-3), 211223 (2005)
[21] Matsuo, Y., Mori, J., Hamasaki, M. (2006) POLYPHONET: An
Advanced Social Network Extraction System from the Web In
Proceedings of 2006 Internet World Wide Web Conference, May 23-26,
Edinburgh, Scotland.
[22] Mislove, A., Marcon, M., Gummadi, K. P., Druschel, P. and
Bhattacharjee, B. (2007) Measurement and Analysis of Online Social
Networks In Proceedings of 2007 Internet Measurement Conference,
October 24-26, 2007, San Diego, California, USA, pp. 29-42
[23] Mitchell, J. C. (1969) Social Networks and Urban Situations
Manchester University Press, England
[24] Mutton, P. (2004) Inferring and Visualizing Social Netwokrs on
Internet Relay Chat In Proceedings of the Eighth International
Conference on Information Visualization, pp. 25-43
[25] Pierrakos, D., Paliouras, G., Papatheodorou, C., Spyropoulos,
C.D.: Web Usage Mining As A Tool for Personalization: A
Survey. User Modelling and User Adapted Interaction 13(4),
311372 (2003)
[26] Sarkar, P., and Moore, A.W. Dynamic Social Network Analysis Using
Latent Space Models SIGKDD Explorations, Vol. 7, Issue 2, pp. 31-40.
[27] Scott, J. (2000) Social Network Analysis: A Hand Book (2nd ed.),
SAGE publication, 2000
[28] Scott, J. (2002) Social Network Analysis: Critical Concepts in
Sociology Routledge, New York, USA
[29] Tang, J., Zhang, D. and Yao, L. (2007) Social Networking Extraction
of Academic Researchers In Proceedings of the Seventh IEEE
International Conference on Data Mining, pp.292-301.
[30] Ting, I. H. (2008) Web Mining Techniques for On-line Social
Networks Analysis In Proceedings of the 5th International Conference
on Service Systems and Service Management, Melbourne, Australia, 30
June-2 July 2008, pp. 696-700
[31] Turoff, M., Hiltz. S. R., Cho, H. K., Li, Z., and Wang, Y. (2002) Social
Decision Support Systems (SDSS) In Proceedings of the 35th Hawaii
International Conference on System Sciences, pp. 1-10.
[32] Wasserman, S., and Faust, K. (2003) Social Network Analysis: Method
and Applications Cambridge University Press, Great Britain 2003
[33] Wellman, B. and Berkowitz, S. D. (ed.), (1988) Social structures: A
network approach, Cambridge University Press, pp. 19-61
[34] Wasserman, B., and Faust, K. Social Network Analysis : Methods and
Applications.New York: Cambridge University Press, 1994.