U.S. patent application number 10/100491 was filed with the patent office on 2002-12-12 for method and system for informing users of subjects of discussion in on -line chats.
Invention is credited to Gruen, Daniel M., Sheldon, Mark A., Vaithaynathan, Shivakumar.
Application Number | 20020188681 10/100491 |
Document ID | / |
Family ID | 22502493 |
Filed Date | 2002-12-12 |
United States Patent
Application |
20020188681 |
Kind Code |
A1 |
Gruen, Daniel M. ; et
al. |
December 12, 2002 |
Method and system for informing users of subjects of discussion in
on -line chats
Abstract
A method for informing a user of topics of discussion in a
recorded chat between two or more people is described. The method
includes the steps of identifying elements from the chat having
similar content, labeling some or the identified elements as
topics, and presenting the topics to the user. Identifying elements
from the chat having similar content includes decomposing the chat
into utterances made by the people involved in the chat and
clustering the utterances using document clustering techniques on
each utterance to identify elements in the utterances having
similar content.
Inventors: |
Gruen, Daniel M.; (Newton,
MA) ; Sheldon, Mark A.; (Arlington, MA) ;
Vaithaynathan, Shivakumar; (San Jose, CA) |
Correspondence
Address: |
BROWN, RAYSMAN, MILLSTEIN, FELDER & STEINER LLP
900 THIRD AVENUE
NEW YORK
NY
10022
US
|
Family ID: |
22502493 |
Appl. No.: |
10/100491 |
Filed: |
March 18, 2002 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
10100491 |
Mar 18, 2002 |
|
|
|
09143075 |
Aug 28, 1998 |
|
|
|
6393460 |
|
|
|
|
Current U.S.
Class: |
709/204 ;
709/246 |
Current CPC
Class: |
H04L 12/1831 20130101;
H04M 2201/40 20130101; H04M 2203/4536 20130101; H04M 3/563
20130101; G10L 15/1815 20130101; G06Q 10/107 20130101; H04M 3/56
20130101; H04L 12/1827 20130101; H04M 3/42221 20130101 |
Class at
Publication: |
709/204 ;
709/246 |
International
Class: |
G06F 015/16 |
Claims
What is claimed is:
1. A method for informing a user of topics of discussion in a
recorded chat between two or more people, the method comprising:
identifying elements from the chat having similar content; labeling
some or all of the identified elements as topics; and presenting
the topics to the user.
2. The method of claim 1, wherein the step of identifying elements
from the chat having similar content comprises decomposing the chat
into a plurality of utterances made by the people involved in the
chat and clustering the utterances to identify elements in the
utterances having similar content.
3. The method of claim 2, wherein the step of identifying elements
from the chat having similar content comprises parsing each
decomposed utterances into one or more tokens and representing each
utterance as a vector comprising a combination of some or all of
the one or more tokens.
4. The method of claim 3, wherein the step of representing each
utterance as a vector comprises removing some of the tokens in the
utterance before representing the utterance as a vector.
5. The method of claim 4, wherein the chat is not ongoing, and
wherein the step of removing some tokens comprises removing tokens
appearing in a percentage of all utterances in the chat which is
below a first percentage or above a second percentage.
6. The method of claim 3, wherein the chat is ongoing, and wherein
the step of representing each utterance as a vector comprises
representing all tokens in the utterance in the vector.
7. The method of claim 3, wherein the step of representing each
utterance as a vector comprises weighting each token in the
vector.
8. The method of claim 7, wherein the step of weighting each token
comprises computing the weight of a each token as the frequency of
occurrence of the token in the utterance divided by the largest
frequency of occurrence for any token in the utterance.
9. The method of claim 7, wherein the step of weighting each token
comprises computing the weight of each token as the frequency.
10. The method of claim 7, comprising normalizing each vector.
11. The method of claim 3, comprising generating a vector space
model comprising a matrix having a plurality of rows and a
plurality of columns, wherein the number of rows equals the number
of utterances represented by vectors and the number of columns
equals the number of tokens contained in the vectors.
12. The method of claim 1, wherein the step of presenting the
topics comprises hyperlinking the topics to documents containing
utterances having the respective identified elements.
13. The method of claim 12, wherein the step of presenting the
topics comprises hyperlinking each utterance in the documents to a
location in the chat in which the respective utterance appears.
14. The method of claim 1, wherein the step of labeling comprises
selecting some of the topics according to a predefined
criteria.
15. The method of claim 14, wherein the step of selecting some of
the topics comprises identifying topics which are nouns or noun
phrases and selecting the topics so identified.
16. The method of claim 1, wherein the chat is ongoing, and wherein
the step of identifying elements from the chat having similar
content comprises: receiving a first set of ongoing chat data from
the ongoing chat; decomposing the first set of ongoing chat data
into a plurality of first utterances; when a first number of first
utterances has been received, clustering the first utterances to
generate a plurality of first clusters; receiving a second set of
ongoing chat data from the ongoing chat after the first set of
ongoing chat data; decomposing the second set of ongoing chat data
into a plurality of second utterances; and when a second number of
second utterances has been received, clustering the second
utterances into the plurality of first clusters.
17. A method for clustering an ongoing chat, the method comprising:
receiving ongoing chat data; decomposing a first set of ongoing
chat data into a plurality of first utterances; when a first number
of first utterances has been received, clustering the first
utterances to generate a plurality of first clusters; decomposing a
second set of ongoing chat data received after the first set of
ongoing chat data into a plurality of second utterances; and when a
second number of second utterances has been received, clustering
the second utterances into the plurality of first clusters.
18. The method of claim 17, comprising parsing each of the first
and second utterances into tokens.
19. The method of claim 18, comprising representing each utterance
as a vector comprising a combination of the tokens in the
utterance.
20. The method of claim 19, wherein the step of representing each
utterance as a vector comprises weighting each token in the
vector.
21. The method of claim 20, wherein the step of weighting each
token comprises computing the weight of each token as the frequency
of occurrence of the token in the utterance divided by the largest
frequency of occurrence for any token in the utterance.
22. The method of claim 17, comprising generating a new cluster
after clustering the second utterances into the first clusters.
23. The method of claim 22, wherein the step of generating the new
cluster comprises identifying a given cluster as larger than all
other clusters and selectively breaking the largest cluster into
two or more smaller clusters.
24. The method of claim 23, identifying a cluster having a centroid
which is further from the largest cluster than all other clusters,
and wherein the step of breaking the largest cluster into two or
more clusters is performed if the largest cluster contains a number
of utterances greater than a predefined number and the distance of
the largest cluster from the centroid of the furthest cluster
exceeds a predefined distance.
25. The method of claim 23, wherein the step of breaking the
largest cluster comprising breaking the largest cluster using a
k-means clustering technique.
26. A method for identifying elements from a chat having similar
content, the method comprising decomposing the chat into a
plurality of utterances made by the people involved in the chat
parsing each decomposed utterance into one or more tokens;
representing each utterance as a vector comprising a combination of
some or all of the one or more tokens; and clustering the
utterances using the vectors to identify elements in the utterances
having similar content.
27. An article of manufacture comprising a computer readable medium
containing a program which when executed on a computer causes the
computer to perform a method for informing a user of topics of
discussion in a recorded chat between two or more people, the
method comprising: identifying elements from the chat having
similar content; labeling some or all of the identified elements as
topics; and presenting the topics to the user.
Description
COPYRIGHT NOTICE
[0001] A portion of the disclosure of this patent document contains
material which is subject to copyright protection. The copyright
owner has no objection to the facsimile reproduction by anyone of
the patent document or the patent disclosure, as it appears in the
Patent and Trademark Office patent files or records, but otherwise
reserves all copyright rights whatsoever.
BACKGROUND OF THE INVENTION
[0002] The invention disclosed herein relates to information
retrieval systems, and particularly, to systems and methods for
providing users with helpful information about the contents of
chats including ongoing on-line chats.
[0003] Real-time textual conversations, commonly known as chats,
have become increasingly popular among both personal and business
computer users. Chats occur as conversations between two people, as
conferences among larger groups, and in persistent chat rooms or
spaces accessible to a larger community who can drop in, read what
was recently written, and contribute if they desire. Chats are
widely available over local and wide area networks, and are
particularly popular among users of on-line services and the
Internet.
[0004] The textual nature of chat makes it particularly valuable in
some settings. Chat can be conducted while people are on the phone,
allowing it to be used as a second channel for exchanging
information. Because of the persistent nature of text, a user can
catch up on anything that was said in a chat if they were
momentarily distracted or interrupted. Chat can be an inexpensive
and lightweight way for people to exchange information in real
time. These and other reasons contribute to the growing use of chat
in business settings and the increasing incorporation of chat into
the offerings of major software manufacturers.
[0005] Chats frequently contain important information that users
will want to access at a later time. This can include specific
details, such as a phone number or address, lists of tasks the user
must remember to perform, and broader discussions and ideas. While
mechanisms have been designed to allow users to save the transcript
of a chat session for later retrieval, these identify the saved
transcripts only by such details as the date, time and/or
participants in the chat or require the user to manually assign a
single name to the conversation. They do not provide an automatic
and convenient way for transcripts to be identified by the topics
they cover.
[0006] Because of the conversational and often informal nature of
chat, a single conversation can concern a number of topics,
intertwined temporally and frequently shifting from one topic to
another. A person presented with a chat transcript, both when
retrieving a past transcript and joining a conversation in
progress, must scan through the entire transcript to know what was
discussed or to find a topic of interest.
[0007] In addition, while some existing systems give others
awareness that people are involved in a currently occurring
conversation in which they could participate, they do not inform
them of the specific topics being discussed. The user must access
the chat transcript and read through it to determine if an issue of
interest is being discussed.
[0008] There is therefore a need for systems and methods for
allowing users to quickly determine the contents of a chat and to
monitor the progress of ongoing chats and the topics being
discussed therein.
SUMMARY OF THE INVENTION
[0009] It is an object of the present invention to solve the
problems described above with existing chat systems.
[0010] It is another object of the present invention to
automatically label a chat transcript by the topics it
includes.
[0011] It is another object of the present invention to allow users
to easily locate the portions of a chat transcript dealing with a
specific topic.
[0012] It is another object of the present invention to allow users
to easily discern the topic under current discussion in an ongoing
chat without monitoring the complete text.
[0013] It is another object of the present invention to
automatically notify potential chat participants when topics of
interest to them are under discussion.
[0014] It is another object of the present invention to allow users
to easily determine the topics discussed in a chat transcript.
[0015] It is another object of the present invention to
automatically categorize and topically index the contents of a chat
session through statistical analysis of its contents.
[0016] The above and other objects are achieved by a method for
informing a user of topics of discussion in a recorded chat between
two or more people. The method includes the steps of identifying
elements from the chat having similar content, labeling some or all
of the identified elements as topics, and presenting the topics to
the user. In some embodiments, identifying elements from the chat
having similar content includes the steps of decomposing the chat
into a plurality of utterances made by the people involved in the
chat and clustering the utterances to identify elements in the
utterances having similar content.
[0017] Furthermore, each decomposed utterance is parsed into one or
more tokens and represented as a vector comprising a combination of
some or all of the one or more tokens. In the case of a previously
recorded chat which is no longer ongoing, some of the tokens in the
utterance may be removed before representing the utterance as a
vector. The tokens removed include tokens appearing in a percentage
of all utterances in the chat which is below a low percentage or
above a high percentage. In the case of an ongoing chat, in which
such percentages cannot be determined because the full chat record
is not yet available, all tokens in the utterance are included in
the vector. The tokens in each vector are weighted by frequency of
their occurrence in the utterance or chat as a whole, and a
vector-space model is generated from all the vectors.
[0018] Standard clustering techniques are used to cluster the
utterances based on the vector space model created from the vectors
and tokens. In the case of a previously recorded chat, clustering
is performed on each utterance. In the case of ongoing chats,
clustering is performed in accordance with a process which accounts
for the dynamically changing nature of the chat content. The
process involves receiving a first set of ongoing chat data from
the ongoing chat, decomposing the first set of ongoing chat data
into a plurality of first utterances, and, when a first number of
first utterances has been received, clustering the first utterances
to generate a plurality of first clusters. As the chat continues, a
second set of ongoing chat data is received and decomposed into a
plurality of second utterances, which utterances are clustered into
the first clusters when a second number of second utterances has
been received. A new cluster is performed under certain conditions
by breaking the largest of the existing cluster into two or more
smaller clusters. The result is an ever changing collection of
topics representing the subject matter under discussion in the
chat.
[0019] Salient phrases and keywords are automatically extracted
from the topics for use in labeling the chat transcript and
creating a dynamic listing of the topics it contains. The listing
serves as an active table-of-contents, allowing users to easily
access the portions of the chat transcript it references. A variety
of textual and graphical displays may be used to provide an
overview at a glance of the locations in a chat transcript in which
each of its topics was discussed. The keywords identifying the
topic under current discussion in an ongoing chat can be displayed,
allowing users to decide at a glance if they wish to participate. A
number of chat conversations may be automatically monitored, and
users notified when a topic in which they have previously, through
deliberate setting or observed actions, indicated an interest.
BRIEF DESCRIPTION OF THE DRAWINGS
[0020] The invention is illustrated in the figures of the
accompanying drawings which are meant to be exemplary and not
limiting, in which like references refer to like or corresponding
parts, and in which:
[0021] FIG. 1 is a block diagram showing a computer system for
processing and clustering a chat in accordance with one embodiment
of the present invention;
[0022] FIG. 2 is a flow diagram showing one process performed by
the system of FIG. 1 for decomposing and clustering a chat in
accordance with the present invention;
[0023] FIG. 3 is a flow chart showing the process of indexing a
previously recorded chat in accordance with one embodiment of the
invention; and
[0024] FIG. 4 is a flow chart showing the process of indexing an
ongoing chat in accordance with one embodiment of the
invention.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0025] In accordance with the invention, automatically clustering
the utterances of a chat identifies elements of the chat with
similar content. Extracting statistically significant labels from
the utterances identifies the topics associated with the clusters.
These labels then act as a summary of topics discussed, a
description of active topics at a specific time, such as the
current topic, or a table of contents into the chat for later
searching. Two general clustering processes are described herein,
one for off-line processing and one for on-line processing.
[0026] A system and method of preferred embodiments of the present
invention are now described with reference to FIGS. 1-4B. Referring
to FIG. 1, a system 10 of one embodiment of the present invention
includes a computer system 12, which may be a personal computer,
networked computers, or other conventional computer architecture,
has a processor 14, volatile and nonvolatile memory devices 16, a
keyboard 18, and a display 20. The system 10 may include additional
or alternative input devices such as a speech recognition system
22, including a microphone and speech recognition software such as
the ViaVoice.TM. software available from IBM Corp., the assignee of
the present application, to translate speech into digital data, as
well as other convention input devices 24 such as a mouse,
electronic stylus, etc. The system 10 may further include
additional output devices such as a voice synthesis system 26,
printer, or other conventional device.
[0027] In accordance with the present invention, a number of
software programs or program modules or routines reside and operate
on the computer system 12. These include a chat application 28, a
chat preprocessor 30, a clustering program 32, a labeler 34, and a
table of contents or index generator 36. The chat application 28
may be any conventional chat application, such as VP-Buddy from
Ubique and Instant Messenger from AOL. The chat application 28
receives input from a local user of the computer system 12 through
one or more of the input devices 18, 22, 24 and receives input from
one or more remote users trough a local or wide area network 38,
including over the Internet, an on-line service, a telephone
network, or other telecommunication connection.
[0028] The system 10 collects chat transcripts, decomposes chats
into individual utterances, cluster treats each utterance as a
separate document, and extracts labels from each cluster. The
preprocessor 30 collects the chat transcripts by using existing
chat logs kept by the particular chat application 28. The
preprocessor 30 decomposes the chat into individual utterances in
several possible ways depending on the application. In one
embodiment, ad hoc parsing techniques specific to the transcript
file format of the chat application 28 to identify each utterance
and write it to a separate file. In a real time chat environment,
the utterances may be individually logged or sent to the input
queue of the clustering engine 32 as each utterance is sent.
[0029] The clustering engine 32, which may be any on-line
clustering algorithm including conventional ones such as the
k-means clustering algorithm described in L. Bottou and Y. Bengio,
Convergence Properties of the K-Means Algorithm, in Advances in
Neural Information Processing Systems 7, pages 585-592 (MIT Press
1995), which is hereby incorporated by reference into this
application. Several examples of additional document clustering
algorithms are described in the following two documents, which are
hereby incorporated by reference into this application. Douglas R.
Cutting, David R. Karger, Jan O. Pedersen, John W. Tukey,
Scatter/Gather: A Cluster-based Approach to Browsing Large Document
Collections. In Proceedings of the 15th Annual International ACM
SIGIR Conference. Association for Computing Machinery. New York.
June, 1992. Pages 318-329. Gerard Salton. Introduction to Modern
Information Retrieval, (McGraw-Hill, New York 1983).
[0030] The clustering engine 32 treats each utterance as a separate
document, and converts each document or utterance to a feature
vector. Features are the words used in the utterance, key phrases,
and other attributes such as time, date, and author. In particular
embodiments, the natural language parsing capabilities of the
Textract.TM. information retrieval program available from IBM Corp.
are used. Textract's ability to locate proper names is described in
the following two articles, which are hereby incorporated by
reference into this application: Yael Ravin and Nina Wacholder,
Extracting Names from Natural-Language Text, IBM Research report RC
20338, T. J. Watson Research Center, IBM Research Division,
Yorktown Heights, N.Y., April 1997; and Nina Wacholder, Yael Ravin,
and Misook Choi, Disambiguation of Proper Names in Text,
Proceedings of the Fifth Conference on Applied Natural Language
Processing, pages 202-208, Washington D.C., March 1997. Textract
may be used only to identify key noun phrases.
[0031] The feature vector for an utterance has a non-zero weight
for every feature present in the utterance. The weight is based on
the frequency of the feature in the document, its type (e.g.,
whether an author field, word, or phrase), and its distribution
over the collection. Once an utterance is represented as a feature
vector, a similarity measure is defined on utterances. The
similarity measure is then used to group related utterances.
[0032] The labeling engine 34 selects the most statistically
significant features to label as clusters. Noun phrases are
advantageously selected as labels because they are typically more
meaningful to users. The table of contents engine 36 organizes the
labels received from the labeling engine 34
[0033] Particular methods for processing and clustering off-line,
completed chat sessions and on-line, ongoing chat sessions are now
described with reference to the flow diagram of FIG. 2 and the flow
charts in FIGS. 3A-3B and 4A-4B. For off-line chats, a chat
transcript file 40 (FIG. 2) is retrieved, step 60 (FIG. 3A), and
the chat transcript broken into separate utterances 42, step 62. An
utterance 42 is a particular entry by a participant in the chat
session, or may be any other convenient logical block or portion of
the chat. The utterances 42 are then parsed into individual tokens
44, step 64, the tokens containing individual words or word
phrases.
[0034] In off-line chat preprocessing, some tokens in each
utterance may be removed from consideration because they are less
relevant or meaningful to users. Tokens that appear in relatively
very few utterances likely do not represent a truly relevant aspect
of the discussion, and tokens that appear in a large percentage of
utterances are likely commonplace words such as articles. Thus, the
preprocessor computes the percentage of utterances in which each
token appears, step 66. Then, each utterance is considered, step
68, and each token in the utterance is considered, step 70. For the
given token, if the percentage associated with that token is either
less than a predefined lower limit percentage L, step 72, or higher
than a predefined upper limit percentage H, step 74, the token is
removed from the utterance, step 75. Alternatively, all tokens may
be retained, and utterances in both off-line and on-line clustering
processes may be subjected to a stop list, which filters the
utterances to remove certain words known to have little value in
information retrieval, such as a, an, but, the, or, etc.
[0035] For each remaining token, a token frequency tf is computed,
step 76, as the frequency of the given token in that utterance, and
compared to tf.sub.max, which is the largest token frequency of any
term in the utterance, initially set to 0 for each utterance. If tf
for a given token exceeds the current value of tf.sub.max for that
utterance, then tf.sub.max is set equal to tf, step 80. Once all
tokens in the utterance have been considered, the current value of
tf.sub.max will represent the maximum token frequency for the
utterance.
[0036] When all tokens in each utterance have been considered, step
82, and all utterances in the chat transcript considered, step 84
(FIG. 3B), each utterance is represented as a vector in a
vector-space model. Thus, each utterance is considered, step 86,
and each token in a given utterance considered, step 88. Each token
is given a weight in each utterance according to the formula
tf/tf.sub.max, step 90. Other possible formulas include a binary
value (1 if the term occurs in the document, 0 if it does not), and
a traditional tfidf measure where the frequency of the term in the
utterance is divided by the number of documents in the collection
that contain the term. In the case of on-line clustering, discussed
further below, the inverse document frequency would simply use the
number of documents up to the present time that contain the
term.
[0037] If all tokens have been assigned weights step 92, a vector
is generated as the combination of the weighted tokens, step 94.
Each vector is then normalized to a unit vector, i.e., a vector of
length 1, step 96. This is accomplished, in accordance with
standard linear algebra techniques, by dividing each token's weight
by the square root of the sum of the squares of the token weights
of all tokens in the vector.
[0038] When all utterances have been considered and converted into
vectors, step 98, the vectors are converted to a vector space
model, step 100, which is a matrix where the number of rows is
equal to the number of utterances and the number of columns is
equal to the number of tokens retained to form the vector-space
representation. This will be referred to as document-token matrix.
In retrospective or off-line clustering of chats, the number of
vectors to be clustered is equal to the number of utterances. The
matrix resulting from the preprocessing is sparse, i.e., very few
of the cells in the document-token matrix are non-zeros.
[0039] The vectors or utterances are then clustered separately,
step 102. This clustering can be performed in several conventional
ways known to those of skill in the art, including in ways
described in the Salton and Cutting references referred to above.
The clustering results in a set of clusters 46 (FIG. 2) which may
then be grouped into groups of clusters 48 based on similar
content. This process of hierarchical clustering is accomplished by
computing a centroid document, which is often a vector where each
token weight is the average of the token weights for that token for
all vectors in the cluster. Each centroid is treated as a document,
and each cluster is represented as a centroid. The process of
clustering is performed again on the centroid representing
clusters, generating a new, cluster containing one or more old
clusters. This process of hierarchical clustering may be performed
a desired number of times or until a predefined criteria is
reached.
[0040] The clusters are then assigned label 50 by selecting some of
the tokens in the cluster 46 or cluster group 48, step 104. The
labeling of document clusters is known to those of skill in the
art, and is described for example in pages 314-323 of Peter G.
Anick and Shivakumar Vaithyanathan, Exploiting Clustering and
Phrases for Context-based Information Retrieval, in Proceedings of
the 20th International ACM SIGIR Conference, Association for
Computing Machinery, July 1997, which document is hereby
incorporated by reference into this application. Problems may arise
in chat clustering since the utterances are usually not very well
formed and there are potentially very large number of spelling
mistakes. The process of labeling chat clustering is restricted to
picking semantically meaningful and important words and phrases in
each cluster, wherein words are considered important when they
satisfy predefined statistical criteria similar to the generation
of token weights.
[0041] Once labels have been assigned, a table of contents or index
52 is generated, step 106, by, in one embodiment, arranging the
labels in an order generally reflective of the order of the tokens
in the utterances or in any other desired order. The table of
contents 40 and chat transcript may be shown together on the
display 20, as shown in FIG. 1, such as by the use of framing
techniques known to those of skill in the art. An additional frame
54 may be provided for displaying a cluster associated with a given
label selected from the index 52. In some embodiment, the labels
may be hyperlinked to documents containing the cluster group
information, such as through the use of HTML anchors. The cluster
group information may contain a list of the utterances in the
group, each utterance being hyperlinked to the same utterance in
the chat transcript 40. As a result, a user may quickly and easily
navigate from the index 52, to the group were utterances are shown,
and ultimately to the position within the chat transcript where the
utterance occurs.
[0042] On-line or ongoing chat clustering (OCC) is a slightly more
difficult problem than of-line clustering. One way of looking at
the OCC problem is that of a constantly evolving problem that the
system attempts to track using a clustering process. The purpose of
the on-line clustering is two-fold:
[0043] 1. Classify an existing utterance into one of existing
clusters/classes.
[0044] 2. Ascertain whether a new utterance should start a new
cluster.
[0045] For on-line clustering, the system has to perform the same
general three steps as retrospective clustering, that is,
preprocessing, clustering and labeling. However, the individual
steps involved differ.
[0046] Referring to FIGS. 4A-4B, preprocessing of the chat sessions
involves the steps of retrieving ongoing chat data, step 120,
either continuously utterance by utterance or in sets, breaking the
chat data retrieved into separate utterances, step 122, and parsing
the utterances into individual tokens, step 124.
[0047] Chat data is retrieved until the number of utterances
reaches a predefined threshold T, step 126, which is sufficient to
create a meaningful data set. When the threshold is reached, each
utterance retrieved is considered, step 12, and each token in each
utterance considered, step 130. In on-line chat clustering, all
tokens are retained in the utterances as they are needed to
represent the vector-space model. As with off-line clustering, each
token is given a weight tf/tf.sub.max by computing tf, step 132,
comparing tf for the given token to the running variable
tf.sub.max, and updating tf.sub.max until the maximum frequency is
found for the utterance. Using inverse document frequencies in the
on-line case is problematic because weights of the same term at
different times are not really comparable because they will have
different document frequencies. On the other hand, updating the
weight of a term in all past documents it occurs in every time it
appears in a new document, that is, re-evaluating its weight, and
reclustering is currently computationally prohibitive. It may also
lead to unstable clustering of prefixes of the chat.
[0048] When all tokens in each utterance, step 138, and all
utterances in the current chat data set, step 140, have been
processed, each utterance is represented as a vector in the
vector-space model, step 142. Each vector is normalized to a unit
vector, as in off-line clustering.
[0049] If clusters do not already exist, step 144, then this set of
utterances represents the first set retrieved from the ongoing
chat. In that case, the utterances are clustered using any
conventional clustering algorithm as with off-line clustering, step
146, tokens from the clusters are selected as labels, step 148, and
a table of contents is generated from the labels, step 150, and
presented to the user.
[0050] The system continues retrieving additional data from the
chat, step 120, and processing the data in the same fashion. When
the number of new utterances is greater than or equal to T, step
126, vectors are created for the new utterances as described above.
Since clusters already exist, step 144, the new vectors are
clustered into the existing clusters, step 152. In one embodiment,
on-line clustering of utterances is performed by computing the
dot-product of the new utterance with each of the centroids of the
existing clusters. A new cluster can be formed in several ways,
including, for example, by forming a new cluster using the on-line
clustering algorithm if the nearest neighbor of all the cluster
centroids is less than a pre-determined threshold t in
distance.
[0051] The system then determines whether to break up any clusters.
First, the system finds the cluster with the greatest number of
utterances, step 154. If that cluster has more utterances than a
predefined threshold N, step 156, then the system finds the
utterance with the furthest distance from the centroid of the
largest cluster, step 158. If this distance exceeds a predefined
distance D, the cluster is broken into two or more smaller clusters
using the k-means clustering algorithm, step 162. The labels are
then revised to reflect the change in clusters, step 164, and the
table of contents or index also revised accordingly, step 166. The
system continues to retrieve new chat data sets and proceeds in an
iterative fashion as described.
[0052] As an alternative, the clustering process may involve use of
the Minimum Description Length (MDL) approach, described in U.S.
Pat. No. 5,787,274, which is hereby incorporated by reference. In
this embodiment, the new utterance is assigned to one of the
existing clusters, by ignoring the pre-determined threshold, based
on a nearest neighbor evaluation using the on-line k-means
clustering algorithm. After a pre-determined number of new
utterances have been collected, the overall likelihood of the data
is evaluated, conditioned on the existing partition. The cluster
with the lowest likelihood is selected and split into two clusters,
using a batch version of the k-means clustering algorithm. The
likelihood of the new partition is then computed, noting that in
this case the number of clusters is one more than in the previous
case. The two computed likelihoods are compared after adding an MDL
penalty, the MDL criterion acting as a regularizer. The number of
clusters to retain is selected based on this computation.
[0053] Labeling is performed identically to labeling in off-line,
retrospective clustering.
[0054] As a result, these various embodiments of the present
invention provide users with a powerful, effective and easy to use
tool to quickly determine the contents of a chat transcript and to
monitor an ongoing chat without having to In addition, the system
may be programmed to accept and store specific words or phrases
which the user desires to monitor, and to inform the user when
these words or phrases are found by virtue of the chat clustering
process described herein.
[0055] While the invention has been described and illustrated in
connection with preferred embodiments, many variations and
modifications as will be evident to those skilled in this art may
be made without departing from the spirit and scope of the
invention, and the invention is thus not to be limited to the
precise details of methodology or construction set forth above as
such variations and modification are intended to be included within
the scope of the invention.
[0056] Appendices
[0057] The following pages in and forming part of this detailed
description contain three appendices with exemplary source and
result data generated in accordance with the present invention.
Appendix A contains a portion of sample chat transcript. Appendix B
contains an exemplary index or list of topics generating from the
entire chat transcript from which Appendix A contains a portion.
The number of utterances associated with each topic is listed in
parentheses following the topic. Appendix C contains several sample
clusters and associated utterances associated with and hyperlinked
to their respective topics. The utterances are hyperlinked to
utterances within the chat transcript for ease of navigation, as
described above. The weight of each token is listed in parentheses
after the token in the cluster.
* * * * *