Information Exploration Systems and Methods
CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] The present application relates to U.S. Patent 6,847,966, "Method and System for
Optimally Searching A Document Database Using a Representative Semantic Space", filed
April 24, 2002, by inventors Matthew Sommer and Kevin Thompson, and hereby
incorporated herein by reference.
BACKGROUND
[0002] Information commonly comes in large quantities. In many cases, such as published
reference works, cataloged library systems, or well-designed databases, the information is organized and indexed for easy information retrieval. In many other cases, such as litigation
discovery, personal document collections, electronic records on local networks, and internet search results (to name just a few), the information is poorly organized and not indexed at all,
making it difficult to locate and retrieve desired information.
[0003] In the past, information providers have scanned documents or otherwise obtained
documents in electronic form and have applied automated searching techniques to aid users in their quest for information. Generally, such information providers employ term searching with Boolean operations (AND, OR, and NOT). Though computationally efficient, this
automated searching technique suffers from a number of drawbacks. The primary drawback is the sensitivity of the search results to the choice of search terms. In a body of documents, the
sought-after information may be hidden by its use of synonyms, misspellings, and different
word forms (e.g., ice, iced, ices, icing, deice, re-ice, ...). A second major drawback is this
search technique's failure to discern differences in term usage, and consequently this search
£fedMique""KUιhis "a" la'rg'e' "percentage of irrelevant results (e.g., "icing" refers to frost
formation, a sugared cake topping, and a hockey penalty).
[0004] These drawbacks can usually be overcome by a person having great familiarity with
the information being sought, e.g., by structuring a query using terms commonly used in the
sought-after document's subject area. Unfortunately, such familiarity is commonly not
possessed by the searcher. Accordingly, information providers seek alternative searching
techniques to offer their users. A searching technique would greatly benefit such information providers if it enabled users to find their desired information without necessitating some
preexisting degree of familiarity with the sought after information or the searching tool itself.
BRIEF DESCRIPTION OF THE DRAWINGS
[0005] A better understanding of the various disclosed embodiments can be obtained when
the following detailed description is considered in conjunction with the following drawings, in which:
Fig. 1 shows an illustrative information exploration system embodied as a desktop computer;
Fig. 2 shows a block diagram of an illustrative information exploration system;
Fig. 3 shows an illustrative set of documents;
Fig. 4 shows an illustrative set of documents represented as vectors in a three dimensional concept space;
Fig. 5 shows a flow diagram of an illustrative information exploration method;
Fig. 6A shows an illustrative hierarchy of clusters in dendrogram form;
Fig. 6B shows an illustrative clustering tree with a branching factor of three;
Fig. 7A shows an illustrative suffix tree;
Fig. 7B shows an illustrative phrase index derived from the suffix tree of Fig. 7 A;
ττg."δ 'snows'arππustrative phrase-to-leaf index;
Fig. 9 shows a flow diagram of an illustrative cluster-naming method; and
Figs. 1OA and 1OB show an illustrative information exploration interface
[0006] While the invention is susceptible to various modifications and alternative forms,
specific embodiments thereof are shown by way of example in the drawings and will herein
be described in detail. It should be understood, however, that the drawings and detailed
description thereto are not intended to limit the invention to the particular form disclosed, but
on the contrary, the intention is to cover all modifications, equivalents and alternatives falling within the spirit and scope of the present invention as defined by the appended claims.
.TERMINOLOGY
[0007] In the following discussion and in the claims, the terms "including" and "comprising"
are used in an open-ended fashion, and thus should be interpreted to mean "including, but not limited to...".
[0008] Also, the term "couple" or "couples" is intended to mean either an indirect or direct
electrical connection. Thus, if a first device couples to a second device, that connection may be through a direct electrical connection, or through an indirect electrical connection via other devices and connections.
[0009] The term "set" is intended to mean one or more items forming a conceptual group. The term "phrase" is intended to mean a sequence of one or more words. The term
"document" refers to a set of phrases. The term "cluster" is intended to mean a set of
documents grouped on the basis of some similarity measure.
DETAILED DESCRIPTION [0010] Information exploration methods and systems are disclosed herein. In some
embodiments, these methods and systems take a set of documents and determine a hierarchy
oi clusters" representing vanous document subsets. The clusters are labeled with phrases that
identify themes common to the documents associated with the cluster. The cluster set's
hierarchical nature enables a user to explore document set information at many levels. For
example, a root cluster is labeled with phrases representing themes or characteristics shared
by the document set as a whole, and subordinate clusters are labeled with phrases
representing themes or characteristics that unify significant document subsets. By exploring
the document set's themes and characteristics in a progression from general to specific, a user becomes familiarized with the document set in a fashion that efficiently guides the user to
sought-after information. In other embodiments and variations, themes and characteristics representative of a selected document set are determined for a user in a dynamic, set-by-set
fashion as the user identifies a document set of possible interest. These dynamic embodiments
enable a user to quickly discern whether a selected document set is worthy of further analysis or not.
[0011] Figure 1 shows an illustrative system 100 for information exploration. System 100 is
shown as a desktop computer 100, although any electronic device having some amount of computing power coupled to a user interface may be configured to carry out the methods disclosed herein. Among other things, servers, portable computers, personal digital assistants (PDAs) and mobile phones may be configured to carry out aspects of the disclosed methods.
[0012] As shown, illustrative system 100 comprises a chassis 102, a display 104, and an input
device 106. The chassis 102 comprises a processor, memory, and information storage devices.
One or more of the information storage devices may store programs and data on removable
storage media such as a floppy disk 108 or an optical disc 110. The chassis 102 may further
comprise a network interface that allows the system 100 to receive information via a wired or
wireless network, represented in Figure 1 by a phone jack 112. The information storage media
ana mrorπiaxion transport "media (i.e., the networks) are collectively called "information
carrier media."
[0013] The chassis 102 is coupled to the display 104 and the input device 106 to interact with
a user. The display 104 and the input device 106 may together operate as a user interface. The
display 104 is shown as a video monitor, but may take many alternative forms such as a
printer, a speaker, or other means for communicating information to a user. The input device
106 is shown as a keyboard, but may similarly take many alternative forms such as a button, a
mouse, a keypad, a dial, a motion sensor, a camera, a microphone or other means for receiving information from a user. Both the display 104 and the input device 106 may be integrated into the chassis 102.
[0014] Figure 2 shows a simplified functional block diagram of system 100. The chassis 102 may comprise a display interface 202, a peripheral interface 204, a processor 206, a modem or
other suitable network interface 208, a memory 210, an information storage device 212, and a bus 214. System 100 may be a bus-based computer, with the bus 214 interconnecting the
other elements and carrying communications between them. The display interface 202 may take the form of a video card or other suitable display interface that accepts information from the bus 214 and transforms it into a form suitable for the display 104. Conversely, the peripheral interface 204 may accept signals from the keyboard 106 and other input devices such as a pointing device 216, and transform them into a form suitable for communication on the bus 214.
[0015] The processor 206 gathers information from other system elements, including input
data from the peripheral interface 204, and program instructions and other data from the
memory 210, the information storage device 212, or from a remote location via the network interface 208. The processor 206 carries out the program instructions and processes the data
aCcόf dffigfyTf he program instructions may further configure the processor 206 to send data to
other system elements, comprising information for the user which may be communicated via
the display interface 202 and the display 104.
[0016] The network interface 208 enables the processor 206 to communicate with remote
systems via a network. The memory 210 may serve as a low-latency temporary store of
information for the processor 206, and the information storage device 212 may serve as a long
term (but higher latency) store of information.
[0017] The processor 206, and hence the computer 100 as a whole, operates in accordance
with one or more programs stored on the information storage device 212. The processor 206
may copy portions of the programs into the memory 210 for faster access, and may switch between programs or carry out additional programs in response to user actuation of the input
device. The additional programs may be retrieved from information the storage device 212 or may be retrieved from remote locations via the network interface 208. One or more of these
programs configures system 100 to carry out at least one of the information exploration
methods disclosed herein.
[0018] Figure 3 shows an illustrative set of documents 302-306. The document set can originate from anywhere. Examples include internet search results, snippets from internet search results, electronic database records, and text document files. Each document includes, one or more "sentences", i.e., one or more groups of words separated by punctuation or some
other form of semantic separators. It is contemplated that the document set as a whole may be
large, i.e., more than 1,000 documents, and potentially may include millions of documents.
With such document sets, the disclosed information exploration methods are expected to greatly reduce the time required for a user to become familiar with the general contents of the
αocumenrseτ, as wen as me lime required for a user to locate a particular document of
interest or documents containing information on a particular topic of interest.
[0019] In various embodiments, the disclosed information exploration methods employ
clustering techniques to group similar documents together. Many such clustering techniques
exist and may be employed. Examples of suitable clustering techniques include suffix tree
clustering (see, e.g., O. Zamir and O. Etzioni, "Web Document Clustering: A Feasibility
Demonstration", SIGIR'98 Proceedings pp 46-54), agglomerative hierarchical clustering, divisive hierarchical clustering, K-means clustering, bisecting K-means clustering (see, e.g.,
M. Steinbach, et al., "A Comparison of Document Clustering Techniques", Technical Report #00-034, Univ. of Minnesota), Buckshot, Fractionation (see, e.g., D.R. Cutting, et al.,
"Scatter/Gather: a cluster-based approach to browsing large document collections", SIGIR'92 Proceedings pp 318-329). Each of the foregoing references is hereby incorporated by reference.
[0020] Each of the foregoing clustering technique examples (except perhaps suffix tree clustering) can employ a variety of similarity measures to evaluate the "distance" or
dissimilarity between any two documents. The similarity measures may be based on term frequency vectors (see, e.g., M. Steinbach, et al., cited previously), concept-space vectors, and other representations. However, clustering quality is an important consideration for the success of the disclosed information exploration methods. Clustering quality has been found
to be significantly better when latent semantic analysis (LSA) principles are applied to obtain
concept-space vectors. LSA-based clustering methods establish a relationship between
documents in a document set and points in a high-dimensional concept space.
[0021] As described in U.S. Patent No. 6,847,966, "Method and System for Optimally
Searching a Document Database Using a Representative Semantic Space", document-to-
''concept" sp'are" relationship can be established by applying singular value decomposition to a
terms-to-document matrix. Briefly summarized, a terms-to-docurnents matrix is created
having a row for each term and a column for each document. (Pervasive terms such as "the",
"and", "in", "as", etc., may be eliminated from consideration.) Each matrix element is the
number of times that row's term can be found in that column's document. In some
embodiments, each row of the matrix is multiplied by a weighting factor to account for the discriminating power of different terms. For example, the weighting factor may be the term's inverse document frequency, i.e., the inverse of the number of documents in which that term
appears. The , term-to-document A matrix is then decomposed using singular value
decomposition into three matrices: a term-to-concept matrix T, a diagonal matrix S, and a
concept-to-document matrix Dτ:
A = T S Dτ (1)
Each column of the concept-to-document matrix Dτ provides the concept-space vector for a
corresponding document. Once the term-to-concept matrix T, and the diagonal matrix S, have
been established, the document-to-concept space relationship can be expressed: dT = S-l rrT a, . (2)
where a is a column vector of term frequencies, i.e., the elements are the number of times that
row's term appear in a given document, and dτ is a resulting column vector of concept coordinates, i.e., the concept space vector for the given document. In embodiments having
weighting factors for the elements of the term-to-document matrix A, those weighting factors
are also applied to the column vector a. The relationship given by equation (2) can be applied
to documents that were not in the set of documents used to derive the matrices S and T, although the concept-space vector may need to be normalized. The term "pseudo-document
vectctf'^'Ηe'rem ifeefl'tό "refer to those concept-space vectors calculated for documents not in
the original set of documents used to derive the S and T matrices. For further details, refer to
U.S. Patent No. 6,847,966, which is hereby incorporated by reference.
[0022] Fig. 4 shows an illustrative document set in a concept space. Three perpendicular axes
are shown (X, Y, and Z), each axis representing a different concept. A unit sphere 402 (i.e., a
sphere of radius 1) is shown centered at the origin. Also shown is a plurality of concept-space
vectors 404, which may include one or more pseudo-document vectors. Each of the vectors
404 is normalized to have unit length, meaning that each vector 404 is drawn from the origin to a point on unit sphere 402. Each of the vectors 404 is derived from a corresponding
document in the document set, with the vector's direction representing some combination of the concepts represented by the illustrated axes. Although only three dimensions are shown, it
is expected that in practice many more dimensions will be used.
[0023] Clustering techniques seek to group together documents concerning similar concepts. In concept space, documents concerning similar concepts should be represented by vectors
having similar orientations. Document similarity may thus be measured by determining the inner ("dot") product of the documents' concept space vectors. The dot product of two unit
vectors equals the cosine of the angle between the vectors, meaning that aligned ("similar") concept vectors have a similarity of 1, while oppositely-aligned concept vectors have a
similarity of -1. Other document similarity measures exist and may be employed. For examples of other similarity measures, see U.S. Patent Nos. 5,706,497, 6,633,868, 6,785,669,
and 6,941,321, which are hereby incorporated herein by reference.
[0024] Fig. 5 shows an illustrative information exploration method that may be implemented by information exploration system 100. The method comprises four phases: obtaining the
documents 502, clustering the documents 504, labeling the clusters 516, and exploring the
dotumait"s'fer'yzs."iiie'"metnod begins with block 502, in which the document set is obtained
or otherwise made accessible. The documents may be obtained and processed incrementally,
meaning that the operations represented by block 502 continue to occur even as the operations
represented by subsequent processing blocks are initiated. Moreover, various processing
operations described below may be serialized or parallelized. For example, the labeling
operations 516 may be completed before exploration operations 528 are initiated.
Alternatively, various labeling operations 516 and exploration operations 528 may occur concurrently. (When parallelized, the information exploration process may exploit available
computational resources more effectively at the cost of added programming complexity.)
[0025] The illustrated clustering operations 504 begin with block 506, which represents the information exploration system's conversion of documents into concept-space vectors. The
conversion relationship may be predetermined from a representative document set, or the
relationship may be calculated anew for each document set. The relationship may be
determined using singular value decomposition in accordance with equation (1), and applied in accordance with equation (2).
[0026] In block 508, the information exploration system 100 determines a number of clusters. This number may be determined in a variety of ways such as having a predetermined number, or a number based on a distribution of the document vectors, hi some preferred embodiments, the number is determined from the size of the document set, and chosen to predispose the
final clusters to have a predetermined target size. For example, a target cluster size may be
chosen, such as 100 documents per cluster. The size of the document set may then be divided
by the target cluster size to and rounded up or down to obtain a number of clusters.
Alternatively, the number of clusters may be chosen as a function of the number of concept
space dimensions, e.g., 2 d or 2d, where d is the number of concept space dimensions. As yet
"anoxner alternative, me target cluster size may be allowed to vary as a nonlinear function of
the document set size, so that the number of clusters is (e.g.)
where N is the document set size.
[0027] In the illustrated clustering operations 504, a bisecting K-means clustering technique
is employed. Initially, the document set is treated as a single cluster. In block 510, the current
number of clusters is compared to the desired number of clusters to determine whether the
clustering operations are complete. Other stopping criteria may be used in addition, or alternatively to, the number of clusters. For example, the clustering operations may be
considered complete if the average clustering error falls below a predetermined threshold. One possible definition for average clustering error is:
2
E _= j 1f V∑ V∑ d AjT - ZdΪiT
*i -«tf (4) keCiek
where C is the set of clusters, d/c is the average concept-space document vector for cluster k
(hereafter termed the "mean cluster vector"), and df is the zth concept-space document
vector in cluster k.
[0028] hi block 512, the information exploration system 100 selects a cluster to be divided. In some embodiments, the largest undivided cluster is selected for division. In other
embodiments, the cluster with the largest clustering error is selected. As the loop iterates, the clusters are iteratively split and split again until the stopping criterion is met.
[0029] In block 514, the selected cluster is processed to determine two sub-clusters. The sub-
clusters may be determined using a K-means algorithm. The determination involves the
random selection of two members of the original cluster as "seeds" for the sub-clusters. Since
"tile ■ffiittai""^tfb-clit'st6rs"h"ave only one document vector, the mean cluster vector for the sub-
clusters equals the corresponding document vector. Each of the remaining members of the
original cluster are in turn compared to the mean cluster vectors and grouped into the sub-
cluster with the closest mean cluster vector. The mean cluster vector is updated as each new
vector is grouped with the sub-cluster. Once each document vector has been processed, a
tentative division of the original cluster has been determined. This process may be repeated
multiple times, and the various resulting tentative divisions may be compared to determine the "best" division. The determination of a best division may be based on sub-cluster sizes,
with more equal sizes being preferred. Alternatively the determination of best division may be based on average clustering error, with the smallest error being preferred. In some
embodiments, the first tentative division is accepted if the disparity in sizes is no greater than 1:4. If the disparity is too great, the process is repeated to obtain a different tentative division
of the cluster.
[0030] Once a tentative division is accepted, the original cluster is replaced with its two sub- clusters (although the original cluster is stored away for later use). The information
exploration system repeats blocks 510-514 until the stopping criterion is met. The iterative subdividing of clusters creates a hierarchy of clusters as illustrated by Fig. 6 A.
[0031] Fig. 6A shows an illustrative hierarchy of clusters in which a document set 602 is iteratively subdivided to obtain leaf clusters 604-616 (also identified by roman numerals I-
Vπ). In the document set and the leaf clusters, letters A through P are used to represent
documents. (Note that intermediate clusters can be reconstructed by combining leaf clusters
as shown by the branches.) The cluster selection rule for this example is to select the largest
remaining cluster, and the stopping criterion was that the largest cluster would have less than four documents. After the first division, the original cluster has been divided into
1A,13;JJ;K,1,K;L,M,JNJ"M'C1 1(J,F,G,H,J,O,P}. The larger of these two clusters is then divided.
At the stage identified by line 620, three clusters exist: {A,E,K,L,M,N}, {B,D,I}, and
{C,F,G,H,J,O,P}. The third of these is now the largest cluster, so it is next to be subdivided.
[0032] The repeated two-way divisions produce a binary clustering tree (i.e., each
intermediate node has two children). However, the divisions occur in a sequential manner,
meaning that any number of children can be identified for each intermediate node. For
example, a ternary clustering tree can be constructed in the following way: the first two divisions of the original document data set produces three sub-clusters, as indicated by line
620. The first of these sub-clusters, after being divided twice, produces another three sub- clusters as indicated by line 622. The second original sub-cluster 610 is a leaf node, and has
no children. The third original sub-cluster is divided twice, producing another three sub- clusters as indicated by line 624. Replacing lines 620-624 with nodes, the ternary tree shown
in Fig. 6B results.
[0033] The desired branching factor for the cluster set hierarchy can be adjusted by moving
lines 620-624 upwards or downwards along the tree (assuming enough divisions have been made to reach the desired branching factor). Thus the original binary cluster hierarchy can be converted into a ternary-, quaternary-, or n-ary cluster hierarchy. This configurability can be exploited in the exploration phase described further below.
[0034] Returning to Fig. 5, the illustrated cluster labeling operations 516 begin in block 518
with the construction of a suffix tree for each of the leaf clusters. A suffix tree is a data
structure that is useful for the construction of a phrase index (block 520). A "true" suffix tree
is an agglomeration of paths beginning at the root node and progressing branch-by-branch
into the tree. Each branch represents one or more words. A path is defined for each sentence
in the document set and for each suffix of those sentences. (A five-word sentence has four
sϊxffix^s'rtM 1IM fdur"wδfa.s"of the sentence, the last three words of the sentence, the last two
words of the sentence, and the last word of the sentence.) Every node of the suffix tree, except
for the root and leaf nodes, has at least two children. If a node does not have two children, the
node is eliminated and the word(s) associated with the links to and from the node are joined
to form a multi-word phrase.
[0035] Before giving an example of a suffix tree, some slight simplifications will be
discussed. The branching factor for a suffix tree can be quite high, leading to a potentially
very large data structure. To improve the system performance, documents may be "cleaned" before being built into the suffix tree. The cleaning involves eliminating "stop words", i.e.,
any words in a predefined set of words. The predefined set includes pervasive words (such as "a", "an", "the"), numbers ("1", "42"), and any other words selected by the system designer as
not being helpful. The cleaning further involves "stemming", a process in which word stems
are retained, but prefixes and suffixes are dropped. Stemming makes "walk", "walker",
"walking" equivalent. Throughout the cleaning process, the position of a given term in the original document is stored for later use. As another simplification, the suffix tree depth is preferably limited to a predetermined maximum (e.g., a maximum of six branches from the root). Taken together, these simplifications preserve feasibility without a significant performance sacrifice.
[0036] Fig. 7 A shows an illustrative suffix tree determined for a document set having three documents {B,D,I}, each with a single sentence:
B: "cat ate cheese"
D: "mouse ate cheese too"
I: "cat ate mouse too"
uocumeiir'ir'nas one sentence and two suffixes, each having a corresponding path in the
suffix tree. The three paths begin at the root node 702, and end at a node labeled by a "B" in a
square. For example, beginning at node 702, taking the branch labeled "cat ate" to node 704,
followed by the branch labeled "cheese" reaches the "B" in the lower left corner. As another
example, beginning at node 702 and taking the branch labeled "ate" to node 706, followed by
the branch "cheese" yields a suffix path that reaches the second "B" four squares from the left
of the figure. Similar paths can be found for the sentences and suffixes in documents D and I
too.
[0037] The suffix tree represents a phrase index. Each node (other than the root node) having
at least two different documents (i.e., squares) in its descendants represents a phrase that is common to two or more documents. Thus node 704 represents the phrase "cat ate", which
appears in documents B and I. Node 706 represents the phrase "ate", which appears in all
three documents. Node 708 represents the phrase "ate cheese", which appears in documents B
and D. These and other shared phrases are shown in the illustrative phrase index of Fig. 7B.
[0038] Fig. 7B shows an illustrative master phrase index having all shared phrases from the document set {B,D,I} . The phrase index may be determined from a suffix tree as described above or it may be determined through other means including, e.g., a suffix array. In practice, the phrase index can get quite lengthy, and accordingly, the phrase indices that are determined
for the leaf clusters in practicing various disclosed method embodiments may include only
phrases that occur in some minimum number of documents of the corresponding leaf cluster.
In block 520 of Fig. 5, the information exploration system 100 constructs a master phrase-to- leaf cluster index by combining the phrase indices for each leaf cluster. For each phrase, the
system 100 identifies the leaf clusters containing that phrase. Although it is not strictly
necessary, the phrase index may also include an indication of one or more documents and
■■"posmons witnm me original documents where the phrase appears. Also for each phrase, the
phrase index includes a score determined by system 100. Different scoring strategies may be
employed. In some embodiments, the score is a product of the phrase's coverage (the fraction
of documents in which it appears), and a logarithm of the phrase's length. For example, the
scores given in Fig. 7B are determined in accordance with:
Score = ( /n / .V) IOg21 , (5) where m is the number of documents in which the phrase appears, N is the number of documents in the document set, and L is the number of words in the phrase.
[0039] In block 520 of Fig. 5, the master phrase index is sorted by score. Fig. S shows an
example of a sorted phrase-to-leaf index for the cluster hierarchy shown in Fig. 6A. hi block
522, the information exploration system 100 iterates through the nodes in cluster hierarchy, selecting representative phrases from the master phrase index to label each node, hi at least
some embodiments, the label selection strategy produces multi-phrase labels that include only
phrases that do not appear in sibling clusters. For example, node 622 in Fig. 6A includes only documents in leaf clusters I, π, and DI. Thus any phrase that the index indicates as appearing in leaf clusters IV, V, VI, or VII, would not be selected as a label for node 622. (Note that in
some embodiments the phrase index construction method may disregard a phrase's appearance in some leaf clusters if that phrase appears in less than some predetermined number of documents. Accordingly, this exclusionary principle may not be absolute.) The
representative phrases are used to label the clusters during exploration operations 528.
[0040] In some alternative embodiments, the information exploration system 100 may determine a node-specific score for each of the phrases in the master phrase index. This
operation is optional, but may be desirable if it is desired to determine which phrase is most
descriptive of the cluster. Though the score determined in equation (5) provides some
""lnϊlibatTbϊϊ"όT'nt)W representative the phrases are of the cluster, node-specific scoring strategies
may be preferred. For example, the phrases may be converted into concept-space document
vectors using equation (2), and scored in accordance with the phrase vector's similarity to the
cluster centroid, e.g., the average document vector for the cluster. (This similarity may be
termed the phrase-vector-to-cluster-centroid similarity.) Further details of labeling strategies
for block 522 are disclosed below in the description of Fig. 9.
[0041] Exploration operations 528 begin at the root of the clustering tree (e.g., node 620 of
Fig. 6B). In block 530, the information exploration system 100 displays representative phrases for the clusters at the current position in the clustering tree. Initially, a single root
cluster would be displayed as a label comprising the representative phrases for that cluster. In block 532, the information exploration system 100 processes user input. Expected user input
includes a termination command and selection of a different node in the clustering tree. A
termination command causes the information exploration operations 528 to halt. Otherwise, in block 534, the information exploration system changes the current position in the
clustering tree, and in block 536, the information exploration system determines whether the current position is a leaf node. If not, the information exploration system returns to block 530.
[0042] If a leaf node has been reached, then in block 538, the information exploration system 100 shows a list of titles of the documents in the cluster, and allows the user to examine (in a
separate window) the contents of documents selected by the user. In block 540, the information exploration system 100 determines whether user input is a termination command
or a position change. As before, a termination command halts the exploration operations 528,
and a position change sends the information control system 100 back to block 534.
[0043] Illustrative display screens are shown in Figs. 1OA and 1OB. In Fig. 1OA, display
screen 1002 shows a label 1004 representing a root cluster. Label 1004 appears on the left
("sibling") side ot display Screen 1002. When a user selects label 1004 (or any other labels on
the sibling side of the screen), the right ("child") side of the display screen shows labels
representing the sub-clusters of the selected cluster, and a "Back" link 1006. When a user
selects a label (e.g., label 1008) on the child side of the display screen, the contents of the
child side of the screen are transferred to the sibling side of the screen, and the sub-clusters of
the selected cluster are shown on the child side of the display. Fig. 1OB shows an example of
display 1002 after label 1008 has been selected. If the selected cluster has no sub-clusters (i.e., the selected cluster is a leaf node), the information exploration system 100 shows on the right side of the display a list of document titles for the selected cluster.
[0044] Except when the current position is the root node, the "Back" link 1006 causes
information exploration system 100 to transfer the contents of the sibling side of the display to the child side of the display, and to display the selected cluster's parent cluster and siblings
of the parent cluster. When the current position is the root node, the Back link 1006 causes
the root node to be de-selected and clears the child side of the display.
[0045] Though cluster quality is important, the quality of the cluster labels is often even more
important to the user. The representative phrases should, at a glance, give the user some understanding of what documents are to be found in the cluster. Some sophistication can therefore be justified in the selection process represented by block 526.
[0046] Fig. 9 shows an illustrative cluster-naming process. The process iterates systematically
through the clustering tree, proceeding either in a bottom-up fashion (processing all nodes at a
given depth in the tree before progressing closer to the root node) or in a top-down fashion.
Beginning with block 902, the information exploration system selects a first node of the clustering tree. Blocks 904-924 represent a loop that is executed to iterate through the master
phrase index, restarting at the beginning of the phrase index each time a new node is selected.
M block 904,' the inlόrmation exploration system tests whether the end of the index has been
reached without selecting enough representative phrases for the current node. If so, a flag is
set in block 906 to waive the exclusivity requirement, and the loop iterations are restarted at
the beginning of the phrase index.
[0047] In block 908, the information exploration system 100 selects the highest scoring,
previously unselected phrase for the current node as a candidate phrase. As discussed
previously, the scoring strategy may be designed to create a selection preference for longer phrases. Alternatively, a strict rule may be enforced on phrase length, restricting representative phrases to a length between two and five significant words, inclusive.
Throughout the phrase indexing process, each indexed phrase may be associated with at least one pointer to a document where the original phrase (i.e., before document cleaning and
stemming operations) appears. When an indexed phrase is selected as a representative phrase,
the information exploration system 100 provides an example of the original phrase as part of the cluster label.
[0048] In block 910, the information exploration system 910 tests the phrase exclusivity, i.e., whether the phrase appears in any leaf nodes that are not descendants of the current node. If the phrase is not exclusive, the information system determines in block 912 whether the exclusivity requirement has been waived and replaced by more relaxed coverage test, e.g.,
whether the phrase's coverage of the leaves of the current node is at least 20% higher than
that phrase's coverage of leaves that are not descendants of the current cluster. If the exclusivity requirement has not been waived or the more relaxed coverage requirement is not
satisfied, then the information exploration system returns to block 904.
[0049] Conversely, if the exclusivity or more relaxed coverage requirements are satisfied,
then in block 914 the information exploration system 100 compares the candidate phrase to
previously selected representative phrases for the current node. If the newly selected
representative phrase is a superstring of one or more previously selected phrases, the one or
more previously selected phrases are dropped unless the difference in cluster coverage
exceeds a predetermined threshold, e.g., 20%. For example, if the newly selected phrase
"chairman of the board" has a leaf cluster coverage of 30%, and the previously selected
phrases "chairman" and "board" have leaf cluster coverages of 35% and 75%, respectively, the previously selected phrase "chairman" would be dropped and the previously selected phrase "board" would be retained due to the difference in leaf cluster coverage.
[0050] In block 914, the information exploration system 100 also determines whether the
newly selected phrase has more than 60% of its significant words (i.e., words that are not stop words) appearing in any one previously selected phrase. If so, the newly selected phrase will
be dropped. Of course, the overlap threshold is programmable and can be set to other values.
If the candidate is dropped, the information exploration system returns to block 904.
[0051] If the previous tests are satisfied, the information exploration system 100 may further apply a test for path- wise uniqueness in optional block 916. In optional block 906, the information exploration system 100 drops candidate phrases that are path- wise non-unique. When the process proceeds in a top-down fashion, the current node represents one end of a path from the root node. To aid in the exploration of the document set, the representative phrases used to label clusters preferably change as the clusters become smaller and more
focused. Accordingly, the information exploration system 100 in block 906 drops phrases
from the phrase index if those phrases have already been selected as representative of a
previous cluster in the path. Thus a user, in following any given path from root node to leaf
node in the clustering tree, will not encounter any representative phrase more than once, making the representative phrases "path-wise" unique.
IUU5ZJ Utner uniqueness tests could be used. For example, in a bottom-up process, the
information system may drop phrases from the phrase index if those phrases have been
selected as representative of any subordinate clusters of the current clusters, i.e.,
"descendants" in the tree such as children, grandchildren, etc. When clustering quality is high,
the uniqueness of the representative phrases is expected to be inherent in the exclusionary
phrase indices, and accordingly, block 916 maybe treated as an optional operation.
[0053] In optional block 918, the information exploration system 100 determines whether one or more sub-clusters are being underrepresented by the selected phrases. For example, if five
representative phrases are to be selected to represent a given cluster, and all of the five phrases that have been selected have less than 10% coverage of a given sub-cluster's leaves,
the newly selected phrase may be dropped in favor of the next highest-scoring phrase having
at least a 25% coverage of the given sub-cluster's leaves. The representation thresholds are programmable and may be allowed to vary based on cluster size. In at least some
embodiments, at least one "slot" is reserved for each sub-cluster to assure that none of the sub-clusters go without representation in the cluster label. The number of reserved slots is
thus equal to the chosen branching factor for the clustering tree. In some implementations of these embodiments, the reserved slot may be released for general use if a previously-selected
phrase provides high coverage of the associated sub-cluster's leaf nodes.
[0054] hi block 920, the information exploration system determines whether the selection of
representative phrases for the current clustering tree node is complete. In some embodiments,
this determination is simply a comparison with the desired number of representational phrases
for the current node. In other embodiments, this determination is a comparison of the overall coverage of the selected representational phrases to a desired coverage threshold, hi still other
embodiments, this determination is a comparison of the coverages of selected phrases with
coverages oi avanaoie pnrases to determine a rough cost-to-benefit estimate for selecting
additional representational phrases.
[0055] If the selection is not complete, the information exploration system loops back to
block 904. Otherwise the information exploration system 100 determines in block 922
whether there are more nodes in the clustering tree. If so, the information exploration system
100 selects the next node in block 924. If not, the information exploration system terminates
the cluster-naming process.
[0056] A number of programmable or user-selectable parameters may be tailored for various
applications of the disclosed information exploration methods and systems. For example, the leaf cluster size may be altered to provide a trade-off between clustering quality and the
number of clusters. The branching factor of the clustering tree and maximum tree depth can be altered to suit user tastes. Similarly, the number of representative phrases can be tailored to
trade off between viewing ease and viewable detail. The cluster-naming processes disclosed
herein are applicable to any cluster set, irrespective of how that cluster set is obtained. Thus, the disclosed cluster-naming processes can be used with various clustering methods, which in turn can each be based on various similarity measures. Both hierarchical and non-hierarchical
clustering methods can be used, though the information exploration system 100 may be expected to perform best with mutually exclusive cluster sets, i.e., sets of clusters that do not have any documents in more than one cluster.
[0057] Numerous other variations and modifications will become apparent to those skilled in
the art once the above disclosure is fully appreciated. It is intended that the following claims
be interpreted to embrace all such variations and modifications.