3.

Automatic Indexing

3.1 Classes of Automatic Indexing

− Automatic indexing is the process of analyzing an item to extract the information to be
permanently kept in an index.
− An index is the data structure created to support the search strategy.

− Search strategies can be classified as

− Statistical strategies.
− natural language approach
− concept indexing

− Statistical strategies
− use of frequency of occurrence of events is the basis for a statistical approach.
− events are related to occurrences of processing tokens (words/phrases) within
documents and within the database.
− words/phrases are the domain of searchable values.

− statistics that are applied to the event data are

− Probabilistic indexing stores the information that are used in calculating a
probability that a particular item satisfies a particular query.
− Bayesian and vector approaches store information used in generating a relative
confidence level of an item’s relevance to a query.
− Neural networks are dynamic learning structures that are used to determine
concept classes.

− Natural Language approaches

− perform processing token identification similarly as in statistical techniques
− Additionally perform varying levels of natural language parsing of the item that
generalizes to more abstract concepts within an item (e.g., present, past, future
actions).
− To enhance the search precision, this additional information is stored within the
index.

Concept indexing
− uses the words within an item to correlate to concepts discussed in the item.
− This is a generalization of the specific words to values used to index the item.
− a special class of indexing can be defined by creation of hypertext linkages.
 3.2 Statistical Indexing
− Statistical indexing

o Uses frequency of occurrence of events to indicate the relevance of an item.

▪ Probabilistic systems: calculate a probability value (Probabilistic Weighting)

▪ Bayesian Model and Vector Weighting approaches: Calculate a relative relevance
value.

3.2.1 Probabilistic Weighting

− based upon the theory of probability to information retrieval systems.

− Logistic Regression method

− Logistic Reference model
− Logistic Inference method

− Logistic Regression method

− starts by defining a “Model 0” system which exists before specific probabilistic
models are applied.
− In a retrieval system there exist query terms and document terms, which have a set of
attributes (v1…vn) from the query (e.g., counts of term frequency in the query), from the
document (e.g., counts of term frequency in the document ) and from the database (e.g.,
total number of documents in the database divided by the number of documents indexed by
the term).

− Logistic Reference model

− uses a random sample of query-document-term for which binary relevance results
have been made from the training sample.
− Log O:
▪ logarithm of the odds (logodds) of relevance for term tk which is present
in document Dj and query Qi.
− Logarithm that the ith Query is relevant to the jth Document:
▪ the sum of the log odds for all terms.

− O(R):
▪ Odds that a document chosen at random from the database is relevant to
query Qi.
− Probability of relevance of a document to a query:
 ▪ obtained by applying the inverse logistic transformation:

− Coefficients of the equation for logodds:

▪ derived for a particular database using a random sample of query-document-
term-relevance quadruples
▪ used to predict odds of relevance for other query-document pairs.

− logodds formula:

− QRF = QAF/ (total number of terms in the query),

− DRF = DAF/(total number of words in the document) and
− RFAD = (total number of term occurrences in the database)/ (total number of all
words in the database).

− Logistic Inference method

− applied to the test database along with the vector system (which uses traditional
term frequency, inverse document frequency and cosine relevance weighting
formulas).
− The logistic inference method outperformed the vector method.
− The index that supports the calculations for the logistic reference model contains
the O(R) constant value along with the coefficients.
− it needs to maintain the data to support DAF, DRF, IDF and RFAD.
− The values for QAF and QRF are derived from the query.
 3.2.2 Vector Weighting

− The semantics of every item are represented as a vector.

− Vector
− It is a one-dimensional set of values, where the order/position of each value in the
set is fixed and represents a particular domain.
− In information retrieval, each position in the vector typically represents a
processing token.
− There are two approaches. Binary and Weighted.

− Example. Figure shows how an item that discusses petroleum refineries in Mexico
would be represented.

− binary approach
− The domain contains the value of one or zero. One represents the existence of the
processing token in the item.
− Binary vectors require a decision process to determine that a particular processing
token represents the semantics of an item is sufficient to include it in the vector.
− Concepts of “Tax” and “Shipment” are below the threshold of importance (e.g.,
assume threshold is 1.0) and they not are included in the vector.

− weighted approach
− The domain is typically the set of all real positive numbers.
− The value for each processing token represents the relative importance of that
processing token in representing the semantics of the item.
− A weighted vector provides a range of values that contains a variance in the value
of the relative importance of a processing token in representing the semantics of
the item.
− Weights provide a basis for determining the rank of an item.
 − The vector approach allows for a mathematical and a physical representation using a
vector space model.
− Each processing token can be considered another dimension in item representation space.
− Figure 5.3 shows a three-dimensional vector representation assuming there were only
three processing tokens Petroleum, Mexico and Oil.

Figure 5.3 Vector Representation

Algorithms used in calculating the weights of a processing token.

− Simple Term Frequency Algorithm (STFA)
− Inverse Document Frequency (IDF)
− Signal Weighting (SW)
− Discrimination Value (DV)

Simple Term Frequency Algorithm

− In a statistical system, the data that are available for calculating a weight are
− Term Frequency (TF): frequency of occurrence of the processing token
in an existing item.
− Total Frequency (TOTF): frequency of occurrence of the processing
token in the existing database.
− Item frequency (IF) or Document Frequency (DF): number of unique
items in the database that contain the processing token.

− Approaches for different document lengths when determining the value of Term
Frequency are
− maximum term frequency (MTF)
▪ Term frequency for each word is divided by the maximum frequency of
the word in any item.
▪ This normalizes the term frequency values to a value between zero and
one.
 − logaritmetic term frequency (LTF)
▪ log of the term frequency plus a constant is used to replace the term
frequency.
▪ The log function will perform the normalization when the term
frequencies vary significantly due to size of documents.

Inverse Document Frequency

− The weight assigned to an item should be inversely proportional to the frequency of

occurrence of an item in the database.
− The un-normalized weighting formula is:

− where WEIGHTij is the vector weight that is assigned to term j in item i.

− Example:
− Assume that the term “oil” is found in 128 items, “Mexico” is found in 16 items and
“refinery” is found in 1024 items.
− If a new item arrives with all three terms in it, “oil” found 4 times, “Mexico” found 8
times, and “refinery found 10 times and there are 2048 items in the total database.
− Using a simple unnormalized term frequency, the item vector is (4, 8, 10).
− Figure 5.4 shows the weight calculations using inverse document frequency.

Figure 5.4 Example of inverse document frequency

− with the resultant inverse document frequency item vector = (20, 64, 20)
Signal Weighting

− Based upon the number of items that contain the term in the existing database.
− Inverse document frequency adjusts the weight of a processing token for an item.

− Example:
− Assume the terms “SAW” and “DRILL” are found in 5 items with the following
frequencies defined in Figure.

− Both terms are found a total of 50 times in the five items.

− The term “SAW” does not give any insight into which item is more likely to be relevant
to a search of “SAW”.
− The distribution of the frequency of processing tokens within an item can affect the
ability to rank items.

− Theoretical basis for the algorithm.

− In Information Theory, the information content value of an object is inversely
proportional to the probability of occurrence of the item.
− This is represented as INFORMATION = -Log2 (p)
− p is the probability of occurrence of event “p”.

− If there are many independent occurring events then the calculation for the average
information value across the events is:

− The value of pk can be defined as TFik/TOTFk, the ratio of the frequency of occurrence of
the term in an item to the total number of occurrences of the item in the data base.
− The weighting factor, Signal:
 − producing a final formula of:

Discrimination Value

− Another approach to create a weighting algorithm is based upon the discrimination value
of a term.
− To achieve the objective of finding relevant items, it is important that the index
discriminates among items.
− discrimination value for each term “i”is:

− AVESIM is the average similarity between every item in the database.

− AVESIMi is the same calculation except that term “i” is removed from all items.

− There are three possibilities with the DISCRIMi value

o Positive
▪ indicates that removal of term “i” has increased the similarity between
items.
o close to zero
▪ implies that the term’s removal or inclusion does not change the similarity
between items
o negative
▪ term’s effect on the database is to make the items appear more similar
since their average similarity decreased with its removal
− Once the value of DISCRMi is normalized as a positive number, it can be used in the
standard weighting formula as:
3.2.3 Bayesian Model

− Basis
− based upon conditional probabilities (Probability of Event 1 given Event 2
occurred).
− Bayesian formula
− probability of relevance (REL) to a search statement given a particular document
and query, P(REL/DOCi, Queryj).
− Uses/Application
− Bayesian model is applied to the search function and to creating the index to the
database.
− Bayesian formula can be used in determining the weights associated with a
particular processing token in an item.
− Bayesian network can be used to determine the final set of processing tokens
(called topics) and their weights.

− Bayesian Term Weighting process

− Figure shows a simple view of the process.

− Ti represents the relevance of topic “i” in a particular item

− Pj represents a frequency of occurrence associated with the event of processing
token “j” being present in the item.
− The “m” topics would be stored as the final index to the item.
− Assumption of Binary Independence:
− The topics and the processing token statistics are independent of each other.
− The existence of one topic is not related to the existence of the other topics.
− The existence of one processing token is not related to the existence of other
processing tokens.

− Issues
− Some topics are related to other topics and some processing tokens related to
other processing tokens.
− For example, the topics of “Politics” and “Economics” are in some instances
related to each other and in many other instances totally unrelated.
− There are two approaches to handling this problem.

− The first approach

− assume that there are dependencies, but the errors resulting by mutual
independence do not affect the determination of relevance of an item nor its
relative rank associated with other retrieved items.

− A second approach
− extends the network to additional layers to handle interdependencies.

− additional layer of Independent Topics (ITs) can be placed above the Topic layer.
− A layer of Independent Processing Tokens (IPs) can be placed above the
processing token layer.
 3.3 Natural Language

− Goal of natural language processing

o To enhance the indexing of the item by using the semantic information in addition
to the statistical information.
o To improve the precision of searches
o To reduce the number of false hits.

3.3.1 Index Phrase Generation

3.3.2 Natural Language Processing

3.3.1 Index Phrase Generation

− The goal of indexing is to represent the semantic concepts of an item in the information
system to support finding relevant information.

− Single words have conceptual context, but frequently they are too general to help the user
find the desired information.

− Term phrases allow additional specification and focusing of the concept to provide better
precision and reduce the user’s overhead of retrieving non-relevant items.

− COHESION factor determines term phrases between terms:

− SIZE-FACTOR is a normalization factor based upon the size of the vocabulary.

− PAIRFREQk,h is the total frequency of co-occurrence of the pair Term-k , Term-h in the
item collection.
− Co-occurrence may be defined in terms of adjacency, word proximity, sentence
proximity, etc.
3.3.2 Natural Language Processing

− It is assumed that Lexical analysis (which determines verb tense, plurality, and part of
speech) have been completed prior to the additional processing.
− produces more accurate term phrases
− provides higher level semantic information that identifies relationships between concepts.
− is not just determining the topic statement(s) but also assigning semantic attributes to the
topic such as time frame (past, present, future).

− Each sentence is evaluated and assigned weights associated with its possible inclusion in
the different components.

− The next level of semantic processing is the assignment of terms to components,

classifying the purpose of the terms in the text and identifying the topical statements.

− The next level of natural language processing identifies inter- relationships between the
concepts.

− The relationships are typically visualized as triples with two concepts and a relationship
between them.

− The weights are based upon a combination of statistical information and values assigned
to the actual words used in establishing the linkages.

− Passive verbs would receive less weight than active verbs.

− The additional information beyond the indexing is kept in additional data structures
associated with each item.
 3.4 Concept Indexing
− goal
− To use concepts instead of terms as the basis for the index
− To produce a reduced dimension vector space.

− Concept indexing
− can start with a number of unlabeled concept classes
− Let the information in the items define the concepts classes.

− Automatic creation of concept classes

− Similar to the automatic generation of thesaurus classes.
− The basis behind the generation of the concept approach is a neural network
model.
− Concepts would be independent in an “N” dimensional vector-space model.
− Example
− A term such as “automobile” could be associated with concepts such as “vehicle”,
“transportation”, “mechanical device”, “fuel”.
− But “automobile” is strongly related to “vehicle”, lesser to “transportation” and
much lesser to the other terms.
− Thus a term in an item needs to be represented by many concept codes with
different weights for a particular item.
− The Convectis system
− uses neural network algorithms for determining the terms that are related.
− A term can have different weights associated with different concepts.
− The definition of a similar context is typically defined by the number of non-stop
words separating the terms.
− The farther apart terms are, the less coupled the terms are associated within a
particular concept class.
− Existing terms already have a mapping to concept classes.
− New terms can be mapped to existing classes by applying some rules to the
classes that terms near the new term are mapped.
− Special rules must be applied to create a new concept class.

− Latent Semantic Indexing (LSI)

− The algorithms associated with vectors (e.g., inverse document frequency) can be
used to perform the merging of concepts.
− Example of this process is Latent Semantic Indexing (LSI).
− It’s assumed that there is a hidden or “latent” structure represented by
interrelationships between words.
− The index contains representations of the “latent semantics” of the item.
3.5 Hypertext Linkages

− Hypertext data structure

− is a new class of information representation.
− generated manually although user interface tools may simplify the process.
− hypertext linkages
− Create an additional information retrieval dimension.
− Traditional items can be viewed as two dimensional constructs.
− The text of the items is one dimension which represents the information in the
items.
− Imbedded references are a logical second dimension that has had minimal use in
information search techniques.
− The imbedding of the linkage allows the user to go immediately to the linked item
for additional information.

− There are three classes of mechanisms to help find information:

− manually generated indexes
− automatically generated indexes
− web crawlers (intelligent agents).

− E.g for manually generated indexes

− YAHOO (http://www.yahoo.com), where information sources (home pages) are
indexed manually into a hyperlinked hierarchy.
− The user can navigate through the hierarchy by expanding the hyperlink on a
particular topic to see the more detailed subtopics.
− At some point the user starts to see the end items.

− E.g for automatic generated indexes

− LYCOS (http://www.lycos.com) and Altavista (http://www.altavista.digital.com)
automatically go out to other Internet sites and return the text at the sites for
automatic indexing.
 − Lycos returns home pages from each site for automatic indexing while Altavista
indexes all of the text at a site.
− None of these approaches use the linkages in items to enhance their indexing.

− Web crawlers and intelligent agents

− are tools that allow a user to define items of interest and they automatically go to
various sites on the Internet searching for the desired information.
− e.g., WebCrawler, OpenText, Pathfinder, Coriolis Groups’ NetSeeker™
− They are better described as a search tool than an indexing tool.

− Hypertext linkages as an extension of the concepts

− An index algorithm that looks at the hypertext linkages as an extension of the
concepts being presented in the item is needed.
− Hypertext links can map to another item or to a specific location within an item.
− The current concept is defined by the information within proximity of the location of
the link.

− The weight of processing tokens:

− where is the Weight associated with processing token “j” in item “i”
and processing token “l” in item “k” that are related via a hyperlink.
− Linkik is the weight associated with strength of the link.
− α,β,ƴ are weighting/normalization factors.
− The values stored in an expanded index structure or calculated dynamically.
 3.6 Document and Term Clustering

− Clustering index terms

− to create a statistical thesaurus
− to increase recall by expanding searches with related terms.
− Clustering items
− to create document clusters.
− Search can retrieve items similar to an item of interest.

3.6 Introduction to Clustering

− Goal of clustering
− to assist in the location of information.
− to group similar objects (e.g., terms or items) into a “class” under.
− To allow linkages between clusters to be specified.

− Steps in the process of clustering:

(a)
− Define the domain for the clustering effort.
− If a thesaurus is being created, this equates to determining the scope of the thesaurus
such as “medical terms”.
− If document clustering is being performed, it is the determination of the set of items
to be clustered.
(b)
− Once the domain is determined, determine the attributes of the objects to be
clustered.
− If a thesaurus is being generated, determine the specific words in the objects to be
used in the clustering process.
− If documents are being clustered, the clustering process may focus on specific
zones within the items that are to be used to determine similarity.
©
− Determine the strength of the relationships between the attributes whose co-
occurrence in objects suggest those objects should be in the same class.
− For thesauri this is determining which words are synonyms and the strength of
their term relationships.
− For documents it may be defining a similarity function based upon word co-
occurrences that determine the similarity between two items.
 (d)
− At this point, the total set of objects and the strengths of the relationships between
the objects have been determined.
− The final step is applying some algorithm to determine the class(s) to which each
item will be assigned.

− Guidelines on the characteristics of the classes:

− A well-defined semantic definition should exist for each class.

− The size of the classes should be within the same order of level.
− Within a class, one object should not dominate the class.
− Whether an object can be assigned to multiple classes or just one must be decided
at creation time.
 3.7 Thesaurus Generation

− 3.7.1 Manual Clustering

− 3.7.2 Automatic Term Clustering
− 3.7.2.1 Complete Term Relation Method
− 3.7.2.2 Clustering Using Existing Clusters
− 3.7.2.3 One Pass Assignments

3.7.1 Manual Clustering

− The first step is to determine the domain for the clustering.

− Defining the domain reduces ambiguities.
− Usually existing thesauri, concordances from items that cover the domain.
− Dictionaries are used as starting points for generating the set of potential words to be
included in the new thesaurus.
− A concordance is an alphabetical listing of words from a set of items along with their
frequency of occurrence and references of which items in which they are found.
− Words that are unrelated to the domain of the thesaurus or those that have very high
frequency of occurrence which have no value are not included.

− If a concordance is used, other tools such as KWOC, KWIC or KWAC may help in
determining useful words.
− Key Word Out of Context (KWOC) is another name for a concordance.
− Key Word In Context (KWIC) displays a possible term in its phrase context.
− It is structured to easily identify the location of the term under consideration in the
sentence.
− Key Word And Context (KWAC) displays the keywords followed by their context.
− Figure 6.1 shows the various displays for “computer design contains memory chips”.
− The phrase is assumed to be from doc4; the other frequency and document ids for KWOC
were created for this example.
− In the Figure 6.1 the character “/” is used in KWIC to indicate the end of the phrase.
− The KWIC and KWAC are useful in determining the meaning of homographs.
− The term “chips” could be wood chips or memory chips.
− In both the KWIC and KWAC displays, the editor of the thesaurus can read the sentence
fragment associated with the term and determine its meaning.

− Once the terms are selected they are clustered based upon the word relationship
guidelines and the interpretation of the strength of the relationship.
− The resultant thesaurus undergoes many quality assurance reviews by additional editors
using some of the guidelines already suggested before it is finalized.
 3.7.2 Automatic Term Clustering

− The basis for automatic generation of a thesaurus (Terms)

o is a set of items that represents the vocabulary to be included in the thesaurus.
o Selection of this set of items is the first step of determining the domain for the
thesaurus.
o The processing tokens (words) in the set of items are the attributes to be used to
create the clusters.
− The basis for automated method of clustering documents (Items)
o is based upon the polythetic clustering where each cluster is defined by a set of
words and phrases.
o inclusion of an item in a cluster is based upon the similarity of the item's words
and phrases to those of other items in the cluster.
− Methods for Automatic generation of a thesaurus
o Complete Term Relation Method
o Clustering Using Existing Clusters
o One Pass Assignments

Complete Term Relation Method

Basis
− similarity between every term pair is calculated for determining the clusters.

o The vector model is represented by a matrix where the rows are individual items
and the columns are the unique words (processing tokens) in the items.
o The values in the matrix represent how strongly that particular word represents
concepts in the item.

o Figure 6.2 provides an example of a database with 5 items and 8 terms.

 o To determine the relationship between terms, a similarity measure is required.

o where “k” is summed across the set of all items.

o the formula takes the two columns of the two terms being analyzed,
multiplying and accumulating the values in each row.
− The results can be placed in a resultant “m” by “m” matrix, called a Term-Term
Matrix, where “m” is the number of columns (terms) in the original matrix.
o Using the data in Figure 6.2, the Term-Term matrix produced is shown
in Figure 6.3.

− There are no values on the diagonal since that represents the autocorrelation of
a word to itself.
− The next step is to select a threshold that determines if two terms are considered similar
enough to each other to be in the same class.
o In this example, two terms are considered similar if the similarity
value between them is the threshold value of 10 or greater.
o This produces a new binary matrix called the Term Relationship
matrix (Figure 6.4) that defines which terms are similar.
 o A ‘1’ in the matrix indicates that the terms specified by the column
and the row are similar enough to be in the same class.
− The final step in creating clusters is to determine when two objects (words) are in the
same cluster.

− The following algorithms are the most common:

o cliques
o single link
o stars
o Connected components.
 Clustering Using Existing Clusters
− It starts with a set of existing clusters.
− It reduces the number of similarity calculations required to determine the clusters.
− The initial assignment of terms to the clusters is revised by revalidating every term
assignment to a cluster.
− The process stops when minimal movement between clusters is detected.

− To minimize calculations, centroids are calculated for each cluster.

− A centroid will equate to the average of all of the vectors in a cluster.

− The first assignment of terms to clusters produces centroids that are not related to the
final clustering of terms.
− The similarity between all existing terms and the centroids of the clusters can be
calculated.
− The term is reallocated to the cluster(s) that has the highest similarity.
− This process is iterated until it stabilizes.
− Calculations using this process are of the order O(n).
− Example:
− A graphical representation of terms and centroids illustrates how classes move after the
initial assignment.
− The solid black box represents the centroid for each of the classes.
− In Figure 6.6b. the centroids for the first three arbitrary class are shown.
− The ovals in Figure 6.6b. show the ideal cluster assignments for each term.
− During the next iteration the similarity between every term and the clusters is performed
reassigning terms as needed.
− The resulting new centroid for the new clusters is again shown as black squares in Figure
6.6a.
− The new centroids are not yet perfectly associated with the ideal clusters, but they are
much closer.
− The process continues until it stabilizes.

One Pass Assignments

− This technique has the minimum overhead.

− Only one pass of all of the terms is used to assign terms to classes.
− The first term is assigned to the first class.
− Each additional term is compared to the centroids of the existing classes.
− A threshold is chosen.
− If the item is greater than the threshold, it is assigned to the class with the highest
similarity.
− A new centroid has to be calculated for the modified class.
− If the similarity to all of the existing centroids is less than the threshold, the term is the
first item in a new class.
− This process continues until all items are assigned to classes.
− Using the system defined in Figure 6.3, with a threshold of 10 the following classes
would be generated:

− This process does not produce optimum clustered classes.

− Items that would have been in the same cluster could appear in different clusters due to
the averaging nature of centroids.
 3.8 Item Clustering

− Clustering of items
o similar to term clustering for the generation of thesauri.
− Manual item clustering
o someone reads the item and determines the category or categories to which it
belongs.
o each item is usually assigned to one category.

− Automatic clustering of the items

o Similarity between documents is based upon two items that have terms in
common versus terms with items in common.
o similarity function is performed between rows of the item matrix.

− Using Figure 6.2 as the set of items and their terms and similarity equation:

− as k goes from 1 to 8 for the eight terms, an Item-Item matrix is created (Figure 6.9).

− Using a threshold of 10 produces the Item Relationship matrix shown in Figure 6.10.
− Algorithms used are
o Clique algorithm
o Single link technique
o Star technique
o String technique
 3.9 Hierarchy of Clusters

− Hierarchical clustering in Information Retrieval focuses on the area of Hierarchical

Agglomerative Clustering Methods (HACM).
− Agglomerative means the clustering process starts with unclustered items and performs
pair wise similarity measures to determine the clusters.
− Divisive is the term applied to starting with a cluster and breaking it down into smaller
clusters.

− The objectives of creating a hierarchy of clusters are to:

o Reduce the overhead of search
o Provide a visual representation of the information space
o Expand the retrieval of relevant items.

− Dendograms
− Search overhead is reduced by performing top-down searches of the centroids of the
clusters in the hierarchy and trimming those branches that are not relevant.
− It is difficult to create a visual display of the total item space.
− Use of dendograms along with visual cues on the size of clusters and strengths of the
linkages between clusters (e.g., dashed lines indicate reduced similarities) allows a user
to determine alternate paths of browsing the database.
− The dendogram allows the user to determine which clusters to be reviewed are likely to
have items of interest.
− Even without the visual display of the hierarchy, a user can use the Figure 6.12
Dendogram logical hierarchy to browse items of interest.
− A user, once having identified an item of interest, can request to see other items in the
cluster.

− HACM approaches can be defined as the dissimilarity D between any existing cluster Ck
and a new cluster Cij created by combining clusters Ci and Cj.

− By proper selection of α,∞, and ƴ

− Group average method produced the best results on document collections.

− The similarity between two clusters can be treated as the similarity between all objects in
one cluster and all objects in the other cluster.

− Voorhees showed that the similarity between a cluster centroid and any item is equal to
the mean similarity between the item and all items in the cluster.

− Ward’s Method chooses the minimum square Euclidean distance between points (e.g.,
centroids) normalized by the number of objects in each cluster.

− Automatic clustering techniques can also be used to create a hierarchy of objects (items
or terms).
− Centroids were used to reduce computation required for adjustments in term assignments
to classes.
− For both terms and items, the centroid has the same structure as any of the items or terms
when viewed as a vector from the Item/Term matrix.
− A term is a vector composed of a column whereas an item is a vector composed of a row.