Comparative analysis of c99 and topictiling text segmentation algorithms

IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
_______________________________________________________________________________________
Volume: 02 Issue: 09 | Sep-2013, Available @ http://www.ijret.org 529
COMPARATIVE ANALYSIS OF C99 AND TOPICTILING TEXT
SEGMENTATION ALGORITHMS
Sukhpreet Kaur, Kamaljeet Kaur Mangat
M.TECH (CSE) Student, Punjabi University Regional Centre of IT and Management, Mohali, Punjab, India,
sukhpreetkaur125@gmail.com
Assistant Professor (CSE), Punjabi University Regional Centre of IT and Management, Mohali, Punjab, India,
kamalmangat@yahoo.co.in
Abstract
In this paper, the work done includes the extraction of information from image datasets which contain natural text. The difficulty
level of segmenting natural text from an image is too high and so precision is the most important factor to be kept in mind. To
minimize the error rates, error filtration technique is provided, as filtration is adopted while doing image segmentation basically
text segmentation present in images. Furthermore, a comparative analysis of two different text segmentation algorithms namely
C99 and TopicTiling on image documents is presented. To assess how well each algorithm works, each was applied on different
datasets and results were compared. The work done also proves the efficiency of TopicTiling over C99.
Index Terms: Text Segmentation, text extraction, image documents,C99 and TopicTiling.
--------------------------------------------------------------------***----------------------------------------------------------------------
1. INTRODUCTION
In computational linguistics (CL) and natural language
processing (NLP), text segmentation is the task of splitting
text into a set of segments which are coherent about a
specific topic. The coherently segmented text enables
effective querying, analysis and usage. There are two main
approaches for Text segmentation namely, lexical cohesion
based approaches and feature based approaches. Lexical
cohesion based approaches depend on the tendency of topic
units to hang together. In feature based approaches, features
like cue phrases, full proper nouns and named entities are
used to detect boundaries between topics [1].Segmentation
is used as a pre-processing step by NLP systems that
perform various tasks such as video and audio retrieval,
question answering, subjectivity analysis, automatic
summarization, etc. Apart from these tasks, there are
applications where text has to be segmented from the image
data. The image documents are segmented to extract natural
text [2]. The algorithms from lexical cohesion based
approach are studied for segmentation. The algorithms
undertaken for this study are C99 and TopicTiling to
segment text from images. C99 segments a document
horizontally in blocks and segmentation is done
sequentially. The basic unit for C99 is a block which is a
group of words. In TopicTiling, a document is treated as a
hierarchical distribution of topics and thus performs
segmentation in a hierarchical manner. The basic unit in this
algorithm is a topic which can be a word or a group of
words. Each unit is assigned topic id, based upon which text
segmentation is done. Extensive research has been done on
English literatures. Basically C99(presented by choi) and
Topic Tiling are extensively used to segment text present in
long literature documents but very few have tried to segment
text from document of images. Presented systems are based
on measuring the lexical cohesion between textual units.
This research paper is organized as follows: section 2
presents related work, section 3 presents an overview of the
analysed systems; results and discussion are reported in
section 4 and finally section 5 concludes the paper.
2. RELATED WORK
Lots of work has been done on text segmentation techniques
implemented on different literature corpus. Hearst [3]
introduced the first topic based text segmentation algorithm
named as: TextTiling which segments texts in linear time by
calculating the similarity between two blocks of words
based on the cosine similarity. The calculation is
accomplished by two vectors containing the number of
occurring terms of each block. Galley [4] presented a
TextTiling based algorithm known as LcSeg in comparison
to TextTiling, it uses tf-idf term weights, which improves
TS results. Choi [5] introduced an algorithm called C99 that
uses a matrix-based ranking and a clustering approach in
order to relate the most similar textual units. Similar to the
previous introduced algorithms, C99 uses words. He again
improves C99 related to work undertaken , modified C99
algorithm, in 2001 that uses the term-representation matrix
in latent space of LSA in combination with a term frequency
matrix to calculate the similarity between sentences.
Utiyama and Isahara [6] introduced one of the first
probabilistic approaches using Dynamic Programming (DP)
called U00. DP is a paradigm that can be used to efficiently
find paths of minimum cost in a graph. Text Segmentation
algorithms using DP, represent each possible segment (e.g.

_______________________________________________________________________________________
every sentence boundary) as an edge. Providing a cost
function that penalizes common vocabulary across segment
boundaries, DP can be applied to find the segments with
minimal cost. Misra et al. and Sun et al. [7] presented two
new DP approaches: here, topic modeling is used to
alleviate the sparsity of word vectors. The algorithm of Sun
et al. follows the approach described in Fragkou et al., but
uses a combination of topic distributions and term
frequencies. Misra et al. [7] extended the DP algorithm from
Utiyama and Isahara (2001) using topic models. Instead of
using the probability of word co-occurrences, they use the
probability of co-occurring topics. Segments with many
different topics have a low topic-probability, which is used
as a cost function in their DP approach. (Misra et al., 2009)
trained the topic model on a collection of the Reuters corpus
and a subset of the Choi dataset, and tested on the remaining
Choi dataset. The topics for this test set have to be generated
for each possible segment using Bayesian inference
methods, resulting in high computational cost.
Mulbregt et al. [8] introduced a further approach for text
segmentation i.e. the usage of Hidden Markov Model
(HMM). Blei and Moreno introduced an Aspect Hidden
Markov Model (AHMM) which combines an aspect model
(Hofmann) with a HMM. The limiting factor of both
approaches is that a segment is assumed to have only one
topic. This problem has been solved by Gruber et al. who
extends LDA to consider the word and topic ordering using
a Markov Chain. In contrast to LDA, not a word is assigned
to a topic, but a sentence, so the segmentation can be
performed sentence-wise. Yaari [9] proposed the first
hierarchical algorithm using the cosine similarity and
agglomerative clustering approaches. A hierarchical
Bayesian algorithm based on LDA is introduced by
Eisenstein.
The current paper also includes image processing techniques
by which natural text gets extracted from images. Lots of
work has been done on different domains of image
processing. Ohya et al. [10] presented a method for the
extraction and recognition of characters in real scene
images. They work with grey-level images, and proposed
binarising image using threshold of sub-blocks, then those
thresholds interpolate whole image, but it generates lot of
noise. Wu et al. [11] proposed a two-step method to segment
an image into text regions, and a robust (but not very
generic) way to binarize the extracted text regions. First, a
texture segmentation algorithm takes advantage of certain
characteristics of text to segment areas possibly containing
text lines. The second phase focuses on the previously
detected areas and constructs “chips” from strokes taking
into account text characteristics. Messelodi and Modena
[12] proposed a method which better addresses problems of
different orientation angles of text lines in real scene
images. The method proposed, does address skew problems,
but it is only described and tested on images of book covers.
Clark and Mirmehdi [13] go one step further, and suggest
two approaches to the location and recovery of text in real
scenes. The first method is focused on situations where text
lies on the surface of an object and the edges of the object
are rectangular and aligned with the text. This is the case for
paper documents, posters, signs, stickers etc. Lopresti and
Zhou [14] proposed two methods to locate text in images,
as well as methods to recognize text. It was the main
contribution to the specific problem of Web Image text
extraction. And although they make a number of
assumptions that do not always hold, their approaches can
produce satisfactory results to a significant sub-group of
Web Images, namely images stored as GIF files (8-bit
palettized colour).
3. IMPLEMENTATION
Design of this research work is basically implemented into
two phases – Text Segmentation from Images and
Segmentation of texts based on lexical cohesion which is
received from phase-I.
3.1 Phase I-Text segmentation from images
The first phase concerns with extracting natural text from
images which is proposed and implemented, passing
through stages shown in Figure 1.
Block Diagram:
Fig- 1: Extraction of natural text from image
Input – Images containing natural text can be black-white
or colored image, but it is important that it must have natural
text so that natural text gets segmented.
Gray Scale Conversion – It is basically a preprocessing
stage which takes images as input which are usually
contaminated with noise and are often in the Red-Green-
Blue (RGB) color space, and convert the 3 color
components (i.e. R, G and B) to the intensity component by
[15]:
Y=0.299R+0.587G+0.114B
This is the actual Value component of the Hue-Saturation-
Value (HSV) color space and the noise of the images is
reduced using a weighted median filter that is applied on
this component using the mask of:
1 1 1
1 3 1
1 1 1
After this filtering step, a great part of noise will be removed
while the edges in the image are still preserved.
Input –
Images
containing
natural text
Gray Scale
Conversion
Calculating
Threshold
Output –
Natural
Text from
Images
Text
Extraction
after
segmenting
image
Boundary
or Edge
Detection
of
Similar
Objects

_______________________________________________________________________________________
Calculating Threshold - “Thresholding is an image point
operation which produces a binary image from a gray scale
image. A binary one is produced on the output image
whenever a pixel value on the input image is above a
specified minimum threshold level. A binary zero is
produced otherwise[16].” Herein Multi - Thresholding is
used which is defined as a point operator employing two or
more thresholds. Pixel values which are in the interval
between two successive threshold values are assigned an
index associated with the interval”. Multi-thresholding is
described mathematically as [16]:
S(x, y) = k, if Tk-1 ≤ f (x, y) < Tk, k=0, 1, 2, …, m
where (x, y) is the x and y co-ordinate of a pixel, S(x, y),
f(x, y) are the segmented and the characteristic feature
functions of (x, y) respectively, T0, …, Tm are threshold
values with T0 equal to the minimum and Tm the maximum
and m is the number of distinct labels assigned to the
segmented image [17].
Boundary Detection - Edge detection based segmentation
techniques are based on the detection of discontinuities in
the image. Discontinuities in a grey-level image are defined
as abrupt changes in grey-level intensity values. Edges are
therefore defined as the points where significant
discontinuities occur. The importance of edge detection lies
to the fact that most of the information of an image lies on
the boundaries between different regions, and that biological
visual systems seem to make use of edge detection rather
than thresholding [18]. Ideally, edge detection should yield
pixels lying only on the boundary between the regions one
tries to segment. In practice, the identified pixels alone
rarely define a correct boundary sufficiently because of
noise, breaks in the boundary and other effects. More often
than not, the edge detection stage is followed by an edge
linking and boundary detection stage, intending to combine
pixels into meaningful boundaries.
Edge detection methods can be classified as parallel and
sequential [19].In parallel techniques which is herein used,
the decision whether a pixel is an edge pixel or not is not
dependent on the result of the operator on any previously
examined pixels, whereas in sequential techniques it is. The
edge detection operator in parallel techniques can therefore
be applied simultaneously everywhere in the image.
Text Extraction after segmenting image - The procedure
of extracting natural text from images done in this research
paper is based on sobel edge detection of boundaries and
then comparing it with probable values or it can be
understood as aspect ratio of alphabets and digits as
mentioned standardly in matlab. Simply for sobel detection
matlab provides fx(imread) which reads binary image and
detected boundaries are compared with probable values.
Example for identifying „A‟ probable value is 36. After
comparing if probabilistic chance or aspect ratio suggests
that the text is „A‟ (for example) then output is declared as
„A‟.
Original Segmented Segmented
Natural Text
Characters
Fig-2: Example of text segmentation
Output – Natural text including all alphabets and digits
which will act as an input for phase two where comparison
between text segmentation algorithms will be done.
3.2 Phase II-Text Segmentation
The first phase concerns with extracting natural text from
images which is proposed and implemented, passing
through stages.
Fig- 3: Block diagram of text segmentation
Pre Processing Phase - TS algorithms which are evaluated
in this research have basic block units as “words”, so there
is a need to understand concept of breaking words from
sentences. But before segmenting sentences into words,
rules defining lexical cohesion must be made. In a sentence
apart from words there are symbols like punctuation marks
(“” , . ! & ; : „ ?) etc.and abbreviations. Words are also of
different categories based on parts of speech like noun,
pronoun, adjectives, verbs and conjunctions. A rule
repository having all these semantic rules has been created
prior which will further help in assigning topic ids to words
and dividing sentences into topics. To segment the text line
under consideration into words, a threshold has to be
determined to distinguish pairs of connected components
that belong to one word from these belonging to two
separate words. For this purpose, the line width „w1‟, the
median stroke width „ws‟ and the median distance between
Pre
Processing
Phase:
Creation of
Lexical
and Semantic
Rules
C99
Implementation
Conversion
of
Segmented
Words
into Topics
TopicTiling
Implementation
Output:
Segmented Text

_______________________________________________________________________________________
two vertical strokes „ds‟ are considered. The line width „w1‟
is defined as follows [20]:
w1=max(p(x,y)=black) – min(p(x,y) = black)
It represents the horizontal distance between the leftmost
and rightmost black pixel in the considered line of text. The
median stroke width „ws‟ is defined according to the
following formula:
Ws= median (h[y])
Using the horizontal projection histogram h[y] it represents
the median of the number of black pixels, taken over all
rows of the image of the considered text line. Finally the
threshold tseg for line segmentation is given by the following
formula:
tseg = alpha * (w1 – ws) /t
where, alpha is scaling constant that needs to be
experimentally determined. By deleting connections in the
minimal spanning tree which are longer than the threshold
tseg the text line is segmented into single words. For each
text line, the threshold tseg is computed separately.
From Words to Topics – This phase instead of using words
directly as features to characterize textual units, uses their
topic IDs as assigned by Bayesian inference. LDA inference
assigns a topic ID to each word in the text document in each
inference iteration step, based on lexical rules created in
preprocessing phase.
Implementation of C99 - The topic based version of the
C99 algorithm [21], called C99LDA, divides the input text
into minimal units on sentence boundaries. A similarity
matrix Sm x m is computed, where m denotes the number of
units (sentences). Every element sij is calculated using the
cosine similarity [22] between unit i and j. For these
calculations, each unit i is represented as a T-dimensional
vector, where T denotes the number of topics selected for
the topic model. Each element tk of this vector contains the
number of times topic ID k occurs in unit i. Next, a rank
matrix R is computed to improve the contrast of S: Each
element rij contains the number of neighbors of sij that have
lower similarity scores then sij itself. This step increases the
contrast between regions in comparison to matrix S. In a
final step, a top-down hierarchical clustering algorithm is
performed to split the document into m segments. This
algorithm starts with the whole document considered as one
segment and splits off segments until the stop criteria are
met, e.g. the number of segments or a similarity threshold.
At this, the ranking matrix is split at indices i, j that
maximize the inside density function D [20].
D=
TopicTiling Implementation - This section introduces Text
Segmentation algorithm called TopicTiling which is based
on TextTiling, but conceptually simpler. TopicTiling
assumes a sentence si as the smallest basic unit. Between
each position p between two adjacent sentences, a coherence
score cp is calculated. To calculate the coherence score, we
exclusively use the topic IDs assigned to the words by
inference: Assuming an LDA model with T topics, each
block is represented as a T-dimensional vector. The tth
element of each vector contains the frequency of the topic
ID t obtained from the respective block. The coherence
score is calculated by cosine similarity for each adjacent
“topic vector”. Values close to zero indicate marginal
relatedness between two adjacent blocks, whereas values
close to one denote a substantial connectivity. Next, the
coherence scores are plotted to trace the local minima.
These minima are utilized as possible segmentation
boundaries. But rather using the cp values itself, a depth
score dp is calculated for each minimum [3]. In comparison
to TopicTiling, TextTiling calculates the depth score for
each position and then searches for maxima. The depth
score measures the deepness of a minimum by looking at the
highest coherence scores on the left and on the right and is
calculated using the following formula:
dp = 1/2 _ (hl(p) − cp + hr(p) − cp)
The function hl(p) iterates to the left as long as the score
increases and returns the highest coherence score value. The
same is done, iterating in the other direction with the hr(p)
function. If the number of segments n is given as input, the n
highest depth scores are used as segment boundaries.
Otherwise, a threshold is applied. This threshold predicts a
segmentation if the depth score is larger than μ − σ/2, with μ
being the mean and σ being the standard variation calculated
on the depth scores. The algorithm runtime is linear in the
number of possible segmentation points, i.e. the number of
sentences: for each segmentation point, the two adjacent
blocks are sampled separately and combined into the
coherence score.
4. RESULTS AND DISCUSSION
4.1 TopicTiling Results on different datasets
(images)
Table 1 shows the outcomes of 5 experiments performed
over 5 different datasets and they are evaluated on three
performance parameters: Precision, Recall and WindowDiff.
Table -1: Results of TopicTiling Algorithm
Precision Recall WindowDiff
92.4242 13.2461 0.21639
96.2963 17.4623 0.20769
94.3662 12.7276 0.21194
93.75 11.0938 0.21333
95.1807 10.9856 0.21013

_______________________________________________________________________________________
4.2 C99 Results on different datasets (images)
Table 2 shows the outcomes of 5 experiments performed
over 5 different datasets and they are evaluated on three
performance parameters: Precision, Recall and WindowDiff.
Table -2: Results of C99 Algorithm
Precision Recall WindowDiff
72.549 11.4802 0.27568
88.0952 19.7846 0.22703
80.3571 12.3852 0.24889
67.6923 7.1834 0.29545
79.4118 9.6194 0.25185
4.3 Comparison of TopicTiling and C99 algorithms
Table 3 shows the comparison of the two algorithms
TopicTiling and C99 based on the results obtained and
shown in Table 1 and Table 2.
Table -3: Comparison of TopicTiling and C99 Algorithms
Parameters TopicTiling C99
Average
Precision
94.397 77.62
Average Recall 12.92 12.092
Average
WindowDiff
0.21 0.259
The experiments show that the average precision of
TopicTiling algorithm is 94.397 as compared to that in C99
which is 77.62. High precision value in TopicTiling clearly
shows improvement in text segmentation when
segmentation is done topic based. There is one more trend
need to be observed that is, precision is directly proportional
to Recall and inversely proportional to WindowDiff. More
the precision more recall and less is the window diff. It
means less is the value for WindowDiff , better is the
algorithm. So in this case TopicTiling algorithm performs
better since the value for precision in case of TopicTiling is
94.397% which is more than that in case of C99 with
77.62% as precision. The value for recall in TopicTiling is
12.92 which again is more than the recall value 12.092 in
case of C99. The value for WindowDiff in case of
TopicTiling is 0.21 and is marginally less than in case of
C99 with 0.259 as value for WindowDiff. Although the
difference in values for Recall and WindowDiff is marginal,
still they contribute in proving TopicTiling as a better
algorithm than C99. As C99 does not assign topic ids and
works on matrix rank calculation system, so precision in
extracting text from sentence drops down marginally
whereas TopicTiling uses topic ids to identify topics and
calculate coherence factor for each sentence and so
the precision in extracting text from sentence rises up
marginally.
CONCLUSIONS
The Research was started with the objective of comparing
and evaluating two text segmentation algorithms: C99 and
TopicTiling and has gradually reached the conclusive phase.
Both the text segmentation algorithms were performed and
implemented on image datasets which contains text. Till
now this area was scarce. The motive of this research, to
find a better solution to segment text which minimizes
computational cost and windowDiff and improve precision
and recall and thus performance, has been accomplished.
The work performed and implemented showed that
replacing words in documents by topic IDs, leads to better
results in the Text Segmentation task. This technique is
applied in the TopicTiling and C99 algorithms. In
comparison to other TS algorithms using topic models the
runtime of TopicTiling is linear in the number of sentences.
During implementation of TopicTiling, repeating the
inference several times and using the most frequently
assigned topic IDs in the last iteration not only reduces the
variance, but also improves overall results. The important
part of implementation is that performance is equal and at
the same time minimize the computational cost. This
method is not only applicable to Text Segmentation, but in
all applications where performance crucially depends on
stable topic ID assignments per token as this research shows
with datasets having images. The higher precision and less
windowDiff justifies the objective of research but in future
there is still scope of improvement. More granual testing in
different scenarios is needed and filling pitfalls of C99 or
TopicTiling will always be welcomed. The concern is to
devise a method to detect the optimal setting for the window
parameter w automatically. Moreover, research can be done
to extend the usage of the algorithms undertaken to more
realistic corpora. Equipped with a highly reliable
segmentation mechanism, application of text segmentation
as a writing aid to assist authors with feasible segmentation
boundaries and work in future would also assist in
segmenting and classifying web documents with precision
as getting consistent and relevant information is need of
hour. This could be applied in an interactive manner by
giving feedback about the coherence during the writing
process.
REFERENCES
[1] Michael A. El-Shayeb, Samhaa R. El-Beltagy and
Ahmed Rafea “Comparative Analysis of Different text
Segmentation Algorithms on Arabic News Stories,” in
proceedings of Information Reuse and Integration-IRI,
pp.441-446, 2007
[2] Shohreh Kasaei, Roshanak Farhoodi, “Text
Segmentation from Images with Textured and Colored

_______________________________________________________________________________________
Background”, 13th
Iranian conference on Electirc
Engineering, 1384.
[3] Hearst, M. A., “TextTiling: Segmenting Text into
Multiparagraph Subtopic Passages”, Computational
Linguistics, 1997.
[4] Galley, M., McKeown, K., Fosler-Lussier, E., Jing,
H.Discourse, “Segmentation of multi-party conversation.”
In Proceedings of the 41st Annual Meeting on Association
for Computational Linguistics, volume 1, pages 562–569,
Sapporo, Japan.
[5] Choi, F. Y. Y., “Advances in domain independent linear
text segmentation”, Proceedings of the 1st North American
chapter of the Association for Computational Linguistics
conference, Seattle, WA, USA, 2000.
[6] Utiyama, M. Isahara H., “A statistical model for domain-
independent text segmentation” Proceedings of the 39th
Annual Meeting on Association for Computational
Linguistics, Toulouse, France, 2001.
[7] Misra H., Yvon F., Jose, J. M., Cappe, O., “Text
Segmentation via Topic Modeling: An Analytical Study”,
Proceedings of the 18th ACM Conference on Information
and Knowledge Management, Hong Kong, 2011.
[8] Mulbregt, P. v., Carp, I., Gillick, L., Lowe, S., Yamron,
J., “Text segmentation and topic tracking on broadcast news
via a hidden markov model approach.” In Proceedings of
5th International Conference on Spoken Language
Processing, Sydney, Australia.
[9] Yaari, Y.,” Segmentation of expository texts by
hierarchical agglomerative clustering.” In Proceedings of
the Conference on Recent Advances in Natural Language
Processing, Tzigov Chark, Bulgaria.
[10] J. Ohya, A. Shio, S. Akamatsu, "Recognizing
Characters in Scene Images", Proceedings of IEEE
Transactions on Pattern Analysis and Machine Intelligence.
[11] V. Wu, R. Manmatha, E. M. Riseman, "Finding Text in
Images", Proceedings of 2nd
ACM International Conference
on Digital Libraries, Philadephia.
[12] S. Messelodi, C. M. Modena, "Automatic Identification
and Skew Estimation of Text Lines in Real Scene Images,"
Pattern Recognition, vol. 32, pp. 791-810, 1999.
[13] P. Clark, M. Mirmehdi, "Recognising Text in Real
Scenes", Proceedings of International Journal on Document
Analysis and Recognition.
[14] D. Lopresti, J. Zhou, "Document Analysis and the
World Wide Web," Proc. of the Workshop on Document
Analysis Systems, Marven, Pennsylvania, October 1996, pp.
417-424.
[15] Shohreh Kasaei, Roshanak Farhoodi, “Text
Segmentation from Images with Textured and Colored
Background”, 13th
Iranian conference on Electirc
Engineering, 1384.
[16] R. M. Haralick, "Glossary of Computer Vision Terms,"
Pattern Recognition, Vol. 24, pp.69-93, 1991.
[17] K. S. Fu, J. K. Mui, "A Survey on Image
Segmentation", Pattern Recognition,
Vol. 13, Issue 1, pp. 3-16.
[18] A. Rosenfeld, A. C. Kak, “Digital Picture Processing”,
Proceedings of New York
Academic Press, 1969.
[19] L. S. Davis, "A Survey of Edge Detection
Techniques", Computer Graphics
and Image Processing, Vol. 4, pp. 248-270, 1975.
[20] Riedl, Martin, and Chris Biemann. "Text Segmentation
with Topic Models." Herausgegeben von: 47.

Comparative analysis of c99 and topictiling text segmentation algorithms

Recommended

More Related Content

What's hot (17)

Similar to Comparative analysis of c99 and topictiling text segmentation algorithms (20)

More from eSAT Journals (20)

Recently uploaded (20)

Comparative analysis of c99 and topictiling text segmentation algorithms