SciPlore Xtract: Extracting Titles from Scientific PDF
Documents by Analyzing Style Information (Font Size)
Jöran Beel1,2, Bela Gipp1,2, Ammar Shaker1, Nick Friedrich1


1 Otto-von-Guericke University, Computer Science/ITI/VLBA-Lab, Magdeburg, Germany
2 UC Berkeley, Berkeley, California, USA
{beel, gipp, shaker, friedrich}@sciplore.org
Abstract. Extracting titles from a PDF‟s full text is an important task in
information retrieval to identify PDFs. Existing approaches apply complicated
and expensive (in terms of calculating power) machine learning algorithms such
as Support Vector Machines and Conditional Random Fields. In this paper we
present a simple rule based heuristic, which considers style information (font
size) to identify a PDF‟s title. In a first experiment we show that this heuristic
delivers better results (77.9% accuracy) than a support vector machine by
CiteSeer (69.4% accuracy) in an „academic search engine‟ scenario and better
run times (8:19 minutes vs. 57:26 minutes).
Keywords: header extraction, title extraction, style information, document
analysis
1 Introduction
Extracting the title from PDF documents is one of the prerequisites for many tasks in
information retrieval. Among others, (academic) search engines need to identify PDF
files found on the Web. One possibility to identify a PDF file is extracting the title
directly from the PDF‟s metadata. However, often the PDF metadata is incorrect or
missing. Therefore, what is often tried is to extract the title from the PDFs‟ full text.
Usually, machine learning approaches such as Support Vector Machines (SVM),
Hidden Markov Models and Conditional Random Fields are used for extracting titles
from a document‟s full text. According to studies, the existing approaches achieve
excellent accuracy, significantly above 90%, sometimes close to 100% [1, 2, 3].
However, all existing approaches for extracting titles from PDF files have two
shortcomings. First, they are expensive in terms of runtime. Second, they usually
convert PDF files to plain text and lose all style information such as font size.
For our academic search engine SciPlore.org we developed SciPlore Xtract, a tool
applying rule based heuristics to extract titles from PDF files. In this paper we present
this tool, the applied heuristics and an evaluation.
Preprint of: Joeran Beel, Bela Gipp, Ammar Shaker, and Nick Friedrich. SciPlore Xtract: Extracting Titles from Scientific PDF
Documents by Analyzing Style Information (Font Size). In M. Lalmas, J. Jose, A. Rauber, F. Sebastiani, and I. Frommholz, editors,
Research and Advanced Technology for Digital Libraries, Proceedings of the 14th European Conference on Digital Libraries
(ECDL‟10), volume 6273 of Lecture Notes of Computer Science (LNCS), pages 413–416, Glasgow (UK), September 2010. Springer.

2 SciPlore Xtract
SciPlore Xtract is an open source Java program that is based on pdftohtml1 and runs
on Windows, Linux and MacOS. The basic idea is to identify a title based on the rule
that it will be the largest font on the upper first third on the first page.


Figure 1: Example PDF

Figure 2: Example XML Output
In the first step, SciPlore Xtract converts the entire PDF to an XML file. In contrast to
many other converters, SciPlore Xtract keeps all layout information regarding text
size and text position. Figure 2 shows an example XML output file of the PDF
showed in Figure 1. Lines 6 to 12 of the XML file show all font sizes that are used in
the entire document (in this case it is all “Times” in a size between 7 and 22 points).
Below this, each line of the original PDF file is stated including layout information
such as the exact position in which the line starts, and which font is used.
SciPlore Xtract now simply needs to identify the largest font type (in the example
the font with the ID=0). Which text uses this font type on the first page is then
identified and to assumed to be the title.
3 Methodology
In an experiment, titles of 1000 PDF files were extracted with SciPlore Xtract.
Then, titles from the same PDFs were extracted with a Support Vector Machine from
CiteSeer [1] to compare results. CiteSeer‟s tool is written in Perl and based on SVM
Light2 which is written in C. As CiteSeer‟s SVM needs plain text, the PDFs were

1 http://www.pdftohtml.sourceforge.net
2 http://svmlight.joachims.org/

converted once with PDFBox3 and once with pdftotext4 as these are the tools
recommended by CiteSeer. It was then checked for each PDF if the title was correctly
extracted by SciPlore Xtract and CiteSeer‟s SVM (for both the pdftohtml text file and
the PDFBox text file). If the title contained slight errors the title was still considered
as being identified correctly. „Slight errors‟ include wrongly encoded special
characters or, for instance, the inclusion of single characters such as „*‟ at the end of
the title.
The PDFs analyzed were a random sample from our SciPlore.org database, a
scientific (web based) search engine. A title was seen as being correctly extracted
when either the main title or both the main title and the sub-title (if existent) were
correctly extracted. The analyzed PDFs were not always scientific. It occurred that
PDFs represented other kind of documents such as websites or PowerPoint
presentations. However, we consider the collection to be realistic for an academic
search engine scenario.
4 Results
From 1000 PDFs, 307 could not be processed by SciPlore Xtract. Apparently,
SciPlore Xtract (respectively pdftohtml) struggles with PDFs that consist of scanned
images on which OCR has been applied. For further analysis only the remaining 693
PDFs were used. We consider this legitimate as the purpose of our experiment was
not to evaluate SciPlore Xtract, but the applied rule based heuristic.
For 54 of the 693 PDFs (7.8%), titles could neither be extracted correctly by
SciPlore Xtract nor CiteSeer‟s SVM. Only 160 (23.1%) of the titles were correctly
identified by all three approaches. Overall, SciPlore Xtract extracted titles of 540
PDFs correctly (77.9%). CiteSeer‟s SVM applied to pdftotext identified 481 titles
correctly (69.4%). CiteSeer‟s SVM applied to PDFBox extracted 448 titles correctly
(64.6%). Table 1 shows all these results in an overview.
Table 1: Title Extraction of 693 PDFs
SciPlore Xtract 528 76.2% 12 1.7% 540 77.9%
CiteSeer SVM +
pdftotext 406 58.6% 75 10.8% 481 69.4%
CiteSeer SVM +
PDFBox 370 53.4% 78 11.3% 448 64.6%
Correct TotalSlight Errors


When only completely correct titles are compared, SciPlore Xtract performs even
better. It extracted 528 (76.2%) titles completely correct, while CiteSeer‟s SVM
extracted only 406 (58.6%) respectively 370 (53.4%) completely correct.
SciPlore Xtract required 8:19 minutes for extracting the titles. SVM needed 57:26
minutes for extracting the titles from the plain text files (this does not include the time
to convert the PDFs to text), which is 6.9 times longer. However, we need to

3 http://pdfbox.apache.org/
4 http://www.foolabs.com/xpdf/download.html

emphasize that these numbers are only comparable to a limited extent. CiteSeer‟s
SVM extracts not only the title but also other header data such as the authors and
CiteSeer‟s SVM is written in C and Perl while SciPlore Xtract is written in Java.
5 Discussion & Summary
All three tests show significantly worse results than the often claimed close-to-100%
accuracies. Our tests showed (1) that style information such as font size is suitable in
many cases to extract titles from PDF files (in our experiment in 77.9%).
Surprisingly, our simple rule based heuristic performed better than a support vector
machine. However, it could be that with other text to PDF converters, better results
may be obtained by the SVM. CiteSeer states to use a commercial tool to convert
PDFs to text and recommends PDFBox and pdftotext only as secondary choice. Our
tests also showed (2) that runtime of the rule based heuristic was better (8:19 min)
than SVM (57:26). However, these numbers are only limitedly comparable due to
various reasons.
In next steps, we will analyze why many PDFs could not be converted (30.7%) and
in which cases the heuristics could not identify titles correctly. The rule based
heuristic also needs to be compared to other approaches such as Conditional Random
Fields and Hidden Markov Models. We also intend to take a closer look at the other
studies and investigate why they achieve accuracies of around 90%, while in our test
the SVM achieved significantly lower accuracies. In the long run, machine learning
algorithms probably should be combined with our rule based heuristic. We assume
that this will deliver the best results. It also needs to be investigated how different
approaches with different languages. Existing machine learning approaches mostly
are trained with English documents. It might be that our approach will outperform
machine learning approaches even more significantly with non-English documents as
style information is language-independent (at least for western languages).
Summarized, despite the issue that many PDFs could not be converted, the rule
based heuristic we introduced in this paper, delivers good results in extracting titles
from scientific PDFs (77.9% accuracy). Surprisingly, this simple rule based heuristic
performs better than a Support Vector Machine based approach.
Our dataset (PDFs, software, results) is available upon request so that other
researchers can evaluate our heuristics and do further research.
References
[1] H. Han, C.L. Giles, E. Manavoglu, H. Zha, Z. Zhang, and E.A. Fox. Automatic document
metadata extraction using support vector machines. In Proceedings of the 3rd ACM/IEEE-CS
Joint Conference on Digital libraries, pages 37–48. 2003.
[2] Y. Hu, H. Li, Y. Cao, L. Teng, D. Meyerzon, and Q. Zheng. Automatic extraction of titles
from general documents using machine learning. Information Processing and Management,
42(5):1276–1293, 2006.
[3] F. Peng and A. McCallum. Accurate Information extraction from research papers using
conditional random fields. Information Processing and Management, 42(4):963–979, 2006.