Layout-aware text extraction from full-text PDF of scientific articles
© Ramakrishnan et al.; licensee BioMed Central Ltd. 2012
Received: 24 April 2012
Accepted: 28 May 2012
Published: 28 May 2012
Skip to main content
© Ramakrishnan et al.; licensee BioMed Central Ltd. 2012
Received: 24 April 2012
Accepted: 28 May 2012
Published: 28 May 2012
The Portable Document Format (PDF) is the most commonly used file format for online scientific publications. The absence of effective means to extract text from these PDF files in a layout-aware manner presents a significant challenge for developers of biomedical text mining or biocuration informatics systems that use published literature as an information source. In this paper we introduce the ‘Layout-Aware PDF Text Extraction’ (LA-PDFText) system to facilitate accurate extraction of text from PDF files of research articles for use in text mining applications.
Our paper describes the construction and performance of an open source system that extracts text blocks from PDF-formatted full-text research articles and classifies them into logical units based on rules that characterize specific sections. The LA-PDFText system focuses only on the textual content of the research articles and is meant as a baseline for further experiments into more advanced extraction methods that handle multi-modal content, such as images and graphs. The system works in a three-stage process: (1) Detecting contiguous text blocks using spatial layout processing to locate and identify blocks of contiguous text, (2) Classifying text blocks into rhetorical categories using a rule-based method and (3) Stitching classified text blocks together in the correct order resulting in the extraction of text from section-wise grouped blocks. We show that our system can identify text blocks and classify them into rhetorical categories with Precision1 = 0.96% Recall = 0.89% and F1 = 0.91%. We also present an evaluation of the accuracy of the block detection algorithm used in step 2. Additionally, we have compared the accuracy of the text extracted by LA-PDFText to the text from the Open Access subset of PubMed Central. We then compared this accuracy with that of the text extracted by the PDF2Text system, 2commonly used to extract text from PDF. Finally, we discuss preliminary error analysis for our system and identify further areas of improvement.
LA-PDFText is an open-source tool for accurately extracting text from full-text scientific articles. The release of the system is available at http://code.google.com/p/lapdftext/.
The field of Biomedical Natural Language Processing (BioNLP) is maturing, with specific fields of software development in response to user requirements: e.g., links between databases and literature, better tool interactivity and integration and the development of high-quality NLP resources [1, 2]. NLP techniques such as Named Entity Recognition  and Semantic Relation Extraction  have been shown to be very useful to biologists studying protein-protein interactions  and Gene-Disease-Phenotype relations . Given the ubiquity of the ‘Portable Document Format’ (PDF) as a means of distributing scientific publications and since access to information in full-text documents is vital for developing effective text-mining applications , it is essential to the general BioNLP community that developers of such applications can extract the textual content from PDF files accurately with open-source tools. Many past biomedical text mining studies have used either the abstracts of scientific papers [8–11] or relatively small collections of full-text articles sampled from the Open Access subset of PubMed Central . It is likely that certain content of journals of interest in a particular task is not distributed as a part of the Open Access subset.
A long-standing promise of BioNLP has been to help accelerate the vital process of literature-based biocuration, where published information is carefully translated into the knowledge architecture of biomedical databases, using specific BioNLP tools [1, 8, 13]. The identification of all papers relevant to the specific database being populated can be considered as a document classification problem . Subsequent steps have been cast as Information Extraction (IE) problems that leverage context dependent features [14, 15]. A key consideration is that the well-crafted manual workflows, developed by expert curators in biomedical databases, typically use rules based on context and rhetorical structure-dependent clues found only in the full-text of an article. Thus, it is important for the developers of BioNLP applications to have access to an accurate representation of the full-text of papers derived from PDF files, see .
Our goal is to provide an open-source software mechanism for automated decomposition and conversion of PDF files of research articles into a simple text format that other NLP groups can easily incorporate into their toolsets. In the most widely used text extraction programs (e.g., Adobe Acrobat, Grahl PDF Annotator, IntraPDF, PDFTron and PDF2Text), the flow of the main narrative from a file may be broken in mid sentence by errors derived from the reading order of individual text blocks and interruptions such as the inclusion of figure captions, footnotes and headers.
The variation in styles and formats of research articles (even within a single journal) can cause errors in terms of the ordering and splicing of text between pages and blocks. Any software that performs such decomposition and extraction should be adaptable with minimal human effort to new styles and formats. Driven by these needs, our system focuses on providing an open source PDF-to-text conversion capability meeting the following requirements: (1) the extraction mechanism should be able to adapt to single-column, two-column or mixed single and double column layouts, (2) extracted text should be error-free and grouped according to specific section headings used in the paper and (3) formatting artifacts such as, headers, footers, figures, tables and floating boxes (used in author summaries) should not interrupt the narrative-flow within each section. Thus, we have developed a three-step approach for extracting text from PDF files. The first step is the identification of contiguous text blocks. The second step is the classification of these text blocks into rhetorical categories (such as ‘Introduction’, ‘Results’ and ‘Discussion’) using logical rules that are easy to generate as ‘decision tables’ in a spreadsheet. The third step utilizes the classification results to ‘stitch’ appropriate text blocks together for extracting the text, while ignoring blocks that contain formatting embellishments so as to minimize flow-disruption of the extracted text. Our system provides programmatic, open-source access to each one (or to all three) of these capabilities for individual files or large collections of files.
The first step in LA-PDFText is to identify contiguous text blocks. In addition to the frequently-used two-column and single-column formats, journals also often use a mixed format where the title, authors, affiliation and abstract span the entire page width (single-column format) while all other sections of the article use a two-column format. We have observed these changes in format by manually inspecting papers from all available issues of the journal Brain Research. We denote these periodic changes in formatting over the lifetime of a given journal as ‘epochs’.
Per page word block parameters symbols and their definitions
Word block height
Horizontal space between words
Vertical space between words
max i f (δw height )
Most Popular Word block height in a page i
max i f (δw horizontal )
Most Popular Horizontal space between word blocks in a page i
max i f (δw vertical )
Most Popular Vertical space between word blocks in a page i
ϕ EW = max i f (δw height ) + max i f (δw horizontal )
east–west word block expansion parameters in page
ϕ NS = max i f (δw height ) + max i f (δw vertical )
north–south word block expansion parameters in page
Figure 1 is an example of how the block detection algorithm decides which word blocks to coalesce. Examples of the parameters δwhorizontal, δwvertical and wheight are shown in Figure 1. The distributions of these parameter values are calculated for each page and the most popular values for these parameters are chosen from these distributions to calculate ϕEW and ϕNS. We intentionally do not use most popular word width since biomedical text uses many long words and the most popular word width will make ϕEW too large thereby making the block subsumption algorithm too greedy. Consider the words ‘Introduction’ in the section heading, the word ‘antimicrobial’ in the first line of the first column and the word ‘the’ in the third line of the second column in Figure 1. Each word-block (shown in red) is surrounded by an expanded bounding-box (shown using a blue dotted line). All word-blocks (shown in red) that intersect with this expanded bounding-box are treated as words blocks to be merged. The block merge procedure is a greedy algorithm and will combine a section heading, subheading and the sections content into a single block based on the ϕEW and ϕNS parameters.
The next phase of LA-PDFText is based on ‘DROOLS’, a business rule management system and an enhanced Rules Engine implementation, ReteOO, based on the Rete algorithm  tailored for the Java language distributed as part of the open-source JBoss Enterprise Platform (http://labs.jboss.com/portal/jbossrules/). DROOLS provides a way for the LA-PDFText user to declaratively specify characteristics of a text block that make it a part of a particular section in the paper. We include the rule files for two epochs within the PLoS Biology dataset in both the DROOLS format as well as Microsoft Excel (Additional files 12 and 3).
The final goal of LA-PDFText is to accurately extract the text of any given section(s) in the correct sequence. As an implementation of this capability the last component of the LA-PDFText iterates over the classified blocks and stitches the classified blocks together to produce contiguous sections along with section and sub-section headings appropriately demarcated. LA-PDFText provides mechanisms to output the text of these PDF as XML formatted using PubMed Central’s OpenAccess DTD.
We have evaluated the three steps of our system independently of each other. In the following sections we will present our evaluation methods for each of the three steps of LA-PDFText and their results.
Example scenarios describing conversion operations and their corresponding costs
Operation in Gold Standard Representation
Block is split
Split the gold standard block into the required number
Big block is subsuming n small blocks
Delete the involved blocks in the gold standard and add one big block
n + 1
n block are intersecting
Delete all the blocks in the gold standard, whose area is common with the intersecting blocks, in the system output
Number of blocks deleted from gold standard +1
Per-section Precision (P), Recall(R), and F1 scores for section classification
PDF2Text is a widely used approach to extract text from PDF files. However, it is unable to distinguish between formatting embellishments and the main narrative of a scientific article. PDF2Text treats the entire document as one string, introducing errors within individual sentences, at column breaks and page breaks. LA-PDFText classifies each text block and (provided the classification is accurate) stitches text blocks belonging to the same section together, in order to extract contiguous rhetorical sections of the input articles. We have compared the text extraction capabilities of both systems to evaluation step 3 of LA-PDFText. Although PDF2Text is a simpler tool to use, we evaluate LA-PDFText’s text extraction capability against that of PDF2Text to show the benefit of our three-stage approach to text extraction.
LA-PDFText is designed to be a baseline system as a precursor for further improvements to the block detection, classification and text extraction stages. In this section, we discuss the results of each stage of LA-PDFText presenting error analyses and identify proposed future improvements.
LA-PDFText’s block detection algorithm is fairly accurate (see Spatial Segmentation Score in Additional file 4: Tables S4, S5, S6 and S7). Over the PLoS Biology dataset, block detection results in alignment scores with mean (μ) = 9.5 and standard deviation (σ) = 5.7. The algorithm depends on the accuracy of JPedal at identifying word blocks. Although it is expected that using a commercial version of JPedal will reduce these scores and improve block detection, we want LA-PDFText to be available for use without the need for users to purchase the commercial version (although we may release a version of our systems that can also work with the commercial version of JPedal).
We have identified specific errors in the rules that were responsible for poor performing categories (Table 3). Within PLoS Biology, the classification recall for the section titled ‘Supporting Information’ is only 0.224 (Table 3). Close inspection of our dataset reveals that most supporting information sections contain figure legends, which belong to two categories namely ‘Figure Legends’ and ‘Supporting Information’. The system correctly classifies the blocks as figure legends but not as supporting information. Both the precision and recall of the section titled ‘References’ are 0.532 and 0.632 respectively (Table 3). We attribute the low score to the fact that the font used in tables in many papers is the same as that used in references. Since our baseline rule-set did not contain a rule to identify tables they get wrongly identified as references resulting in poor recall and precision.
The quality of text extraction is best determined by the usability of the text by downstream text mining applications. We have presented evaluations that show the ability of LA-PDFText to extract text with fewer flow interruptions than text extracted by PDF2Text. It should be noted that the evaluation of text extraction was done on full text of papers explicitly to contrast LA-PDFText with PDF2Text. LA-PDFText also provides the user with the additional capability to extract text on a per-section basis; a capability that PDF2Text does not support.
Since the introduction of Portable Document Format in 1993 and the widespread development of online journals in the late 1990s, many archival documents published earlier have been scanned and converted into PDF. Furthermore, the scientific community and publishers have adopted PDF as the de facto standard format for scientific communication. In this paper we therefore do not focus on the Optical Character Recognition (OCR) problem but instead assume that we are given PDF documents that include the text, fonts, images, and 2D vector graphics. We are primarily concerned with related work in development of PDF extraction systems that support BioNLP work in the academic community.
Discovering the logical structure of documents is a well-studied problem. However most past efforts were aimed a logical-structure discovery [19, 20] and not explicitly aimed at text extraction from PDF documents. Furthermore, these past efforts used OCR to produce images of document pages, which are then segmented and the segments are classified to discover logical structure. Summers et al. present a survey of methods for the document-logical-structure discovery problem . While some methods surveyed by the author perform joint segmentation and classification, other methods separate these steps into distinct phases. Certain methods use a multi-level form of bounding boxes as the basis of their joint segmentation and decision-tree based classification  for logical-structure discovery. All of the above methods are aimed at inducing some hierarchical representation of the document content from document images. The method presented in this paper uses bounding boxes as well but separates the segmentation and classification phases.
One recent effort aimed at recovering the logical structure of the scholarly articles using Nuance OmniPage 16 to identify bounding boxes of words . The bounding box information is represented in XML that includes markup indicating each line and paragraph within the input PDF. The words, lines and paragraph information along with font information of each word are used as features to train a Conditional Random Field (CRF)  model to classify each line into one of 23 predetermined classes corresponding to rhetorical categories. The method proposed in  relies on a commercial tool; a feature we seek to avoid here. The authors performed tests on two datasets: one comprising 40 scientific papers in the field of computer science and the other from their previous work comprising 211 Association of Computing Machinery (ACM) papers. We downloaded the second dataset3 and manually inspected the PDF documents. We observed that formatting across the 211 papers from ACM is fairly regular using a two-column format. In contrast, we have tested LA-PDFText on articles from the journal Brain Research spanning volumes 1 to 1155. Manually we have identified 10 significant formatting changes from 1966 to 2007. In order to deal with all articles within PubMed,4a PDF extraction system will have to deal with these formatting variations. The system developed by  also produces XML similar to the LA-PDFText system and can therefore produce text on per-section basis. Upon close inspection of their results, we observed that formatting embellishments interrupt the flow of text extracted by their system in much the same way as it is in PDF2Text’s results. We believe that this is due to the fact that their system does not use a rule-based classification of text blocks, and may not be flexible enough to incorporate this change without substantial effort in feature engineering and retraining.
PDF extraction was used in the Mouse Genome Informatics (MGI) system to generate text input for text-mining software in-situ. They used a collection of commercial software (IntraPDF, PDFTron and specifically ProMiner) to extract text from PDF files but did not describe the process or outcome in detail, making it difficult to compare with our current work. Another toolset of particular interest is the Utopia documents platform [24, 25]. Utopia uses PDF as the base framework for constructing an entire toolset within the familiar architecture of a paper. As a first step, the Utopia system performs the text extraction process with a high accuracy, but it does so directly within the rubric of the Utopia system. Our system is a library that provides low-level control of multiple components of the text extraction process and is designed specifically for use by other text mining developers.
LA-PDFText is built using non-commercial components, making it freely available under the LGPL license. We believe that it is a very useful tool for the BioNLP community owing to its flexibility and adaptability to a variety of journal formats with minimal rule-development effort. We plan to extend this work by extracting text and structure from tables , graphs, figures  and citations (C ). The systems framework is designed in a modular fashion and can incorporate different methods for block detection and block classification. LA-PDFText will be put to immediate use in the development of a variety of biocuration applications. The next version of LA-PDFText will output annotations in compliance with ontologies such as Annotation Ontology [29, 30] and ontologies about bibliographic records, citations, evidence and discourse relationships.
In addition to open-source software distribution of LA-PDFText, we also provide the data set that was used in the evaluation presented in this paper (see Additional file 4). During our evaluation process each phase of our systems three-stage process produces intermediate files meant specifically for use by developers to monitor performance. For instance, the block classification phase produces images each page showing color-coded word blocks grouped using chunk block bounding boxes. This has been an invaluable tool for debugging rule files used in the classification process. Further details about verifying our systems output are forthcoming at the project page listed below. Our code contains unit tests that show how to programmatically invoke our system in all its modes of operation. We invite the reader to download the data set from the location indicated in the supplemental file and reconstruct our evaluation.
Project name: LA-PDFText – Layout-Aware Text Extraction from Full-text PDF of Scientific Articles
Project home page: http://code.google.com/p/lapdftext/
Current Version: 1.7
Operating system: MacOSX 10.6.7, Linux and Windows XP
Programming language: Java 1.6
Other requirements: none.
License: GNU General Public License
1 Average taken over all class labels in the section classification task
This research is funded in part by:
· U.S. National Science Foundation under the SciKnowMine project5 (grant #0849977)
· NIGMS under the BioScholar project (NIGMS: RO1-GM083871)
· NIH under the NeuArt project (NIH: 1R01MH079068-01A2)
· BIRN project (U24 RR025736-01).
We wish to acknowledge Marcelo Tallis and Thomas Russ for the discussions regarding evaluations. We would also like to acknowledge the contributions of Mark Shirley in helping with the development of early proof-of-concept prototypes. The authors would like to specially thank Dr. Drashti Dave for her help in reviewing the manuscript.
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.