Document layout analysis

How to cite: Page, R. (2023). Document layout analysis. https://doi.org/10.59350/z574z-dcw92

Some notes to self on document layout analysis.

I’m revisiting the problem of taking a PDF or a scanned document and determining its structure (for example, where is the title, abstract, bibliography, where are the figures and their captions, etc.). There are lots of papers on this topic, and lots of tools. I want something that I can use to process both born-digital PDFs and scanned documents, such as the ABBYY, DjVu and hOCR files on the Internet Archive. PDFs remain the dominant vehicle for publishing taxonomic papers, and aren’t going away any time soon (see Pettifer et al. for a nuanced discussion of PDFs).

There are at least three approaches to document layout analysis.

Rule-based

The simplest approach is to come up rules, such as “if the text is large and it’s on the first page, it’s the title of the article”. Examples of more sophisticated rules are given in Klampfl et al., Ramakrishnan et al., and Lin. Rule-based methods can get you a long way, as shown by projects such as Plazi. But there are always exceptions to rules, and so the rules need constant tweaking. At some point it makes sense to consider probabilistic methods that allow for uncertainty, and which can also “learn”.

Large language models (LLMs)

At the other extreme are Large language models (LLMs), which have got a lot of publicity lately. There are a number of tools that use LLMs to help extract information from documents, such as LayoutLM (Xu et al.), Layout Parser, and VILA (Shen et al.). These approaches encode information about a document (in some case including the (x,y) coordinates of individual words on a page) and try and infer which category each word (or block of text) belongs to. These methods are typically coded in Python, and come with various tools to display regions on pages. I’ve had variable success getting these tools to work (I am new to Python, and am also working on a recent Mac which is not the most widely used hardware for machine learning). I have got other ML tools to work, such as an Inception-based model to classify images (see Adventures in machine learning: iNaturalist, DNA barcodes, and Lepidoptera), but I’ve not succeeded in training these models. There are obscure Python error messages, some involving Hugging Face, and eventually my patience wore out.

Another aspect of these methods is that they often package everything together, such that they take a PDF, use OCR or ML methods such as Detectron to locate blocks, then encode the results and feed them to a model. This is great, but I don’t necessarily want the whole package, I want just some parts of it. Nor does the prospect of lengthy training appeal (even if I could get it to work properly).

The approach that appealed the most is VILA, which doesn’t use (x,y) coordinates directly but instead encodes information about “blocks” into text extracted from a PDF, then uses an LLM to infer document structure. There is a simple demo at Hugging Face. After some experimentation with the code, I’ve ended up using the way VILA represents a document (a JSON file with a series of pages, each with lists of words, their positions, and information on lines, blocks, etc.) as the format for my experiments. If nothing else this means that if I go back to trying to train these models I will have data already prepared in an appropriate format. I’ve also decided to follow VILA’s scheme for labelling words and blocks in a document:

• Title
• Author
• Abstract
• Keywords
• Section
• Paragraph
• List
• Bibliography
• Equation
• Algorithm
• Figure
• Table
• Caption
• Header
• Footer
• Footnote

I’ve tweaked this slightly by adding two additional tags
from VILA’s Labeling Category Reference, the “semantic” tags “Affiliation” and “Venue”. This helps separate information on author names (“Author”) from their affiliations, which can appear in very different positions to the author’s names. “Venue” is useful to label things such as a banner at the top of an article where the publisher display the name of the journal, etc.

Conditional random fields

In between masses of regular expressions and large language models are approaches such as Conditional random fields (CRFs), which I’ve used before to parse citations (see Citation parsing tool released). Well known tools such as GROBID use this approach.

CRFs are fast, and somewhat comprehensible. But it does require Feature engineering, that is, you need to come up with features of the data to help train the model (for the systematists among you, this is very like coming up with characters for a bunch of taxa). This is were you can reuse the rules developed in a rules-based approach, but instead of having the rules make decisions (e.g., “big text = Title”), you just a rule that detects whether text is big or not, and the model combined with training data then figures out if and when big text means “Title”. So you end up spending time trying to figure out how to represent document structure, and what features help the model get the right answer. For example, methods such as Lin’s for detecting whether there are recurring elements in a document are great source of features to help recognise headers and footers. CRFs also make it straightforward to include dependencies (the “conditional” in the name). For example, a bibliography in a paper can be recognised not just by a line having a year in it (e.g., “2020”), but there being nearby lines that also have years in them. This helps us avoid labelling isolated lines with years as “Bibliography” when they are simply text in a paragraph that mentions a year.

Compared to LLMs this a lot of work. In principle with an LLM you “just” take a lot of training data (e.g., text and location on a page) and let the model to the hard work of figuring out which bit of the document corresponds to which category (e.g., title, abstract, paragraph, bibliography). The underlying model has already been trained on (potentially) vast amounts of text (and sometimes also word coordinates). But on the plus side, training CRFs is very quick, and hence you can experiment with adding or removing features, adding training data, etc. For example, I’ve started training with about ten (10) documents, training takes seconds, and I’ve got serviceable results.

Lots of room for improvement, but there’s a constant feedback loop of seeing improvements, and thinking about how to tweak the features. It also encourages me to think about what went wrong.

Problems with PDF parsing

To process PDFs, especially “born digital” PDFs I rely on pdf2xml, originally written by Hervé Déjean (Xerox Research Centre Europe). It works really well, but I’ve encountered a few issues. Some can be fixed by adding more fonts to my laptop (from XpdfReader), but others are more subtle.

The algorithm used to assign words to “blocks” (e.g., paragraphs) seems to struggle with superscripts (e.g., ¹), which often end up being treated as separate blocks. This breaks up lines of text, and also makes it harder to accurately label parts of the document such as “Author” or “Affiliation”.

Figures can also be problematic. Many are simply bitmaps embedded in a PDF and can be easily extracted, but sometimes labelling on those bitmaps, or indeed big chunks of vector diagrams are treated as text, so we end up with story text blocks in odd positions. I need to spend a little time thinking about this as well. I also need to understand the “vet” format pdftoxml extracts from PDFs.

PDFs also have all sorts of quirks, such as publishers slapping cover pages on the front, which make feature engineering hard (the biggest text might now be not be the title but some cruff from the publisher). Sometimes there are clues in the PDF that it has been moodier.! For example, ResearchGate inserts a “rgid” tag in the PDF when it adds a cover page.

Yes but why?

So, why I am doing this? Why battle with the much maligned PDF format. It’s because a huge chunk of taxonomic and other information is locked up in PDFs, and I’d like a simpler, scalable, way to extract some of that. Plazi is obviously the leader in this are in terms of the amount of information they have extracted, but their approach is labour-intensive. I want something that is essentially automatic, that can be trained to handle the idiosyncracities of the taxonomic literature, and can be applied to both born digital PDFs and OCR from scans in the Biodiversity Heritage Library and elsewhere. Even if we could simply extract bibliographic information (to flesh out the citation graph) and the figures, that would be progress.

References

Déjean H, Meunier J-L (2006) A System for Converting PDF Documents into Structured XML Format. In: Bunke H, Spitz AL (eds) Document Analysis Systems VII. Springer, Berlin, Heidelberg, pp 129–140 https://doi.org/10.1007/11669487_12

Klampfl S, Granitzer M, Jack K, Kern R (2014) Unsupervised document structure analysis of digital scientific articles. Int J Digit Libr 14(3):83–99. https://doi.org/10.1007/s00799-014-0115-1

Lin X (2003) Header and footer extraction by page association. In: Document Recognition and Retrieval X. SPIE, pp 164–171 https://doi.org/10.1117/12.472833

Pettifer S, McDERMOTT P, Marsh J, Thorne D, Villeger A, Attwood TK (2011) Ceci n’est pas un hamburger: modelling and representing the scholarly article. Learned Publishing 24(3):207–220. https://doi.org/10.1087/20110309

Ramakrishnan C, Patnia A, Hovy E, Burns GA (2012) Layout-aware text extraction from full-text PDF of scientific articles. Source Code for Biology and Medicine 7(1):7. https://doi.org/10.1186/1751-0473-7-7

Shen Z, Lo K, Wang LL, Kuehl B, Weld DS, Downey D (2022) VILA: Improving Structured Content Extraction from Scientific PDFs Using Visual Layout Groups. Transactions of the Association for Computational Linguistics 10:376–392. https://doi.org/10.1162/tacl_a_00466

Xu Y, Li M, Cui L, Huang S, Wei F, Zhou M (2020) LayoutLM: Pre-training of Text and Layout for Document Image Understanding. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. pp 1192–1200

Written with StackEdit.

Using the hypothes.is API to annotate PDFs

Sometimes it's the small victories. Using the @hypothes_is API to annotate a scientific name in the PDF for https://t.co/WU9IAAztE5 it shows up when the PDF is viewed online https://t.co/bMDyqDL8np , when the PDF file is viewed on a local computer in Chrome...

— Roderic Page (@rdmpage) September 22, 2020

With somewhat depressing regularity I keep cycling back to things I was working on earlier but never quite get to work the way I wanted. The last couple of days it's the turn of hypothes.is.

One of the things I'd like to have is a database of all taxonomic names such that if you clicked on a name you would get not only the bibliographic record for the publication where that name first appeared (which is what I've bene building for animals in BioNames) but also you could see the actual publication with the name highlighted in the text. This assumes that the publication has been digitised (say, as a PDF) and is accessible, but let's assume that this is the case. Now, we could do this manually, but we have tools to find taxonomic names in text. And in my use case I often know which page the name is on, and what the name is, so all I really want is to be able to highlight it programmatically (because I have millions of names to deal with).

So, time to revisit the hypothes.is API. One of the neat "tricks" hypothes.is have managed is the ability to annotate, say, a web page for an article and have that annotation automagically appear on the PDF version of the same article. As described in How Hypothesis interacts with document metadata this is in part because hypothes.is extracts metadata from the article's web page, such as DOI and link to the PDF, and stores that with the annotation (I say "in part" because the other part of the trick is to be able to locate annotations in different versions of the same text). If you annotate a PDF, hypothes.is stores the URL of the PDF and also a "fingerprint" of the PDF (see PDF Fingerprinting for details). This means that you can also add an annotation to a PDF offline (for example, on a file you have downloaded onto your computer) and - if hypothes.is has already encountered this PDF - that annotation will appear in the PDF online.

What I want to do is have a PDF, highlight the scientific name, upload that annotation to hypothes.is so that the annotation is visible online when anyone opens the PDF (and ideally when they look at the web version of the same article). I want to do this programmatically. Long story short, this seems doable. Here is an example annotation that I created and sent to hypothesis.is via their API:

{
    "uri": "http://www.scielo.br/pdf/isz/v105n3/1678-4766-isz-105-03-00372.pdf",
    "document": {
        "highwire": {
            "doi": [
                "10.1590/1678-476620151053372375"
            ]
        },
        "dc": {
            "identifier": [
                "doi:10.1590/1678-476620151053372375"
            ]
        },
        "link": [
            {
                "href": "urn:x-pdf:6124e7bdb33241429158b11a1b2c4ba5"
            }
        ]
    },
    "tags": [
        "api"
    ],
    "target": [
        {
            "source": "http://www.scielo.br/pdf/isz/v105n3/1678-4766-isz-105-03-00372.pdf",
            "selector": [
                {
                    "type": "TextQuoteSelector",
                    "exact": "Alpaida venger sp. nov.",
                    "prefix": "imens preserved in 75% ethanol. ",
                    "suffix": " (Figs 1-9) Type-material. Holot"
                },
                {
                    "type": "TextPositionSelector",
                    "start": 4834,
                    "end": 4857
                }
            ]
        }
    ],
    "user": "acct:xxx@hypothes.is",
    "permissions": {
        "read": [
            "group:__world__"
        ],
        "update": [
            "acct:xxx@hypothes.is"
        ],
        "delete": [
            "acct:xxx@hypothes.is"
        ],
        "admin": [
            "acct:xxx@hypothes.is"
        ]
    }
}

In this example the article and the PDF are linked by including the DOI and PDF fingerprint in the same annotation (thinking about this I should probably also have included the PDF URL in document.highwire.pdf_url[]). I extracted the PDF fingerprint using mutool and added that as the urn:x-pdf identifier.

The actual annotation itself is described twice, once using character position (start and end of the text string relative to the cleaned text extracted from the PDF) and once by including short fragments of text before and after the bit I want to highlight (Alpaida venger sp. nov.). In my limited experience so far this combination seems to provide enough information for hypothes.is to also locate the annotation in the HTML version of the article (if one exists).

You can see the result for yourself using the hypothes.is proxy (https://via.hypothes.is). Here is the annotation on the PDF (https://www.scielo.br/pdf/isz/v105n3/1678-4766-isz-105-03-00372.pdf)

and here is the annotation on HTML (https://doi.org/10.1590/1678-476620151053372375)

If you download the PDF onto your computer and open the file in Chrome you can also see the annotation in the PDF (to do this you will need to install the hypothes.is extension for Chrome and click the hypothes.is symbol on your Chrome's toolbar).

In summary, we have a pretty straightforward way to automatically annotate papers offline using just the PDF.

Towards BioStor articles marked up using Journal Archiving Tag Set

A while ago I posted BHL to PDF workflow which was a sketch of a work flow to generate clean, searchable PDFs from Biodiversity Heritage Library (BHL) content:


I've made some progress on putting this together, as well as expanded the goal somewhat. In fact, there are several goals:

1. BioStor articles need to be archived somewhere. At the moment they live on my server, and metadata is also served by BHL (as the "parts" you see in a scanned volume). Long term maybe PubMed Central is a possibility (BHL essentially becomes a publisher). Imagine PubMed Central becoming the primary archival repository for biodiversity literature.
2. BioStor articles could be more useful if the OCR text was cleaned up and marked up (e.g., highlighting taxon names, localities, extracting citations, etc.).
3. If BioStor articles were marked up to same extent as ZooKeys then we could use tools developed for ZooKeys (see Towards an interactive taxonomic article: displaying an article from ZooKeys) for a richer reading experience.
4. Cleaned OCR text could also be used to generate searchable PDFs, which are still the most popular way for people to read articles (see Why do scientists tend to prefer PDF documents over HTML when reading scientific journals?). BioStor already generates PDFs, but these are simply made by wrapping page images in a PDF. Searchable PDFs would be much friendlier.

For BioStor articles to be archived in PubMed Central they would need to be marked up using the Journal Archiving and Interchange Tag Suite (formerly the NLM DTDs). This is the markup used by many publishers, and also the tag suite that TaxPub build upon.

The idea of having BioStor marked up in JATS is appealing, but on the face of it impossible because the all we have is page scans and some pretty ropey OCR. But because the NLM has also been heavily involed in scanning the historical literature they are used to dealing with scanned literature, and JATS can accommodate articles ranging from scans to fully marked up text. For example, take a look at the article "Microsporidian encephalitis of farmed Atlantic salmon (Salmo salar) in British Columbia" which is in PubMed Central (PMC1687123). PMC has basic metadata for the article, scans of the pages, and two images extracted from those pages. This is pretty much what BioStor already has (minus the extracted images).

With this in mind, I dusted off some old code, put it together and created an example of the first baby steps towards BioStor and JATS. The code is in github, and there is a live example here.


The example takes BioStor article 65706, converts the metadata to JATS, links in the page scans, and also extracts images from the page scans based on data in the ABBYY OCR files. I've also generated HTML from the DjVu files, and this HTML includes hOCR tags that embed information about the OCR text. This format can be edited by tools such as (see Jim Garrison's moz-hocr-edit discussed in Correcting OCR using hOCR in Firefox). This HTML can be processed to output a PDF that includes the page scans but also has the OCR text as "hidden text" so the reader can search for phrases, or copy and paste the text (try the PDF for article 65706).

I've put the HTML (and all the XML and images) in github, so one obvious model for working on an article is to put it into a github repository, push any edits made to the repository, then push that to a web server that displays the articles.

There are still a lot of rough edges, and I think we can buld nicer interfaces than moz-hocr-edit (e.g., using the "contenteditable" attribute in the HTML), althogh moz-hocr-edit has the nice feature of being able to save the edits straight back to the HTML file (saving edited HTML to disk is a non-trivial task in most web browsers). I also need to add the code for building the initial JATS file (currently this is hidden on the BioStor server). There are also issues about PDF quality. At the moment I output black and white PNGs, which look nice and clean but can mangle plates and photos. I need to tweak that aspect of the process.

One application of these tools would be to take a single journal and convert all the BioStor articles into JATS, then make it available for people to further clean and markup as needed. There is an extraordinary amount of information locked away in this literature, it would be nice if we made better use of that treasure trove.

BHL to PDF workflow

Just some random thoughts on creating searchable PDFs for article extracted from BHL.

Mapping names to literature: closing in on 250,000 names

Following on from my earlier post Linking taxonomic names to literature: beyond digitised 5×3 index cards I've been slowly updating my latest toy:

http://iphylo.org/~rpage/itaxon

This site displays a database mapping over 200,000 animal names to the primary literature, using a mix of identifiers (DOIs, Handles, PubMed, URLs) as well as links to freely available PDFs where they are available. Lots still to do as about a third of the 1.5 million names in the database have citations that my code hasn't been able to parse. There are also lots of gaps that need to be filled in, for example missing DOIs or PubMed identifiers, and a lot of the earlier names are linked by "microcitations" to names, and I'll need to handle those (using code from my earlier project Nomenclator Zoologicus meets Biodiversity Heritage Library: linking names directly to literature).

The mapping itself is stored in a database that I'm constantly editing, so this is far from production quality, but I've found it eye-opening just how much literature is available. There is a lot of scope for generating customised lists of papers, for example, primary taxonomic sources for taxa currently on the IUCN Red List, or those taxa which have sequences in GenBank (building on the mapping of NCBI taxa onto Wikipedia). Given that a lot of the relevant literature is in BHL, or available as PDFs, we could do some data mining, such as extracting geographical coordinates, taxonomic names, and citations. And if linked data is your thing, the 110,000 DOIs and nearly 9,000 CiNiii URLs all serve RDF (albeit not without a few problems).

I've set a "goal" of having 250,000 names mapped to the primary literature, at which point the database interface will get some much-needed attention, but for now have a look for your favourite animal and see if it's original description has been digitised.

Viewing scientific articles on the iPad: the PLoS Reader

Continuing on from my previous post Viewing scientific articles on the iPad: towards a universal article reader, here are some brief notes on the PLoS iPad app that I've previously been critical of.

There are two key things to note about this app. The first is that it uses the page turning metaphor. The article is displayed as a PDF, a page at a time, and the user swipes the page to turn it over. Hence, the app is simulating paper on the iPad screen.

But perhaps more interesting is that, unlike the Nature app discussed earlier, the PLoS app doesn't use a custom API to retrieve articles. Instead the app uses RSS feeds from the PLoS site. PLoS provides journal-specific RSS feeds, as well as subject-specific feeds within journals (see, for example, the PLoS ONE home page). The PLoS Reader app takes these feeds and uses them to create a list of articles the reader can choose from.

A nice feature of the PLoS ATOM feeds is the provision of links to alternative formats for the article (unlike many journal RSS feeds, which provide just a DOI or a URL). For example, the feed item for the article "Transmission of Single HIV-1 Genomes and Dynamics of Early Immune Escape Revealed by Ultra-Deep Sequencing" doi:10.1371/journal.pone.0012303 contains links to the PDF and XML versions of the article:

<link rel="related" 
   type="application/pdf" 
   href="http://www.plosone.org/article/fetchObjectAttachment.action?uri=info:doi/10.1371/journal.pone.0012303&representation=PDF" 
   title="(PDF) Transmission of Single HIV-1 Genomes and Dynamics of Early Immune Escape Revealed by Ultra-Deep Sequencing" />
<link rel="related" 
   type="text/xml" 
   href="http://www.plosone.org/article/fetchObjectAttachment.action?uri=info:doi/10.1371/journal.pone.0012303&representation=XML" 
   title="(XML) Transmission of Single HIV-1 Genomes and Dynamics of Early Immune Escape Revealed by Ultra-Deep Sequencing" />

This makes the task of an article reader much easier. Rather than attempt to screen scrape the article web page, or rely on a rule for constructing the link to the desired file, the feed provides an explicit URL to the different available formats.

I've not seen this feature in other journal RSS feeds, although article web pages sometimes provide this information. BMC journals, for example, provide <link rel="alternate"> tags in the web page for each article, from which we can extract links to the XML and PDF versions, and some journals (BMC included) provide the Google Scholar metadata data tag <meta name="citation_pdf_url"> to link to the PDF. Hence, a generic article reader will need to be able to extract metadata tags from article web pages as it seeks formats suitable to display.

BioStor gets PDFs with XMP metadata - bibliographies made easy with Mendeley and Papers

The ability to create PDFs for the articles BioStor extracts from the Biodiversity Heritage Library has been the single most requested feature for BioStor. I've taken a while to get around to this -- for a bunch of reasons -- but I've finally added it today. You can get a PDF of an article by either clicking on the PDF link on the page for an article, or by appending ".pdf" to the article URL (e.g., http://biostor.org/reference/570.pdf). In some ways the BioStor PDFs are pretty basic - they contain page images, not the OCR text, so they tend to be quite large and you can't search for text within the article. But what they do have is XMP metadata.

XMP metadata
One of the great bugbears about organising bibliographies is the lack of embedded metadata in PDFs, in other words Why can't I manage academic papers like MP3s? (see my earlier post for some background). Music files and digital photos contain embedded metadata that store information such as song title and artist in the case of music, or date, exposure, camera model, and GPS co-ordinates in the case of digital images. This means software (and webs sites such as Flickr) can automatically organise your collection of media based on this embedded metadata.

Wouldn't it be great if there was an equivalent for PDFs of papers, whereby the PDF contains all the relevant the bibliographic details (article title, authorship, journal, volume, pages, etc.), and reference managing software could read this and automatically put the PDF into whatever categories you chose (e.g., by author, journal, or date)? Well, at least two software programs can do this, namely the cross-platform Mendeley, and Papers, which supports Apple's Macintosh, iPhone, and iPad platforms. Both programs can read bibliographic metadata in Adobe's Extensible Metadata Platform (XMP), which has been adopted by journals such as Nature, and CrossRef has recently been experimenting in providing services to add XMP to PDFs.

One reason I put off adding PDFs to BioStor was the issue of simply generating dumb PDFs for which users would then have to retype the corresponding bibliographic metadata if they wanted to store the PDF in a reference manager. However, given that both Papers and Mendeley support XMP, you can simply drag the PDF on to either program and they will extract the details for you (including a list of up to 10 taxonomic names found in the article). Both Papers and Mendeley support the notion of a "watched folder" where you can dump PDFs and they will "automagically" appear in your reference manager's library. Hence, if you use either program you should be able to simply download PDFs from BioStor and add them to your library without having to retype anything at all.

Technical details
This post is being written as I'm waiting to catch a plane, so I haven't time to go into all the gory details. The basic tools I used to construct the PDfs were FPDF and ExifTool, which supports injecting XMP into PDFs (I couldn't find another free tool that could insert XMP into a PDF that didn't already have any XMP metadata). I store basic Dublin Core and PRISM metadata in the PDF. The ten most common taxonomic names found in the pages of the article are stored as subject tags.

Initially it appeared that only Papers could extract XMP, Mendeley failed completely (somewhat confirming my prejudices about Mendeley). However, I sent an example PDF to Mendeley support, and they helpfully diagnosed the problem. Because XMP metadata can't always be trusted, Mendeley compares title and author values in the XMP metadata with text on the first couple of pages of the PDF. If they match, then the program accepts the XMP metadata. Because my initial efforts as creating PDfs just contained the BHL page images and no text, they wouldn't pass Mendeley's tests. Hence, I added a cover page containing the basic bibliographic metadata for the article, and now Mendeley is happy (the program itself is growing on me, but if you're an Apple fanboy like me, Papers has native look and feel, and syncing your library with your iPhone is a killer feature). There are a few minor differences in how Papers and Mendeley handle tags. Papers will take text in the Dublin Core "Subject" tag and use those as keywords, whereas to get Mendeley to extract tags I had to store them using the "Keywords" tag (implemented using FPDF's SetKeywords function). But after a bit of fussing I think the BioStor PDFs should play nice in both programs.

XMP

Following on from the previous post, as Howison and Goodrum note, Adobe provides XMP as a way to store metadata in files, such as PDFs. XMP supports RDF and namespaces, which means widely used bibliographic standards such as Dublin Core and PRISM can be embedded in a PDF, so the article doesn't become separated from its metadata. Adobe provides a developers kit under a BSD license.

The vision of managing digital papers being as easy as managing digital music is really compelling. Imagine auto populating bibliographic management software simply by adding the PDF! Note also that the Creative Commons supports XMP.