SPRUCE Mashup: Batch File Identification using Apache Tika

My last post discussed the benefits of collabaration, centred around a SCAPE hackathon. I argued that, in general, it was the collaborative, collocated nature of the developers working together that made demo development quicker; more people staring at the same problem results in multiple and varied viewpoints, ideas, and solutions. Developers can easily and quickly learn from one another, sharing information in an ad-hoc manner, and avoid reinventing the wheel. This communication is important and needs to be encouraged, but it needs to include practioners as well – they are, after all, the target audience for the tools developed. This communication and collaboration is exactly what the SPRUCE project is trying to do.

Spruce Mashup

Over the last 3 days I attended the first SPRUCE Digital Preservation Mashup in Glasgow; a mix of practioners and techies thrown together to discuss digital content management and preservation, identify real-world challenges and prototype solutions. Everything is free (you just have to get there and commit to the 3 days), so you don’t have to worry about anything other than digital preservation; as a techie this is especially useful when, on the second day, you realise a few extra hours coding will make all the difference and you couldn’t possibly go out for dinner – instead, food is brought to you!

The agenda was very well organised, starting with the usual lightening talks. In particular, practioners highlighted the sample data set they brought with them and the challenges they have with them, whereas developers discussed their background and digital preservation interests. Devs were then paired with practioners based on matches in challenges and interests.  As there were slightly more practioners than techies, I was paired with 3 practioners, Rebecca Nielson from the Bodleian Library, Hannah Green from Seven Stories and Richard Freeston from the University of Sheffield, who all had similar challenges of identifying content within their collections.

Scenario and Requirements

Initial brainstorms with this sub-group generated a lot of discussion about their collections, and in particular what challenges they faced with them.  The common theme that seemed to arise was the challenge in manually working out exactly what content they had in their collections. This was hindered by access issues, such as content contained in ISO files, and problems such as strangely named file extensions (.doc’s renamed as .tree). One particular directory in a sample set had a number of oddly named files which had manually (through a lot of hard work of trying various applications to open the files with!) been determined to be Photoshop files. DROID apparently had problems identifying these files, so I was keeping a close eye on how well Tika performed in identifying them!

There was also an interest in gathering additional metadata about the files, content authors, creation dates, etc., and summarising this information. Relatedly, being able to pull out keywords to summarise the content of a document was also of interest, but not considered the priority. 

So with these requirements in mind…

Let the Hacking Begin!

Knowing the promising results shown by Apache Tika™ in file identification, and having good development experience with it, I chose this tool to develop a prototype solution with.

I broke the problem down into several chunks/steps:

  1. Batch identification and metadata extraction of a directory of files using Tika
  2. Aggregation of identification/metadata information into a CSV file (for importing into Excel)
  3. Summarisation of the aggregated CSV file, e.g. #’s of each type of file format
  4. Automated mounting of ISO files on a Windows platform to enable the above steps to operate on ISO contents
  5. Extracting the top N highest frequency words from each text document (for semantically classifying documents)

I knew it may be tough to get through everything on that list in the alloted time, but it’s good to have a plan at least. It’s also worth mentioning that although Tika is Java based, for speed and simplicity I chose to script everything in Python. Python is more than capable of instantiating a Java program, so this wasn’t really an issue.

Batch processing of the files in a directory was reasonably trivial. Just a simple routine to walk a user-specified directory, pull out all file paths, and run Tika over each. To keep the solution modular, I ended up creating a user-specified output directory which contained one JSON formatted output file per input file (also maintaining the same sub-directory structure as the input directory). The output file was simply the output supplied by Tika (it has an option to return metadata in JSON format).

Next was to run through the output directory, reading in each output file and aggregating all the information into one CSV. Again, reasonably straightforward, although it did require some fiddling to make sure the file path specified in the CSV reflected the actual input file (rather than the output file). I’d initially just picked a subset of metadata information to return, creation dates, authors, application, number pages, word count, etc., but after showing it to the practioners, agreement was that it would be useful to output everything possible. This goes to highlight that these tools are being developed for practioners to use, and their input is vital to the development process in order to provide them with the tools they need!

The following table gives an idea of the aggregated output generated.  These results have been anonymised. The number and variety of headings is much larger than shown here, and depends on the types of files being parsed, for example image files often present data on width and height, emails give subject, from and to fields.

Filename Content-Length Application Author title Last-Author Creation-Date Page-Count Revision-Number Last-Save-Date Last-Printed Content-Type
C:\SPRUCE\input\file1.DOC 295424 Microsoft Word 6.0  Author A.  title A  Author A.  1997-09-28T21:56:00Z  74  27  1999-08-27T17:05:00Z  1998-02-12T18:31:00Z  application/msword 
C:\SPRUCE\input\file2.doc 297472 Microsoft Word 6.0 Author A. some text Author A.


73  1997-11-26T17:28:00Z  1601-01-01T00:00:00Z  application/msword
C:\SPRUCE\input\file3 12544                   text/plain
C:\SPRUCE\input\file4 11392                   application/octet-stream
C:\SPRUCE\input\prob_ps1                     image/vnd.adobe.photoshop
C:\SPRUCE\input\prob_ps12.psd                     image/vnd.adobe.photoshop

The final step, which I started somewhere around 9/9.30pm on the penultimate day, was to summarise all those results into a small summary CSV, outputing the number of files per format type, the creation date ranges, and contributing authors. This summary list was based on a practioner’s requirements, but it wouldn’t be challenging to adjust it to summarise other information.

That was it, pretty much. A modular solution resulting in three python scripts for automated batch file identification, metadata aggregation and summarising.  There was no time to consider keyword extraction, although through talking to other techies I did get some useful tool suggestions to look into (Apache Solr and elasticsearch). Nor was there really any time to focus on accessing the ISO images, although I did manage to find a bit of time after all the presentations on Wednesday to find a tool (WinCDEmu) which had a command line interface to mount an ISO file to a drive letter (enabling automated ISO access on Windows); thankfully my scripts seem to work fine using this mounted drive letter.


I was particularly interested in how well Tika would perform on identifying the problematic Photoshop files. I’m pleased to say it managed to get them all right, indicating them as image/vnd.adobe.photoshop formats.

Overall, for the sample set I tested on (primarily word documents), it was taking just over 1 second to evaluate each file on an old Dell Latitude laptop sporting a Core 2 Duo 1.8GHz processor with 1GB RAM; and roughly 4 minutes to complete the sample as a whole. As such I modified the script to provide an indication of expected duration to the user. Running over an CD ISO file took 30-40 minutes to complete.

Aggregating the results and summarising was extremely quick by comparison, taking mere seconds for the original sample, and slightly longer for the CD ISO.

Problems Encountered and Next Steps

A few notable problems were encountered during development, and investigating workarounds exhausted some of the development time.

  • Some input files caused Tika to crash during parsing.
    • This resulted in no output from Tika at all (not even identification information)
    • Workaround was to reuse a Tika API wrapper, with slight modification, to enable a 2 phase identification approach; the first phase tries to run Tika normally, if that fails, it uses the wrapper just to do identification.
    • Needs thorough investigation to work out why Tika crashes.
  • Output from some files could not be processed.
    • This seems to relate to the character encoding used in the file and returned by Tika.
    • Some approaches were tried to solve this, but no adequate solution was found.
    • Currently, some lines in the aggregated CSV are empty (except for filename), although the metadata itself should exist in the JSON output files.
    • Again, needs investigation to work out a good solution.
  • Some files were only identified as application/octet-stream
    • This is the Tika default when it doesn’t know what the file is
    • These files need further investigation as to why they’re not identified

Beyond these problems, another area for improvement would be performance.  A command line call to run Tika is made to evaluate every file, suffering a JVM initialisation performance hit every time.  Perhaps translating the tool to Java and making use of the Tika API wrapper would be a better approach (single JVM instantiation) as well as creating a more consolidated tool (that only depends on Java). Another approach would be parallelisation, making use of multi-core processors to evaluate multiple files at the same time.


It was fantastic to get a chance to talk with practioners, find out the real-world challenges they face, and help develop practical solutions for them. In particular it was useful to be able to go back to them after only a few hours of development, show them the progress, get their feedback, come up with new ideas, and really focus the tool on something they need. Without their scenarios and feedback, tools which we develop could easily miss the mark, having no real-world value.

At the same time, through this development, I have found problems that will feed back into the work I am doing on the SCAPE project. In particular, the test set I operated on highlights some robustness issues in Tika that need addressing (parsing crashes and output formatting), and some areas where its detection capabilities could be improved (application/octet-stream results). Solving these problems will improve Tika and ultimately increase robustness and performance of the tool I created here.

Ultimately, attending this event has been a win-win situation! Practioners have got prototypes of useful tools and, from my perspective at least, I have new insights into areas of improvement for SCAPE project tools based on real-world scenarios and data sets. As such, this event has proved invaluable, and I would encourage anyone with an interest in digital preservation to attend.

Hopefully I’ll see you at the next mashup!

By Peter May, posted in Peter May's Blog

19th Apr 2012  10:48 AM  16228 Reads  No comments


There are no comments on this post.

Leave a comment