The GutenTag project is project aimed at assisting with computational analysis of the Project Gutenberg text corpus, particularly by those working in the digital humanities. For more information on the goals of the project, please visit the main website  or read the paper which introduced it .
The current version of GutenTag takes the form of a downloadable tool (though there is a web demo, and there is the possibility future iterations might be primarily or entirely web-based), and assumes that you already have a copy of the Project Gutenberg corpus (2010 image), which is available officially on the Project Gutenberg website , and unofficially here ; our version is smaller because we have removed files that are not useful for text analysis, but it is still a fairly large download of over 4 gigabytes. Once you have the corpus (and have unzipped it, if necessary) and can access it from within GutenTag, you shouldn't need to modify it yourself.
Our goal is to help the full range of users that might be interested in accessing GutenTag's functionality. There are thus three distinct ways to run GutenTag, depending on your situation (and the version you have downloaded):
- You can run an executable (for PCs; see download page on the main website)
- You can run the python script
- You import GutenTag into other python code and access its functionality through an API.
The first two are functionally equivalent in the sense that both access the GutenTag web interface. The main difference is that the python script requires you have Python (2.7)  installed, whereas the exectuable is standalone and does not require any additional software. If you are accessing through an API, you can mostly skip the next section, though even API users may use the web interface for the purposes of creating configuration files which access specific subcorpora or do XML (TEI) tagging.
When you run GutenTag (either by clicking on it or running it from the command line), a brower window should open automatically. The first time you run GutenTag, it will prompt you for the location of the directory which contains the Project Gutenberg corpus: on a GUI like Windows or Mac OS X, you can usually grab the exact full path from a file finder/explorer window opened to that directory using copy and paste. (If you download the verion of the main Project Gutenberg from our website, and place it (after unzipping) in a directory parallel to (in the same folder as) the GutenTag directory, the path you need to enter will simply be "../PGUS/" [without quotation marks]). Click check to see if you've done this right, at which point it will prompt you to refresh your browser.
The main page of the interface allows you to configure everything about a particular run of GutenTag. By clicking on the first button on the screen, you can load an existing (previously saved) parameter file from a list, which will then populate the HTML form with those previous settings. At the bottom of the screen, there is a corresponding button that allows you to enter a filename to save your parameters; these will be available when you start your next run of GutenTag. If you wish, you can also run GutenTag from the command line with the -c command (configure), which will allow you to save the parameter file without actually runing GutenTag. The saved parameter files can be found in the saved_parameters subdirectory.
The large box on the screen allows you to configure a subcorpus that you're interested in, focusing on particular types of texts, time periods, authors, etc. Note that you must select one or more genres, or you will have no output whatsoever. Most of the other opinions limit the range of texts; if you leave them blank, there will be no restrictions on your texts. Multiple restrictions within a subcorpus mean that a text must satisfy all of them: if you select female authors and then select a birth range between 1800 and 1900, you will get texts written by female authors who were also born between 1800 and 1900. Some of the options available are marked with an asterisk (*), which means that some (in some cases, many) texts are not marked for these aspects and/or that the information was obtained automatically from the texts or Wikipedia and might contain errors. If you select one the restrictions for which the annotation is incomplete, the search will excude all texts for which there is no information about that aspect.
For texts and authors, you can create a predefined list which is accessible through the interface. In order to define such a list, create a new text file in the "user_lists" subdirectory, with one item on each line. Note that what you enter must exactly match the author/title in the PG database. If you want to check to see exactly what name is used for a particular author/text, you can browse the full lists of each attribute in the "full_lists" subdirectory. There are two "user" lists included with GutenTag: a list of authors consider by the Norton Anthology of English Literature as "canonical", and a list of texts which were focused on in the development of GutenTag.
The "Within-Text" filter allows you to exclude particular parts of the texts that meet your search criteria. The most common use of this will be to exclude material outside the main body of the text (that is, exclude front matter such as editorial introductions and coverpages, back matter such as indices, appendices, and sections of endnotes), though you can select specific sections to include or delete. For instance, for fiction you can include only narration or only character speech, and for plays you can include only character speech or only stage directions, etc. The default setting is to include everything. If you want to select a specific set of tags outside of the preset options, first select one of the presets from the drop down menu and then modify them manually by checking or unchecking the box.
There are two major modes in GutenTag: Export and Analyze. Export creates a new text corpus designed to your specifications; Analyze gets statistics about the tags in your subcorpus directly, without the need to save anything. Both modes allow you to include additional lexical tags which are either displayed in the XML (if in TEI mode; see below) or counted. GutenTag includes three major lexical resources: a subset of the tags from the General Inquirer , the ratings from the MRC psycholinguistic database , and a "6-style" lexical model created automatically from the Project Gutenberg corpus using automated methods [9,10], as well as sentiment polarity ratings . Users can also define their lexical tags by putting files into the "user_lexicons" directory, which will be accesible through the interface: an example with words refering to English nobility is included. The format for this is a text file with one word or phrase per line or, optionally, with a tab delimited value for each entry. In Analyze mode, the result for discrete tags will be the number of tags divided by the total number of tokens across all texs, while for tags with values the output is the sum of all the values divided by the number of tokens in each subcorpus. A second major option in the Analyze mode is to choose one of the various included textual measures: the output will be the average of the measure across the applicable texts. The detailed output of any analysis run of GutenTag (including results for individual texts) can be saved to a CSV if desired.
You can choose multiple subcorpora in a single run by clicking the Add Subcorpora button under the first subcorpus box. There are two uses for this: first, you can use this to create a single corpus defined in incompatible ways (for instance, a corpus with female authors from the 19th century and male authors from the 18th). In Analyze mode, often you want to be able to compare results directly across subcorpora; this provides an easy way to do that. However, anything that could be accomplished with multiple subcorpora can also be accomplished by running GutenTag multiple times.
In export mode, the most important option is whether to output in plain text mode (just tokens) or in TEI XML, which displays all the structural tagging for the text as well as any additional lexical tags that have been selected. In both modes, you can choose to include Part of Speech tags in the output, and also use the original or lemmatized version of the word. When exporting, you must give the name of the output directory which will be created and filled with files corresponding to individual texts which match the parameters. For both modes, you can choose to limit each subcorpus to a certain number of texts. With respect to the order in which these texts are chosen, you may decide on Random (which will create different results each time, but will not show ordering effects within the PG) or Fixed (which will have the same results each time).
The Named Entity Recognition module, which tags names corresponding to (individual) people and places, is accessible under Lexical Tags when exporting in TEI mode; we have published a separate paper discussion how it was created . The NER module is intended only for prose fiction; tagging persons will cause a rule-based clustering of characters, a rule-based assignment of said tags to those characters, and the insertion of a automatically-generated character list (including all characters which are assigned speech) at the beginning of the TEI file.
When you have entered all your parameters for a run of GutenTag, hit the button on the bottom of the screen corresponding to your desired mode (Export or Analyze) and GutenTag will begin to scan the corpus. The results screen has a progress bar tracking how far it is through the corpus (or, if you have restricted the number of texts, the percentage of required texts that have been retrieved), and the list of texts included so far will be updated as it goes along. This can take a long time, particularly if the number of texts which match your input parameters is large and you have included lexical tags involving large lexicons. Be patient and go do something else. When the run is over, the progress bar will disappear and the results page will be scrollable. If you want to do another run of GutenTag, you can click on the GutenTag icon at the top of the page, which will load a fresh (blank) version of the main interface.
To access the API, copy the GutenTag.py script as well as the resources directory into the directory with your other python code. GutenTag is not (yet) a Python package, so you need to preserve the basic directory structure. You can then "import GutenTag" in your code, and create a GT_API object.
gt = GutenTag.GT_API(corpus_path, parameter_path)
The corpus_path is the path to the Project Gutenberg corpus. In order to use the API, you need to have an existing parameter file which includes all your settings for GutenTag (this is preferable to having an extrodinarily long argument list). We provide a default one in the saved_parameters directory called "allEnglish" which would allow you to iterate through all the English texts in the corpus (if you only want a small subset, though, we recommend you create your own parameter file, it will be signfiicantly faster). You can create your own by modifying an existing file (it is in human-readable JSON format) or using the main GutenTag HTML interface with the -c command (see explanation above)
There are three methods available through GT_API:
This is an iterator function which will return pairs (2-ples) of GutenTag internal Text objects and corresponding information dictionaries for texts that match the given parameters. A Text object contains a set of ordered set of "tokens" (strings), and a set of "tags" in no particular order. Tags include a index "start" which indicates the starting token of the tag, an "end" index which indicates the end of the tag, the "tag" (the xml tag), a dictionary of "attributes" (key/value pairs), and a some other attributes which control how they are displayed in TEI output (in general, these shouldn't be modified). In general, the tags will encode structural and (if desired) lexical information about the text. The "info" dictionary includes basic information about the text, including author, title, genre, etc. One thing to note: this function makes use of the Python multiprocessing module, and if you are running this on a Windows machine, you must call cycle_through_texts within an if name == "main" block to prevent new processes from spawning their own new processes which spawn their own process, etc.
This iterator function cycles through the info dictionaries for all the texts in the corpus It is much quicker because it does not have to analyze the texts, and is useful for investigating basic facts about the corpus.
If the user wishes to output a text/info pair from cycle_through_texts() into the TEI (XML) format, this function can be called. A string corresponding to a TEI version of the text will be returned.
GutenTag is a work in progress, and will be for the foreseeable future. We welcome bug reports, requests for particular features, suggestions for or implementations of relevant (i.e. potentially useful for literary analysis) tagging algorithms, and general feedback. Please e-mail email@example.com.
- Julian Brooke, Adam Hammond, and Graeme Hirst. 2015. GutenTag: An NLP-driven tool for digital humanities research in the Project Gutenberg corpus. In Proceedings of the 4nd Workshop on Computational Literature for Literature (CLfL ’15).
- Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural Language Processing with Python. O’Reilly Media Inc.
- Philip J. Stone, Dexter C. Dunphy, Marshall S. Smith, and Daniel M. Ogilivie. 1966. The General Inquirer: A Computer Approach to Content Analysis. MIT Press.
- Max Coltheart. 1980. MRC Psycholinguistic Database User Manual: Version 1. Birkbeck College
- Julian Brooke and Graeme Hirst. 2013. Hybrid models for lexical acquisition of correlated styles. In Proceedings of the 6th International Joint Conference on Natural Language Processing (IJCNLP ’13).
- Julian Brooke and Graeme Hirst. 2014. Supervised ranking of co-occurrence profiles for acquisition of continuous lexical attributes. In Proceedings of The 25th International Conference on Computational Linguistics (COLING 2014).
- Julian Brooke, Adam Hammond, and Tim Baldwin. 2016. Bootstrapped text-level named entity recognition for literature. In Proceedings of ACL '16.