Take a MALLET to disciplinary history
Clone or download
agoldst Merge branch 'prep-data-drop-blank-lines' of git://github.com/senderl…
…e/dfr-browser into senderle-prep-data-drop-blank-lines
Latest commit 49fc661 Feb 10, 2017

README.md

A simple topic-model browser

These files use d3 to provide a way to browse some of a topic model of texts in a web browser, relying entirely on static html and javascript files. It is specialized for models of texts from JSTOR's Data for Research service, though it can be (and has been) adapted for models of other corpora. For more information, see the main project page at agoldst.github.io/dfr-browser and the working demo, a browser for a 64-topic model of the journal PMLA.

This software is free to use, copy, modify, and distribute under the MIT license (which is to say, please credit me if you do use it). This project skews to my needs as a literary scholar and to my amateurishness as a programmer. No aspect of this code has been systematically tested. Sorry.

The rest of this file explains how to set up the browser to use with your own topic models.

Generate the model

Creating and saving topic models from R using MALLET

The browser looks for the model data in a series of files. The most convenient way to generate these is via my companion R package, dfrtopics, which supplies an export_browser_data function. You can use the package to create topic models and then export them directly---for details, see the introduction to dfrtopics. The dfrtopics package provides many functions for adjusting the model inputs and parameters, but, very briefly, here is how one might go from a DfR download unzipped in the folder dfr-data to a topic model browser:

library("dfrtopics")
m <- model_dfr_documents(
    citations_files="dfr-data/citations.tsv",
    wordcounts_dirs="dfr-data/wordcounts",
    n_topics=40
)
# optional but recommended: save model outputs
write_mallet_model(m, output_dir="model")
# create data files and copy over dfr-browser sources
export_browser_data(m, out_dir="browser", supporting_files=TRUE)

This will create a 40-topic model of all words in all the documents in dfr-data, and then create a browser folder holding all files necessary to browse the topic model (with supporting_files=TRUE all the other requisite dfr-browser HTML, JavaScript, and CSS files are copied over as well). See help(export_browser_data) in R for details on other parameters to export_browser_data.

If you have created a topic model with command-line mallet, then dfrtopics can still load and export the necessary files:

library("dfrtopics")
m <- load_from_mallet_state(
    mallet_state_file="mallet_state.gz",
    instances_file="docs.mallet",
    metadata_file="dfr-data/citations.tsv")
export_browser_data(m, out_dir="browser", supporting_files=TRUE)

In either case, when the export is done, change to the browser directory, run bin/server in the shell, then visit localhost://8888 in the web browser.

Browser data file specifications

In case you wish to edit the exported files or create them another way, here are the data files expected by this browser:

  • browser info (data/info.json): a text file giving a JSON object with title, meta_info, and optionally VIS and topic_labels members. dfrtopics::export_browser_data will create a stub info.json file for you to edit. meta_info is displayed as part of the "About" page. VIS controls many visualization parameters, and topic_labels gives hand labels for topics; see below.

  • word weights for the n most probable words in each topic (data/tw.json): a text file giving a JSON object with alpha, an array of estimated values for the hyperparameter alpha for each topic, and tw, an array of { words, weights } objects, one for each topic. words and weights are in turn arrays, in order, of the most prominent words in the topic and their weights respectively. The value of n is up to you.

  • document metadata (data/meta.csv.zip): a headerless zipped CSV of document metadata, with the following columns, quoted where required by RFC 4180:

    1. DOI
    2. title
    3. author(s) (separated by VIS.bib.author_delimiter, by default a tab)
    4. journal title
    5. volume
    6. issue
    7. publication date (ISO 8601)
    8. page range

    Additional columns after these are loaded with the default names X1, X2, and so on, and ignored by the display. (The names can be changed with the VIS.metadata.spec.extra_fields array in info.json.) If your documents are not journal articles, this format may not quite fit, and some modification of the metadata parsing or document-display code may be necessary. See "Adapting this project" below.

  • the topic weights for each document (data/dt.json.zip): a zipped text file giving a JSON object specifying the document-topic matrix in sparse compressed-column format in terms of three arrays i, p, and x. specified as follows. The documents d with non-zero weights for topic t are given by i[p[t]], i[p[t] + 1], ... i[p[t + 1] - 1], and the corresponding topic weights for those documents are given by x[p[t]], x[p[t] + 1], ..., x[p[t + 1] - 1]. Where possible, I suggest using unnormalized, unsmoothed topic weights (from a final Gibbs sampling state). "Words in document d assigned to topic t" is a bit more intuitive, if slightly less proper, than "posterior probability of topic t in d." Normalized (fractional) weights will also work; dfr-browser checks to see if the rows of the document-topic matrix all sum to one, and adjusts some of its display labels accordingly.

  • two-dimensional coordinates for each topic (data/topic_scaled.csv): a headerless two-column CSV. Row t gives the position of topic t in some space. dfrtopics generates this with its topic_scaled_2d function (q.v.). This file is optional; if it is available, the browser can draw the "scaled" overview plot, but if not, the rest of the browser will still work.

Except for data/info.json, which is hard-coded, all of the filenames are configurable. They are properties of the VIS.files object and can be modified by adding a files property to VIS in info.json. The configurable file names can be either local paths or URLs (they are passed straight on to d3.xhr). The property names for the last four listed above are dt, tw, meta, and topic_scaled. If a given filename ends in .zip, the browser uses JSZip to unzip the file. It is normally not worth compressing tw.json or topic_scaled.csv.

Preparing data files entirely on the command line

If you prefer not to use dfrtopics, the supplied prepare-data python script can produce the necessary files from model data in multiple formats---except for the scaled topic coordinates. The script can also check your browser data folder to verify that you have what you need: bin/prepare-data check. Use bin/prepare-data without arguments for usage information.

Here is an example:

mallet train-topics ... \   # ...: input parameters
    --output-state mallet_state.gz --output doc-topics dt.csv
# generate tw.json and dt.json.zip
bin/prepare-data convert-state mallet_state.gz \
    --tw data/tw.json --dt data/dt.json.zip
# generate meta.csv.zip from dfr-data/citations.tsv
# dt.csv is needed only for doc id's
cut -f 2 dt.csv > ids.txt
bin/prepare-data convert-citations dfr-data/citations.tsv --ids ids.txt \
    -o data/meta.csv.zip
# generate info.json
bin/prepare-data info-stub -o data/info.json
# edit info.json by hand if needed
vi data/info.json

Sample datafiles

The data files used in the demo (PMLA, 64 topics) reside in a directory on the gh-pages branch of this repository.

Tune the visualization parameters

In the model-info file data/info.json, you can also override some aspects of the visualization by adding a VIS object with properties whose names correspond to those of the VIS object in the program. See VIS.js for the fields of the VIS object and their default values. Specifying VIS properties in info.json changes these defaults.

Many properties are nested. There are parameter settings for most of the model views as well as for metadata categories and bibliographic display. For example, to change the font size and number of words shown in the little circles representing topics in the Grid and Scaled overviews, modify some fields of VIS.model_view.plot:

"VIS": {
    "model_view": {
        "plot": {
            "words": 5,
            "size_range": [6, 14]
        }
    }
}

Five words will be shown in each circle, at font sizes between 6 and 14 points. The default values of other model_view.plot parameters, like aspect, will be conserved. (The layout of the Grid view can be changed by specifying rows and indents arrays in model_view.plot.)

Conditioning on metadata

So far dfr-browser has always had a display of topic proportions per year. But this is just a special case of something more general, a topic distribution conditional on a metadata variable, that is, the probability of a topic within a given metadata category. In addition to topics over time, dfr-browser can display other relations of this kind.

The VIS.condition parameter governs this display. It specifies what kind of variable the covariate is, and how to categorize documents into discrete categories using it. This is simplest to explain by example.

For publication date, the setting might be

"condition": {
    "type": "time",
    "spec": {
      "field": "date",
      "unit": "month",
      "n": 2
    }
}

This indicates that the display of topics over time should use the date field of the document metadata (loaded by default) and should group documents into intervals of two months.

For a categorical variable, the setting might be

"condition": {
    "type": "ordinal",
    "name": "journal",
    "spec": {
        "field": "journal"
    }
}

The topic displays will now show the total proportion of each journal's words assigned to that topic. ("Ordinal" is the d3 name for such variables, a reminder that an ordering is imposed on the category by alphabetization.) The optional name field gives a name to be used on axis labels. The field is used if it is missing, so it's superfluous in this case.

Finally, a continuous variable can be specified as follows. Let's imagine the data include a variable pagelen giving document page lengths. Topic distributions conditional on page length are probably uninteresting, though possibly of some diagnostic use, so this example is purely to illustrate the software mechanism:

"condition": {
    "type": "continuous",
    "name": "page length",
    "spec": {
        "field": "pagelen",
        "step": 10
    }
}

Which says that the pagelen field should be divided into bins of width 10 and topic proportions should be marginalized over those bins. This is a little more complex, because the field named here must exist in the metadata. In order to ensure that the ninth metadata column in the meta.csv[.zip] file has the name pagelen rather than X1, we also need this setting:

"metadata": {
    "spec": {
        "extra_fields": [ "pagelen" ]
    }
}

And of course the metadata column has to actually exist. Page length is not supplied in JSTOR metadata, but it is not difficult to calculate from the pagerange field.

It may also be necessary to adjust some of the graphical parameters specified in the topic_view property (see VIS.js for documentation), the corresponding parameters in model_view.list.spark, and related parameters in model_view.conditional. In particular, the width of the bars in bar charts, as well as the margin left for y-axis labels (topic_view.m.left), often requires manual tuning. Tuning parameters are keyed to variable types, so that, for example, the topic bar chart for a time covariate has settings in VIS.topic_view.time whereas the settings for a categorical covariate are VIS.topic_view.ordinal.

(What is displayed is only a naive estimate of the conditional probabilities, formed by marginalizing the topic distribution over levels or bins of the metadata covariate. Or, if you prefer, the display constitutes a visual posterior predictive check of the topic model: ordinary LDA assumes that documents are exchangeable and hence that metadata categories ought not to matter to topic probabilities.)

Hiding topics

If certain topics are distractingly uninterpretable, they can be hidden from the display by specifying a hidden_topics array as a property of VIS. The topics identified by numbers (indexed from 1) in this array will, by default, not be shown in any view, including aggregate views. Hidden topics can be revealed using the Settings dialog box. Hiding topics can be misleading and should be done cautiously.

Adding topic labels

In order to aid interpreting the model, it is often useful to manually label topics. The browser looks for labels in data/info.json, in a top-level topic_labels property (not a property of VIS) with the following slightly eccentric form:

"topic_labels": {
    "1": "label for topic 1",
    "4": "a very good topic"
}

One-based topic numbers as strings are the keys. This allows for easier editing by hand. Topics for which no label is supplied are automatically labeled by their number.

Launch the browser

The necessary files are index.html, the data files (looked for in data/ by default), and the css, js, lib, and fonts folders. Put all of these files in the path of a web server and go.

To preview locally, you will need a local web server, so that the javascript can ask for the data files from your file system. A widely available option is Python's SimpleHTTPServer module, which I have wrapped in a one-line script, so that you can simply type:

cd dfr-browser
bin/server

Adapting this project to other kinds of documents (or models)

The data-prep is tuned to MALLET and my dfrtopics package, but again altering the setup for other implementations of LDA would not be too challenging. Adapting to other kinds of latent topic models would require more changes.

Since I first released this model-browser, numerous (more than three!) people have been interested in using it to explore LDA models of other kinds of documents. This is more straightforward, since the specialization to JSTOR articles is limited to the expectations about the metadata format, the bibliography sort, and the way documents are cited and externally linked. Though some modifications to the program are necessary in such cases, they should normally be limited to the metadata and bib objects. Let me first explain a little more about the design of the program.

The design

The whole browser is a single webpage, whose source is index.html. This defines components of the layout---the various "views" on the model. It also loads includes the necessary JavaScript, including the d3 library and the dfr-browser scripts.

The browser follows (scrappily) the model-view-controller paradigm. The main controller is created by dfb(), which has the job of loading all the data and responding to user interaction. The data is stored in a model(), which encapsulates the details of storing the model and of aggregating its information in various ways. The details of parsing metadata are handled by separate metadata and bib objects. The controller then passes the necessary data from the model to the various view objects that configure the different views of the model. These view objects do all their work by accessing parts of index.html using CSS selectors and transforming them with d3. (This means that the JavaScript makes many assumptions about the elements that are present in index.html.)

As an example of a simple modification, suppose you wanted to eliminate one of the views, say the word index. You could simply delete the <div id="words_view"> element. The view could, however, still be accessed directly by entering the URL #/words in the web browser. The controller's method for dispatching to this view are set up by the lines reading

my.views.set("words", function (...) {
...
});

Remove the statement and the browser will no longer respond to #/words. You could now remove the file defining view.words. The other views are set up analogously (my.views.set("topic", ...), etc.). These anonymous functions are subsequently invoked in the refresh method of dfb(). To create a new view, you would add another statement within the definition of dfb():

my.views.set("newview", function (...) {
});

URLs of the form #/newview/x/y/z would cause the function to be invoked with parameters x, y, and z. The model data is available within the function as my.m (my is a local variable to the enclosing function dfb() which stores the controller's private member data.) This function should extract the model data needed for the view and then call an external function to actually render it.

In any case: to run the program, initialize a dfb() object and call its load() method. This will trigger a series of requests for data files and set up the event listeners for user interaction (the main one is the hashchange handler). In short, the following line must be executed after all the libraries and scripts have been loaded in index.html:

dfb().load();

Assumptions about the document metadata are restricted to two places: the metadata object and the bib object. A new dfb passes the metadata object the incoming data from the metadata file and then hands it off to the model for storage. When document metadata must be displayed or sorted, the dfb passes that data to the bib object for formatting or sorting.

There are various ways to customize these.

First, and most simply, suppose the metadata format is not quite the same as the DfR format assumed here. Metadata parsing is governed by the from_string method of metadata.dfr. Modify this method to modify parsing. Suppose that in your data, the author column comes before the title column. Then the lines

title: d[1].trim(),
authors: d[2].trim(),

should switch to

title: d[2].trim(),
authors: d[1].trim(),

Let's imagine that the publisher is found in the ninth metadata column, and VIS.metadata.spec.extra_fields = [ "publisher" ]. Documents will now have a publisher field, but to display this information, one must also change the methods of bib. The printing of document citations is governed by the citation method of bib.dfr. Imagine simply adding, before the return at the end of the function:

s += "Published by " + doc.publisher + "."

Similarly, the external document links are created by the url method, so if the URL should not lead to JSTOR, change that method.

If you specify VIS.metadata.type: "base" in info.json, the base class metadata() (defined in src/metadata.js) will be used instead of metadata.dfb(). This class parses the metadata file using d3.csv.parse (which assumes, unlike metadata.dfb, that it has a header naming the columns).

The functions that determine how documents are grouped for the topic proportions over time (or other conditional distribution) are the metadata.key properties, also defined in metadata.js.

To modify the bibliographic sorting, see bib.js (the derived bib.dfr object in bib/dfr.js only adds additional sorting options). If you want to be elaborate about it, you can also derive further objects from bib.dfr or bib or metadata and supply those to the constructor dfb() as follows:

dfb({
    metadata: my_meta(),
    bib: my_bib()
}).load();

I haven't been as consistent as I should have been about my programming idioms, but in general I follow the "functional object" pattern described by Douglas Crockford's Javascript: The Good Parts. The source for metadata.dfr illustrates how to write a derived object.

The "build" process

If you are modifying the code, note that the js/*.js files for the server are minified versions. The source code is found in the src directory. To make the minified scripts, run make uglify. This requires uglifyjs. On a Mac, install homebrew, then:

brew install npm
npm install uglify-js -g

Performance

The ranking and sorting calculations are done on the fly, but nothing else is, and the page holds all the data in memory. I haven't done much to optimize it. It's serviceable but not as fast as it could be. The most demanding calculations are done in a separate thread using an HTML5 Web Worker. Other optimizations would be possible, for example using ArrayBuffers for the big matrices rather than ordinary arrays.

For serving from a web server, a big model means a lot of data has to be sent to the client; keeping the doc-topic matrix sparse saves some room, as does zipping up the datafiles, but there are limits to this. Past a certain limit, it would be necessary to hold the model on the server, presumably in a real database. I haven't implemented this, but because access to the model is abstracted by the methods of the model object (see src/model.js), doing so would not be intractable.

Libraries

This browser uses the code of the following open-source projects by others, under the fonts, css, and lib directories: d3 by Mike Bostock; bootstrap by Twitter, Inc.; JQuery by the JQuery Foundation; and JSZip by Stuart Knightley.

This browser also makes use of code from Andrew Goldstone, Susana Galán, C. Laura Lovin, Andrew Mazzaschi, and Lindsey Whitmore, "An Interactive Topic Model of Signs," in Signs at 40 (source at http://github.com/signs40th/topic-model).