Package eyekit

Eyekit is a Python package for analyzing reading behavior using eyetracking data. Eyekit aims to be entirely independent of any particular eyetracker hardware, experiment software, or data formats. It has an object-oriented style that defines three core objects – the TextBlock, InterestArea, and FixationSequence – that you bring into contact with a bit of coding. Eyekit is currently in the alpha stage of development and is freely available under the MIT License.

Is Eyekit the Right Tool for Me?

  • You want to analyze which parts of a text someone is looking at and for how long.

  • You are mostly interested in fixations, as opposed to, for example, saccades, blinks, or millisecond-by-millisecond eye movements.

  • You need convenient tools for extracting areas of interest from texts, such as specific words, phrases, or letter combinations.

  • You want to calculate common reading measures, such as gaze duration or initial landing position.

  • You need support for arbitrary fonts, multiline passages, right-to-left text, or non-alphabetical scripts.

  • You want the flexibility to define custom reading measures and to build your own reproducible processing pipeline.

  • You would like tools for dealing with noise and calibration issues, and for discarding fixations according to your own criteria.

  • You want to share your data in an open format and produce publication-ready vector graphics.

Installation

Eyekit may be installed using pip:

$ pip install eyekit

Eyekit is compatible with Python 3.6+. Its main dependency is the graphics library Cairo, which you might also need to install if it's not already on your system. Many Linux distributions have Cairo built in. On a Mac, it can be installed using Homebrew: brew install cairo. On Windows, it can be installed using Anaconda: conda install -c anaconda cairo.

Getting Started

Once installed, import Eyekit in the normal way:

>>> import eyekit

Eyekit makes use of three core objects: the FixationSequence, TextBlock, and InterestArea. Much of Eyekit's functionality involves bringing these objects into contact. Typically, you define particular areas of the TextBlock that are of interest (phrases, words, morphemes, letters…) and check to see which fixations from the FixationSequence fall in those InterestAreas and for how long.

The FixationSequence object

Fixation data is represented in a FixationSequence. Usually you will import fixation data from some raw data files, but for now let's just create some pretend data to play around with:

>>> seq = eyekit.FixationSequence([[106, 490, 0, 100], [190, 486, 100, 200], [230, 505, 200, 300], [298, 490, 300, 400], [361, 497, 400, 500], [430, 489, 500, 600], [450, 505, 600, 700], [492, 491, 700, 800], [562, 505, 800, 900], [637, 493, 900, 1000], [712, 497, 1000, 1100], [763, 487, 1100, 1200]])

Each fixation is represented by four numbers: its x-coordinate, its y-coordinate, its start time, and its end time (so, in this example, they all last 100ms). Once created, a FixationSequence can be traversed, indexed, and sliced just like an ordinary list. For example,

>>> print(seq[5:10])
### FixationSequence[Fixation[430,489], ..., Fixation[637,493]]

slices out fixations 5 through 9 into a new FixationSequence object. A FixationSequence (and the Fixation objects it contains) has various properties that you can query:

>>> print(len(seq)) # Number of fixations
### 12
>>> print(seq.duration) # Duration of whole sequence
### 1200
>>> print(seq[0].duration) # Duration of first fixation
### 100
>>> print(seq[-1].xy) # xy coordinates of final fixation
### (763, 487)
>>> print(seq[1].x) # x coordinate of second fixation
### 190

The TextBlock object

A TextBlock can represent a word, sentence, or passage of text. When you create a TextBlock object, it is necessary to specify various details such as its position on the screen and the font:

>>> sentence = 'The quick brown fox [jump]{stem}[ed]{suffix} over the lazy dog.'
>>> txt = eyekit.TextBlock(sentence, position=(100, 500), font_face='Times New Roman', font_size=36)
>>> print(txt)
### TextBlock[The quick brown ...]

Eyekit has a simple scheme for marking up interest areas, as you can see in the above sentence. Square brackets are used to mark the interest area itself (in this case jump and ed) and curly braces are used to provide a unique ID for each interest area (in this case stem and suffix). These interest areas that have been specifically marked up in the raw text are called "zones". We can extract a particular zone as an InterestArea object by using its ID:

>>> print(txt['stem'])
### InterestArea[stem, jump]

or indeed we can iterate over all the marked-up zones using the TextBlock.zones() iterator:

>>> for zone in txt.zones():
>>>   print(zone)
### InterestArea[stem, jump]
### InterestArea[suffix, ed]

Zones are useful if you have a small number of interest areas that are convenient to mark up in the raw text. However, Eyekit also provides more powerful tools for automatically extracting interest areas. For example, you can use TextBlock.words() to iterate over every word in the text as an InterestArea without needing to explicitly mark each of them up in the raw text:

>>> for word in txt.words():
>>>   print(word)
### InterestArea[0:0:3, The]
### InterestArea[0:4:9, quick]
### InterestArea[0:10:15, brown]
### InterestArea[0:16:19, fox]
### InterestArea[0:20:26, jumped]
### InterestArea[0:27:31, over]
### InterestArea[0:32:35, the]
### InterestArea[0:36:40, lazy]
### InterestArea[0:41:44, dog]

You can also supply a regular expression to iterate over words that match a certain pattern; words that end in ed for example:

>>> for word in txt.words('.+ed'):
>>>   print(word)
### InterestArea[0:20:26, jumped]

or just the four-letter words:

>>> for word in txt.words('.{4}'):
>>>   print(word)
### InterestArea[0:27:31, over]
### InterestArea[0:36:40, lazy]

or case-insensitive occurrences of the word the:

>>> for word in txt.words('(?i)the'):
>>>   print(word)
### InterestArea[0:0:3, The]
### InterestArea[0:32:35, the]

You can also collate a bunch of interest areas into a list for convenient access later on. For example, if you wanted to do some analysis on all the three-letter words, you might extract them and assign them to a variable like so:

>>> three_letter_words = list(txt.words('.{3}'))
>>> print(three_letter_words)
### [InterestArea[0:0:3, The], InterestArea[0:16:19, fox], InterestArea[0:32:35, the], InterestArea[0:41:44, dog]]

You can also slice out arbitrary interest areas by using the row and column indices of a section of text. Here, for example, we are extracting a two-word section of text from line 0 (the first and only line) and characters 10 through 18:

>>> brown_fox_ia = txt[0:10:19]
>>> print(brown_fox_ia)
### InterestArea[0:10:19, brown fox]

The InterestArea object

Once you've extracted an InterestArea, you can access various properties about it:

>>> print(brown_fox_ia.text) # get the string represented in this IA
### brown fox
>>> print(brown_fox_ia.width) # get the pixel width of this IA
### 157.974609375
>>> print(brown_fox_ia.center) # get the xy coordinates of the center of the IA
### (328.4365234375, 491.94921875)
>>> print(brown_fox_ia.onset) # get the x coordinate of the IA onset
### 253.94921875
>>> print(brown_fox_ia.location) # get the location of the IA in its parent TextBlock
### (0, 10, 19)
>>> print(brown_fox_ia.id) # get the ID of the IA
### 0:10:19

By default, an InterestArea will have an ID of the form 0:10:19, which refers to the unique position that it occupies in the text. However, you can also change the IDs to something more descriptive, if you wish. For example, you could enumerate the words in the text and give each one a number:

>>> for i, word in enumerate(txt.words()):
>>>   word.id = f'word{i}'

and, since InterestAreas are passed around by reference, the IDs will update everywhere. For example, it we print the three_letter_words list we created earlier, we find that it now reflects the new IDs:

>>> print(three_letter_words)
### [InterestArea[word0, The], InterestArea[word3, fox], InterestArea[word6, the], InterestArea[word8, dog]]

The InterestArea defines the in operator in a special way so that you can conveniently check if a fixation falls inside its bounding box. For example:

>>> for fixation in seq:
>>>   if fixation in brown_fox_ia:
>>>     print(fixation)
### Fixation[298,490]
### Fixation[361,497]
>>> for fixation in seq:
>>>   for zone in txt.zones():
>>>     if fixation in zone:
>>>       print(f'{fixation} is inside {zone}')
### Fixation[430,489] is inside InterestArea[stem, jump]
### Fixation[450,505] is inside InterestArea[stem, jump]
### Fixation[492,491] is inside InterestArea[suffix, ed]

The TextBlock also provides some more convenient methods for asking these kinds of questions, for example:

>>> print(txt.which_word(seq[3])) # which word is fixation 3 inside?
### InterestArea[word2, brown]
>>> print(txt.which_character(seq[3])) # which character specifically?
### InterestArea[0:12:13, o]

An InterestArea can be any sequence of consecutive characters, and, as you can see, it's possible to define several overlapping InterestAreas at the same time: txt['word4'] (jumped), txt['stem'] (jump), txt['suffix'] (ed), and txt[0:24:25] (the p in jumped) can all coexist.

Visualization

Now that we've created a FixationSequence and TextBlock, and we've extracted some InterestAreas, it would be useful to visualize how these things relate to each other. To create a visualization, we begin by creating an Image object, specifying the pixel dimensions of the screen:

>>> img = eyekit.vis.Image(1920, 1080)

Next we render our text and fixations onto this image:

>>> img.draw_text_block(txt)
>>> img.draw_fixation_sequence(seq)

Note that the elements of the image will be layered in the order in which these methods are called – in this case, the fixations will be rendered on top of the text. Finally, we save the image (Eyekit supports PDF, EPS, SVG, or PNG):

>>> img.save('quick_brown.svg')

Sometimes it's useful to see the text in the context of the entire screen, as is the case here; other times, we'd like to remove all that excess white space and focus in on the text. To do this, you need to specify a crop margin; the image will then be cropped to the size of the text block plus the specified margin:

>>> img.save('quick_brown_cropped.svg', crop_margin=2)

There are many other options for creating custom visualizations, which you can explore in the eyekit.vis module. For example, if you wanted to depict the bounding boxes around the two zoned interest areas we defined earlier, you might do this:

>>> img = eyekit.vis.Image(1920, 1080)
>>> img.draw_text_block(txt)
>>> img.draw_rectangle(txt['stem'], color='crimson')
>>> img.draw_rectangle(txt['suffix'], color='cadetblue')
>>> img.draw_fixation_sequence(seq)
>>> img.save('quick_brown_with_zones.svg', crop_margin=2)

Colors can be specified as a tuple of RGB values (e.g. (220, 20, 60)), a hex triplet (e.g. #DC143C), or any standard HTML color name (e.g. crimson). Similarly, we can do the same thing with the list of three-letter words we created earlier:

>>> img = eyekit.vis.Image(1920, 1080)
>>> img.draw_text_block(txt)
>>> for word in three_letter_words:
>>>   img.draw_rectangle(word, fill_color='slateblue', opacity=0.5)
>>> img.draw_fixation_sequence(seq)
>>> img.save('quick_brown_with_3letter_words.svg', crop_margin=2)

Or, indeed, all words in the text:

>>> img = eyekit.vis.Image(1920, 1080)
>>> img.draw_text_block(txt)
>>> for word in txt.words():
>>>   img.draw_rectangle(word, color='hotpink')
>>> img.draw_fixation_sequence(seq)
>>> img.save('quick_brown_with_all_words.svg', crop_margin=2)

Note that, by default, each InterestArea's bounding box is slightly padded by, at most, half of the width of a space character. This ensures that, even if a fixation falls between two words, it will still be assigned to one of them. Padding is only applied to an InterestArea's edge if that edge adjoins a non-alphabetical character (i.e. spaces and punctuation). Thus, when jumped was divided into separate stem and suffix areas above, no padding was applied word-internally. If desired, automatic padding can be turned off by setting the autopad argument to False during the creation of the TextBlock, or it can be controlled manually using the InterestArea.adjust_padding() method.

Taking Measurements

The eyekit.measure module provides various functions for computing common reading measures, including:

These functions take an InterestArea and FixationSequence as input, and return the measure of interest. For example, if we wanted to measure the initial landing position on the stem interest area, we can apply the function like this:

>>> print(eyekit.measure.initial_landing_position(txt['stem'], seq))
### 2

The initial fixation on jump landed on character 2. You can also apply these functions to a collection of interest areas. This will return a dictionary of results in which the keys are the InterestArea IDs. For example, if we wanted to calculate the initial fixation duration on the three-letter words we extracted earlier, we can do this:

>>> print(eyekit.measure.initial_fixation_duration(three_letter_words, seq))
### {'word0': 100, 'word3': 100, 'word6': 100, 'word8': 100}

In this case, we see that the initial fixation on each of the three-letter words lasted 100ms. Likewise, if we wanted to measure the gaze duration on all words, we can do this:

>>> print(eyekit.measure.gaze_duration(txt.words(), seq))
### {'word0': 100, 'word1': 200, 'word2': 100, 'word3': 100, 'word4': 300, 'word5': 100, 'word6': 100, 'word7': 100, 'word8': 100}

The measurement functions provided by Eyekit can be used as-is or you can take a look at the underlying code and adapt them for your own purposes.

Multiline Passages

So far, we've only looked at a single line TextBlock, but handling multiline passages works in largely the same way. The principal difference is that when you instantiate your TextBlock object, you must pass a list of strings (one for each line of text):

>>> txt = eyekit.TextBlock(['This is line 1', 'This is line 2'], position=(100, 500), font_face='Helvetica', font_size=24)

To see an example, we'll load in some real multiline passage data from Pescuma et al. which is included in the Eyekit GitHub repository:

>>> example_data = eyekit.io.read('example/example_data.json')
>>> example_texts = eyekit.io.read('example/example_texts.json')

and in particular we'll extract the fixation sequence for trial 0 and its associated text:

>>> seq = example_data['trial_0']['fixations']
>>> pid = example_data['trial_0']['passage_id']
>>> txt = example_texts[pid]['text']

As before, we can plot the fixation sequence over the passage of text to see what the data looks like:

>>> img = eyekit.vis.Image(1920, 1080)
>>> img.draw_text_block(txt)
>>> img.draw_rectangle(txt[0:32:40], color='orange')
>>> img.draw_rectangle(txt[1:34:38], color='orange')
>>> img.draw_fixation_sequence(seq)
>>> img.save('multiline_passage.svg', crop_margin=4)

First, we might decide that we want to discard that final fixation, where the participant jumped a few lines up right at the end:

>>> seq[-1].discard() # discard the final fixation

A second problem we can see here is that fixations on one line sometimes appear slightly closer to another line due to imperfect eyetracker calibration. For example, the fixation on the word voce on line two actually falls into the bounding box of the word vivevano on line one. Obviously, this will cause issues in your analysis further downstream, so it can be useful to first clean up the data by snapping every fixation to its appropriate line. Eyekit implements several different line assignment algorithms, which can be applied using the snap_to_lines() function from the eyekit.tools module:

>>> original_seq = seq.copy() # keep a copy of the original sequence
>>> eyekit.tools.snap_to_lines(seq, txt)

This process adjusts the y-coordinate of each fixation so that matches the midline of its assigned line. To compare the corrected fixation sequence to the original, we could make two images and then combine them in a single Figure, like so:

>>> img1 = eyekit.vis.Image(1920, 1080)
>>> img1.draw_text_block(txt)
>>> img1.draw_rectangle(txt[0:32:40], color='orange')
>>> img1.draw_rectangle(txt[1:34:38], color='orange')
>>> img1.draw_fixation_sequence(original_seq)
>>> img1.set_caption('Before correction')
>>> 
>>> img2 = eyekit.vis.Image(1920, 1080)
>>> img2.draw_text_block(txt)
>>> img2.draw_rectangle(txt[0:32:40], color='orange')
>>> img2.draw_rectangle(txt[1:34:38], color='orange')
>>> img2.draw_fixation_sequence(seq)
>>> img2.set_caption('After correction')
>>> 
>>> fig = eyekit.vis.Figure(1, 2) # one row, two columns
>>> fig.add_image(img1)
>>> fig.add_image(img2)
>>> fig.set_crop_margin(3)
>>> fig.save('multiline_passage_corrected.svg')

The fixation on voce is now clearly associated with the correct word. Nevertheless, the snap_to_lines() function should be applied with care, especially if the fixation data is very noisy or if the passage is being read nonlinearly. For advice on which method to use, see Carr et al. (2021).

Just like single-line texts, we can extract interest areas from the passage and take measurements in the same way. For example, if we were interested in the word piccolo/piccola in this passage, we could extract all occurrences of this word and calculate the total fixation duration:

>>> piccol_words = list(txt.words('piccol[oa]'))
>>> durations = eyekit.measure.total_fixation_duration(piccol_words, seq)
>>> print(durations)
### {'2:64:71': 253, '3:0:7': 347, '3:21:28': 246, '3:29:36': 319, '7:11:18': 268, '10:43:50': 178}

Furthermore, we could make a visualization to show this information:

>>> img = eyekit.vis.Image(1920, 1080)
>>> img.draw_text_block(txt)
>>> for word in piccol_words:
>>>   img.draw_rectangle(word, color='lightseagreen')
>>>   x = word.onset
>>>   y = word.y_br - 3
>>>   label = f'{durations[word.id]}ms'
>>>   img.draw_annotation((x, y), label, color='lightseagreen', font_face='Arial bold', font_size=4)
>>> img.draw_fixation_sequence(seq, color='gray')
>>> img.save('multiline_passage_piccol.svg', crop_margin=4)

Another way to look at the data is to distribute the fixations across the characters of the passage probabilistically, under the assumption that the closer a character is to a fixation point, the more likely it is that the reader is perceiving that character. This can be performed with the duration_mass() function and plotted in a heatmap like so:

>>> mass = eyekit.measure.duration_mass(txt, seq)
>>> img = eyekit.vis.Image(1920, 1080)
>>> img.draw_heatmap(txt, mass, color='green')
>>> img.save('multiline_passage_mass.svg', crop_margin=4)

Input–Output

Eyekit is not especially committed to any particular file format; so long as you have an x-coordinate, a y-coordinate, a start time, and an end time for each fixation, you are free to store data in whatever format you choose. However, as we have seen briefly above, Eyekit provides built-in support for JSON, where a typical data file might look something like this:

{
  "trial_0" : {
    "participant_id": "John",
    "passage_id": "passage_a",
    "fixations": { "__FixationSequence__" : [[412, 142, 770, 900], ..., [655, 653, 46483, 46532]] }
  },
  "trial_1" : {
    "participant_id": "Mary",
    "passage_id": "passage_b",
    "fixations": { "__FixationSequence__" : [[368, 146, 7, 197], ..., [725, 681, 30331, 31260]] }
  },
  "trial_2" : {
    "participant_id": "Jack",
    "passage_id": "passage_c",
    "fixations": { "__FixationSequence__" : [[374, 147, 7, 283], ..., [890, 267, 31931, 32153]] }
  }
}

This format is compact, structured, human-readable, and flexible. With the exception of the __FixationSequence__ object, you can freely store whatever key-value pairs you want and you can organize the hierarchy of the data structure in any way that makes sense for your project. JSON files can be loaded using the read() function from the eyekit.io module:

>>> data = eyekit.io.read('example/example_data.json')

which automatically instantiates any FixationSequence objects. Similarly, an arbitrary dictionary or list can be written out using the write() function:

>>> eyekit.io.write(data, 'output_data.json', compress=True)

If compress is set to True, files are written in the most compact way; if False, the file will be larger but more human-readable (like the example above). JSON can also be used to store TextBlock objects – see example_texts.json for an example – and you can even store FixationSequence and TextBlock objects in the same file if you like to keep things organized together.

The eyekit.io module also provides functions for importing data from other formats: import_asc() and import_csv(). Once data has been imported this way, it may then be written out to Eyekit's native JSON format for quick access in the future. In time, I hope to add more functions to import data from common eyetracker formats.

Getting Texts Into Eyekit

Getting texts into Eyekit can be a little tricky because their precise layout will be highly dependent on many different factors – not just the font and its size, but also the peculiarities of the experiment software and its text rendering engine.

Ideally, all of your texts will be presented in some consistent way. For example, they might be centralized on the screen or they might have a consistent left edge. Once you specify how a text is positioned on screen, Eyekit calculates the location and bounding box of every character based on the particular font and font size you are using.

The best way to check that the TextBlock is set up correctly is to check it against a screenshot from your actual experiment. Eyekit provides the align_to_screenshot() tool to help you do this. First, set up your text block with parameters that you think are correct:

>>> txt = eyekit.TextBlock(saramago_text, position=(300, 100), font_face='Baskerville', font_size=30, line_height=60, align='left', anchor='left')

Then pass it to the align_to_screenshot() function along with the path to a PNG screenshot file:

>>> eyekit.tools.align_to_screenshot(txt, 'screenshot.png')

This will create a new image file ending _eyekit.png (e.g. screenshot_eyekit.png). In this file, Eyekit's rendering of the text is presented in green overlaying the original screenshot image. The point where the two solid green lines intersect corresponds to the TextBlock's position argument, and the dashed green lines show the baselines of subsequent lines of text, which is based on the line_height argument. You can use this output image to adjust the parameters of the TextBlock accordingly. In this example case, we see that the TextBlock is positioned slightly too high up. If all of your texts are presented in a consistent way, you should only need to establish these parameters once.

An alternative strategy would be to produce your experimental stimuli using Eyekit. For example, you could export images of your TextBlocks, and then display them full-size in some experiment software of your choice.

Multilingual Support

Eyekit aims to offer good multilingual support, and the most common scripts – Arabic, Chinese, Cyrillic, Greek, Hebrew, Japanese, Korean, Latin – should all work out of the box. Right-to-left text (and bidirectional text in general) is supported – all you need to do is set right_to_left=True when creating a TextBlock. This ensures that the text will be rendered correctly and that functions like initial_landing_position() and snap_to_lines() will process the text in right-to-left direction. If you are working with the Arabic script, the text should be shaped prior to passing it into Eyekit using, for example, the Arabic-reshaper package.

Eyekit uses Cairo's "toy font" API to extract character metrics from the fonts available on your system. This API can be somewhat imperfect, especially if you are working with a particularly complex script or advanced typographical features, such as ligatures and kerning. However, in most cases it should be more than sufficient to extract areas of interest fairly accurately. When choosing a font for your experiment, the key thing to do is to make sure it supports all the glyphs in the language you're working with (some software, for example, may fall back to an alternative font in cases where a glyph is missing).

Contributing

Eyekit is still in an early stage of development, but I am very happy to receive bug reports and suggestions via the GitHub Issues page. If you'd like to work on new features or fix stuff that's currently broken, please feel free to fork the repo and/or raise an issue to discuss details. Before sending a pull request, you should check that the unit tests pass using Pytest:

$ pytest tests/

and run Black over the codebase to normalize the style:

$ black eyekit/

Here are some areas of Eyekit that are currently underdeveloped:

  • Additional reading measures (e.g. of saccades and regressions)
  • Awareness of different experimental paradigms
  • Creation of animations/videos
  • More convenient methods for collating results into dataframes etc.
  • Importing data from other eyetracker data formats
  • Synchronization of fixation data with other types of experimental event
  • Support for nontextual objects, such as images or shapes
  • Interactive tools for cleaning up raw data
  • General testing and the addition of more unit tests

Documentation

eyekit.fixation

Defines the Fixation and FixationSequence objects, which are used to represent fixation data.

eyekit.io

Functions for reading and writing data.

eyekit.measure

Functions for calculating common reading measures, such as gaze duration or initial landing position.

eyekit.text

Defines the TextBlock and InterestArea objects for handling texts.

eyekit.tools

Functions for performing common procedures, such as discarding out of bounds fixations and snapping fixations to the lines of text.

eyekit.vis

Defines the Image, Figure, and Booklet objects, which are used to create visualizations.