Process Features¶

Process features capture the affective, cognitive and behavioral dynamics of action (e.g., translation) as they unfold in time. In the context of the TPR-DB 3.0, they comprise keystroke logging measures — such as production pauses and typing bursts — and gaze metrics derived from eye-tracking, including fixation durations, regressive eye movements, and re-reading behavior. Many of those features are described in the TPR-DB version 2.0.

The definitions of many TPR-DB features, such as Dur, Ins, Del, FixS, FixT, etc. have not changed in the TPR-DB 3.0, although - due to a redefinition of word boundaries and the reimplementation of the The TPR Pipeline in Python along with several algorithmic changes - their values may be silightly different. In addition, some features are now differently defined.

This section therefore, mainly focuses on new features or features for which the definition has changed.

Source and Target Groups and IDs¶

The features STid, TTid, SGid, and TGid appear in several TPR-DB tables. STidand TTid are indexes of individual tokens in the source and target text respectively, and are, therefore, integers. SGid and TGid (may) refer to gropus of words and are, therefore, strings, i.e., the concatenation of several token IDs, e.g., "\(\mathtt{8+10}\)".

During keystroke-to-word mapping each keystroke is associated with a unique target word. The Id of this target word is the TTid, an integer, in the KD file. During Bilingual Alignment source words and target words are connected in a way such that one target word can be associated with several source words, i.e., the SGid. Each SGid is, in turn, aligned with one or more target words, which together are the TGid. Thus, the TTid is part of the TGid, but the latter may contain several elements. The STid is smallest number in the source group.

For instance, a target token "\(\mathtt{8}\)" can be aligned with source tokens "\(\mathtt{8}\)" and "\(\mathtt{9}\)", thus the SGid value is "\(\mathtt{8+9}\)" and STid is "\(\mathtt{8}\)". If this source group is aligned to a group of target words, say, target words "\(\mathtt{8}\)" and "\(\mathtt{10}\)", the TGid has the value "\(\mathtt{8+10}\)".

A similar mechanism applies to fixations. While keystrokes can only occur in the target text, fixations are observed on source and target words. Thus, a fixation may occur on a source word "\(\mathtt{7}\)" which is associated with target words "\(\mathtt{5}\)" and "\(\mathtt{6}\)". If this target group is associated with source words "\(\mathtt{7}\)" and "\(\mathtt{8}\)", the SGid of this fixation is "\(\mathtt{7+8}\)", while TTid is "\(\mathtt{5}\)".

This same principle applies to these features also in the other tables, including AU, KU, PU, ST, TT, etc. where the ID refers to the token, while the group feature relates to the alignment group to which the unit refers.

Keystroke Units (KUs) and Production Units (PUs)¶

As for the TPR-DB version 2.0, also the TPR-DB 3.0 fragments the flow of keystrokes into processing units. Keystroke-based processing units are separated by a pre-defined lag of time between successive keystrokes (aka inter keystroke intervals, IKIs). The TPD-DB 3.0 distinguishes between two tresholds which, respecively, separate Keystroke Units (KUs) and Production Units (PUs). KUs consist of at least one keystroke separated from the next KU by an IKI \(\ge\) KUI (the KU Interruption threshold); PUs are separated by an IKI \(\ge\) PUB, a PU Break. Bandaru and others ¹ define these thresholds as:

• KUI : \(2 \times median(\text{within word IKI})\)
• PUB : IKI duration of the keystroke quantile: \(1 - \frac{3 \times \text{text length in characters}}{\text{number of words}}\)

KUs and PUs are enumerated in two separate tables. KU tables enumerate the sequence of KUs and the intervening keystroke pauses (KUI, and PUB) in their sequential order. The Type of a KU can be one of \(\mathtt{I}\), \(\mathtt{C}\) or \(\mathtt{D}\), depending on whether the keystrokes are only insertions, deletions or both insertions and deletions, respecively. A keystroke pause can be of Type \(\mathtt{K}\), (KUI) or a \(\mathtt{P}\) (PUB).

As for each session, KUI \(\le\) PUB, every PU consists of one or more KU(s). Thus, there is no overlapping between KUs and PUs.

Keystroke pauses¶

Lacruz and colleagues² introduce several metrics to quantify the relationship between text production (i.e., sequences of fluent typing or drafting) and pausing, on the assumption that keystroke pauses are good indicators of cognitive demand in monolingual language production and translation.

Their metrics include, among others:

• Pause Ratio (PR): the proportion of segment time spent pausing.

• Average Pause Ratio (APR): the average duration per pause relative to the average word duration.

• Pause-to-Word Ratio (PWR): the number of pauses relative to the number of words in a segment.

The TPR-DB provides basic features for computing these and other pause metrics at the segment level (SG).

The pause metrics rely on a notion of pause, which has been the subject of debate for many years. Following the literature, TPR-DB version 2.0 used fixed Pause thresholds of 500ms, 1000ms, 2000ms, and 5000ms. In TPR-DB 3.0, pausing behavior is instead treated as translator-specific, so thresholds are now also based on the translator-specific values KUI and PUB.

The following segment-level features are used in the TPR-DB pause metrics:

• Dur: production duration for a segment — the time between the first and last keystroke.
• Nedit: the number of times the segment was edited.
• PreGap: the segment-initial keystroke pause — the lag between the last keystroke of the previous segment (or the start of the session) and the first keystroke of the current segment.
• PostGap: the segment-final keystroke pause — the lag between the last keystroke of the current segment and the first keystroke of the next segment (or the end of the session).
• TB_pause: the number of typing bursts, given a pause threshold.
• TD_pause: the total duration in typing bursts, given a pause threshold.
• TG_pause: the total pausing duration, i.e., the gap between bursts, given a pause threshold.

Note that for any pause threshold value, Dur = TD_pause + TG_pause. That is, 1) neither the PreGap preceding the first keystroke event in a segment nor the PostGap following the last keystroke in that segment is included in Dur and 2) depending on the duration of the pause threshold, the total duration is distributed differently across TD_pause and TG_pause: longer thresholds increase the total drafting duration (TD); they decreas the number of typing bursts (TB) and proportinally shorten the total pausing duration (TG). Note that a segment may be edited multiple times; in that case, the relevant features are summed across edits. This means a segment's PostGap may differ from the next segment's PreGap, if a segment was edited more than once.

Depending on the definition adopted, if PostGap is counted as a pause within the segment, the number of pauses equals the number of typing bursts (TB) plus one; otherwise, the number of pauses is simply equal to TB.

Based on these considerations, the pause metrics can be computed as follows:

These equations are also part of the CRITT academy and are explained there in more detail.

Typing Inefficiency (InEff)¶

We adopt the definition of InEff from TPR-DB version 2.0, p.26, who define typing (in)efficiency for a word, chunk or segment as:

They approximate this in terms of number of insertions and deletions:

A number 1 is added to the denominator to prevent division by 0, for instance in case of postediting when a word or segment remains unchanged. In the current version we also add 1 to the nominator, so that if no deletions are recorded, the metric will return 1 irrespectively of how many deletions occurred.

Note that this measure only applies if number of insertions >= number of deletions which ensures that the result >= 1. Otherwise, if there are more deletions than insertions, as might be the case in post-editing, InEff is computed as follows, which provides a number between 0 and 1:

Gaze measures¶

Fixations are often considered the basic units of gaze data analysis. Depending on the sample rate of the eyetracken, fixations aggregate several to many gaze sample points \(SP\) into one unit, the fixation. In the TPR-DB. the FD table enumerates fixations of a translation session. Fixations can be further aggregated into so-called areas of interest (AOI). The ST and TT tables define each word (i.e., a token) an AOI, the AOIs in the AG table are alignment groups and the SG tabels aggregates fixation data for an entire segment.

Gaze data can be described on several levels of granularity:

The key principle is that each level adds more temporal and spatial context around the fixation event. Currently, the TPR-DB provides some measures for levels 0 - 3 but not for level 4.

Level 0 — Sub-fixation / Oculomotor Measures¶

A fixation is a dynamic event that changes in time with respect to the gaze position (micro saccades) and the pupil size. Level 0 measures capture raw oculomotor signal below the fixation level:

• Saccade amplitude, velocity, direction
• Microsaccades during fixation
• Pupil dilation (cognitive load proxy)
• Blink rate and duration

Pupillometry¶

While the pupil's primary function is the regulation of light intake, it also responds to cognitive and emotional states, which modulates arousal and mental effort. When a task becomes more mentally demanding, the pupil dilates (widens) slightly — typically by fractions of a millimeter — even under constant lighting conditions. This response is involuntary and continuous, making it useful for tracking moment-to-moment fluctuations in processing difficulty. This makes pupil diameter a sensitive, non-invasive proxy for cognitive effort and task-related arousal.

In reading research, pupillometry helps reveal where and when comprehension becomes effortful:

• Lexical difficulty: rare or low-frequency words trigger measurable dilation compared to common words
• Syntactic complexity: structurally ambiguous or embedded clauses cause increased load
• Discourse integration: resolving pronouns or bridging inferences produces detectable peaks

Pupil responses are often time-locked to specific words, allowing researchers to pinpoint which linguistic features drive difficulty. However, particular caution needs to be taken in pupillometry since blinks can produce extreme values (outliers) and there are large individual differences: readers with lower working memory capacity tend to show larger and more prolonged dilation responses.

Pupillometry has also become a tool in TPR to assess:

• Source text difficulty: complex syntax, ambiguous terms, or dense domain-specific content.
• Decision points: moments of terminological or stylistic uncertainty show elevated load
• Revision behavior: re-reading and self-correction phases correlate with pupil dilation spikes

Expertise differences — professional translators tend to show more efficient (lower or faster-recovering) pupil responses than novices on equivalent texts, suggesting automatization of sub-processes

Pupillometry in the TPR-DB¶

Pupil measures are a new feature in the TPR-DB 3.0. Pupillometry is the measurement of pupil size and its dynamic changes over time.

Given the heterogeneous nature of the TPR-DB (different eyetrackers with sampling rates, different lighting conditions and recording environments, etc.) and since the pupil size and their changes is specific to every participant, a normalization is required. To address this issue, the TPR-DB computes pupil size baseline as the median pupil diameter (rather than the mean) for every translation session. The pupillometric measures are then based on change in pupil diameter relative to the baseline in each session. For each gaze sample point (\(SP\), depending on the eyetracker sampling rate) the TPR-DB computes pupilometric measures in several steps:

1. Average pupil diameter across both eyes for binocular gaze sample data
2. Compute median pupil diameter across an entire session as a baseline
3. Interpolate over blink gaps, fill edges for sequences of gaze samples
4. Compute deviation from baseline in various ways for each gaze sample
5. Compute pupillometric measures for each fixation or other unit

For each gaze sample point \(SP\), the TPR-DB 3.0 computes an effective pupil size \(SP_p\) as the mean of the left and right pupil diameters when both are available (i.e., diameter \(> 0\)) for binocular tracking, and otherwise falls back to the available monocular diameter. Each \(SP_p\) is then normalised by the session median. The TPR-DB 3.0 computes two measures of dispersion per participant session: a robust median absolute deviation (MAD) and a standard deviation:

• \(\mathtt{baseline} = \text{median}(SP_{p})\)
• \(\mathtt{pupil\_mad} = \text{median}(|SP_{p} - \mathtt{baseline}|)\)
• \(\mathtt{pupil\_std} = \text{std}(SP_{p})\)

An \(SP\) can be said to be in a dilated or constricted state relative to the \(\mathtt{baseline}\), i.e., a dilation if \(SP_{p} > \mathtt{baseline}\) and a constriction if \(SP_{p} <= \mathtt{baseline}\)

Because pupillary responses unfold over several hundred milliseconds, pupil dilatipon measures should generally be aggregated over higher-level units such as words, AUs, PUs, HORF states, translation phases, or segments before substantive interpretation. The TPR-DB computes three sample-level measures 1. percent of change per from the baseline, 2. a mean robust std-score mad 3. a median-centred z-score std:

1. per: percentage change from $\mathtt{baseline} (session median): \(SP_{\text{per}} = 100 \times \frac{SP_{\text{p}} - \mathtt{baseline}}{\mathtt{baseline}}\)
2. mad: median-centred MAD-standardized pupil size: \(SP_{\text{mad}} = \frac{SP_{\text{p}} - \mathtt{baseline}}{\mathtt{pupil\_mad}}\)
3. std: median-centred SD-standardized pupil size: \(SP_{\text{std}} = \frac{SP_{\text{p}} - \mathtt{baseline}}{\mathtt{pupil\_std}}\)

For each of the three measures ([per|std|mad]) the TPR-DB produces the following eight measures in pupillometry research, for each fixation in the FD tables:

We make a distinction between primary pupil measures, the mean measures (PUP_per_mean and PUP_std_mean, PUP_mad_mean), secondary measures, the max and min measures (PUP_per_max; PUP_per_min and PUP_per_sd); and exploratory measures, the variants of AUC, time-normalised AUC, time-normalised AUC for constriction and time-normalised AUC for dilation.

zMAD​= x - median(x) divided by 1.4826 times MAD​ The factor 1.4826 arises from the relationship between the median absolute deviation (MAD) and the standard deviation under normality. It ensures that: if the data are approximately normally distributed, the robust z-score is on the same scale as a conventional z-score; if the data contain outliers, the estimate remains much more stable than the standard deviation.

Level 1 — Fixation measures on a word level¶

Level 1 measures capture what happens at one specific location, in isolation from surrounding context. Fixations are quantified based on their position on the screen (X/Y coordinates), duration, and the character/word/image looked. They reflect early, bottom-up processing:

• First fixation duration: how long the eye rests on first landing in a word
• Single fixation duration: duration when the word receives only one fixation
• Total fixation duration / dwell time: summed time across all fixations on the word
• Number of fixations: how often the word is fixated
• Fixation probability: whether the word was fixated at all

Fixation measures in the TPR-DB¶

These metrics have been the in the main focus of research and are documented in the TPR-DB version 2.0 an are FD and FU tables.

Level 1 metrics include:

• Time offset (in ms) of first fixation on the source token FFtimeS and on the target tokenFFtimeT
• First fixation duration on the source token FFDurS, and on the target token FFDurT
• Total number of fixations the source token FixS and the trget token FixT
• Total fixation duration (reading time) of the source token TrtS, or on the target token TrtT
• Fixation duration (reading time) in the orientation phase of the source token OrtS and target token OrtT
• Fixation duration (reading time) during first micro unit 1 (MU1) of the source token MU1rtS and target token MU1rtT

Level 2 — Transitions between words¶

Level 2 measures capture how the eye moves between words. The unit of analysis is the fixation behavior on a word in relation to the neighboring words:

• First-pass reading time: total fixation time of a word before the eye leaves it for the first time
• Regression path duration: time from first fixation until the eye moves rightward past the word, including any regressions launched from it
• Go-past time: total time including regressions back into earlier AOIs triggered by this region
• Re-reading time: time spent on second and later passes
• Regression rate in / out: number of regressions into (entering) or out of (leaving) the word
• Refixation rate: probability of making more than one fixation before leaving

Fixation transitions measures in the TPR-DB¶

Level 2 metrics include:

• TO BE COMPLETED

Level 3 — Gaze path measures and local reading strategies¶

These measures describe the local reading strategy. The unit of analysis is a sequence of fixations and how the eyes move sequentially.

• Linear reading: sequential, left-to-right, top-to-bottom progression
• Scattered fixations: non-sequential, exploratory, attention-driven
• Regressions / refixations — backward movements, indicating comprehension difficulty or verification
• Scan path similarity — comparing one participant's path to another or to an ideal (e.g., Levenshtein distance on AOI sequences)
• Coverage / revisit rate — what proportion of AOIs were visited, and how often
• Reading order deviation — how much the actual sequence departs from canonical order

Gaze patterns in the TPR-DB¶

While Fixation transitions measures capture transitions to and from a word (a static AOI, e.g., a chunk or segment) Local reading strategies describe sequences of gaze patterns. These gaze patterns may occur on the source text or on the target text. As soon as the eyes switch from the source to the target window or vice versa, a new pattern starts. We make a distinction between three types of gaze patterns:

1. Linear reading: sequences of mainly progressive fixations indicating successive reading, i.e., input of new information.
2. Refixation/regressive reading: sequences of fixations on the same or on neighboring words, probably indicating comprehension difficulties.
3. Scattered gaze patterns: sequences of fixations that do not fit the two categories above
4. No gaze detected: a non-gazing pattern

Each fixation in the \(\mathbb{FD}\) table is tagged with a label L, R, or S depending on whether the fixation is part of a linear, refixation, or scattered gaze pattern. AU tables provide features Dur_L, Dur_R, and Dur_S, which indicate the sum of fixation durations belonging to each of the three patterns, and Dur_N for the duration in which no gaze data was collected. The features RelDur_L, RelDur_R, RelDur_S, and RelDur_N provide the proportion of total gaze time spent in each pattern, computed as:

Level 4 — Global gaze behavior¶

Above the local gaze path level, aggregating across multiple trials or participants:

• Mean gaze behavior per participant (reading speed profile, regression tendency)
• Cluster-based reader types (skimmer vs. careful reader)
• Learning effects across trials

Global gaze behavior in the TPR-DB¶

as of now there are no such measures in the TPR-DB

Translation Phases in the TPR-DB¶

According to Jakobsen (2011) translation sessions can be separated into an orientation phase (O), a drafting phase (D) and a revision phase (R). Drafting starts with the first keystroke and the time before is defined as the orientation phase. We adopt this definition, even though it may not always be entirely correct. Some translators actually start with testing the keyboard, by typing some characters, and then start the actual orientation phase. We will ignore these cases. According to Jakobsen, drafting ends when the last word has been typed. We operationalize this definitionas follows:

1. We take the last word in the target text, rather than the translation of the last source text word.
2. Drafting produces at least 50% of the keystrokes in a translation sessions.
3. Drafting proceeds sequentially, i.e., successive keystrokes are no further than [-5 .. +2] word IDs and no more than [-20 .. +10] cursor positions apart. Drafting ends when five (or more) successive keystrokes not in a sequential.