Automated Extraction of Prosodic Structure from Unannotated Sign Language Video

Antonio F. G. Sevilla

Antonio's name sign.

José María Lahoz-Bengoechea

JM's name sign.

Alberto Díaz Esteban

Alberto's name sign.

Prosody is an important carrier of linguistic information in sign languages. One prominent way this manifests is through the temporal structure of signs, including their rhythm and intensity of articulation. To empirically observe these effects, the velocity of the hands can be computed throughout the execution of a sign.

We propose a method for extracting this information from unlabeled videos of sign language, utilizing CoTracker, a recent advancement in computer vision that can track every point in a video without any calibration or fine-tuning. The dominant hand is identified through clustering of the computed point velocities, and its dynamic profile is plotted to reveal the prosodic structure of signing.

Prosodic Profiles

An MH sign: BSL “LOOK”. Segments are clearly identifiable in the plot, and conform to Liddell and Johnson’s model of movements and holds.

We plot the velocity of articulation (speed of the hand(s)) throughout the video duration. This generates a prosodic profile that helps elucidate the temporal structure of signing: distinct segments produce identifiable regions in the plots. Movement direction (line color), distinguishes changes of regime in articulation and aids in the identification of segments.

Movement segments (M) are characterized by peaks in velocity as the hand moves between target points, while hold segments (H) are plateaus with minimal movement, where the hand is held in position. Circular segments appear as regions of sustained elevated speed, but slow enough to be perceptible.

The segments with higher velocity are not necessarily the most significant. For instance, preparation and accomodation segments occur at the beggining and end of signing, or between components of a compound sign.

This pattern is also observed in signs with double repetition, where are typically three peaks: the two lexical strikes and an accomodation segment in between. Remarkably, this structure is highly consistent across different languages.

MM sign in LSE: “NEVER”. Between the two lexical segments, an accomodation segment can be seen.

A compound sign in LSE: “FIREFIGHTER”. There is a brief segment, which is either a hold or a small movement for contact, for the helment, and then a circular movement for the hose.

Another MM sign in LSE: “COUSIN”. The often called “two syllables” here are similar to those in “NEVER”, but not to the ones in “FIREFIGHTER”.