System(s) and method(s) for training a sign language captioning model and subsequent use thereof
Implementations are directed to training and subsequently utilizing a sign language captioning model. Initially, processor(s) of a system can obtain a plurality of training instances that are generated based on processing sign language video content, sign language conversations, etc. Each of the plurality of training instances can include at least corresponding sign language feature tokens for a sign language video content segment, ground truth caption tokens associated with ground truth sign language captions for the sign language video content segment and ground truth timestamp tokens that align the ground truth sign language captions with respect to the sign language video content segment. Further, the processor(s), can train the sign language captioning model based on the plurality of training instances, and can cause the sign language captioning model to be deployed in an offline manner and/or in an online manner for processing sign language content.
1 . A method implemented by one or more processors, the method comprising:
obtaining a plurality of training instances for training a sign language captioning model,
each of the plurality of training instances including a corresponding training instance input and a corresponding training instance output,
the corresponding training instance inputs including at least corresponding sign language feature tokens for a sign language video content segment, and
corresponding alignment indicators that indicate whether ground truth sign language captions for the sign language video content segment are well-aligned or misaligned; and
the corresponding training instance outputs including ground truth caption tokens associated with the ground truth sign language captions for the sign language video content segment and ground truth timestamp tokens that align the ground truth sign language captions with respect to the sign language video content segment;
training, based on the plurality of training instances, the sign language captioning model, wherein training the sign language captioning model based on a given training instance, of the plurality of training instances, comprises:
processing, using the sign language captioning model, at least the corresponding sign language feature tokens for the sign language video content segment and the corresponding alignment indicator, included in the corresponding training instance input for the given training instance, to generate sign language captioning model output;
determining, based on the sign language captioning model output, predicted caption tokens associated with predicted sign language captions for the sign language video content segment and predicted timestamp tokens that are predicted to align the predicted sign language captions with respect to the sign language video content segment;
generating, based on a comparison of (i) the predicted caption tokens associated with the predicted sign language captions for the sign language video content segment and the ground truth caption tokens associated with the ground truth sign language captions for the sign language video content segment, and/or (ii) the predicted timestamp tokens that are predicted to align the predicted sign language captions with respect to the sign language video content segment and the ground truth timestamp tokens that align the ground truth sign language captions with respect to the sign language video content segment, one or more losses; and
updating, based on one or more of the losses, the sign language captioning model; and
subsequent to training the sign language captioning model:
causing the sign language captioning model to be deployed.
2 . The method of claim 1 , wherein the corresponding training instance inputs further include corresponding current caption tokens for current captions from the sign language video content segment.
3 . The method of claim 1 , wherein the corresponding training instance inputs further include corresponding previous caption tokens for previous captions from a previous sign language video content segment that precedes the sign language video content segment.
4 . The method of claim 1 , wherein the corresponding training instance inputs further include corresponding next caption tokens for next captions from a next sign language video content segment that follows the sign language video content segment.
5 . The method of claim 1 , wherein the corresponding sign language feature tokens for the sign language video content segment comprise one or more of: a corresponding video embedding for the sign language video content segment, corresponding image embeddings for the sign language video content segment, or corresponding vectors for skeletonized representations of the sign language video content segment.
6 . The method of claim 1 , wherein the corresponding training instance outputs further include a ground truth language token associated with a ground truth language for the sign language video content segment.
7 . The method of claim 6 , further comprising:
determining, based on the sign language captioning model output, a predicted language token associated with a predicted language for the sign language video content segment,
wherein generating one or more of the losses is further based on a comparison of: (iii) the predicted language token associated with the predicted language for the sign language video content segment and the ground truth language token associated with the ground truth language for the sign language video content segment.
8 . The method of claim 1 , wherein the corresponding training instance inputs further include corresponding instructions for the sign language captioning model to determine the predicted sign language captions for the sign language video content segment and to determine the predicted timestamp tokens that are predicted to align the predicted sign language captions with respect to the sign language video content segment.
9 . The method of claim 1 , further comprising:
prior to obtaining the plurality of training instances for training the sign language captioning model:
generating the plurality of training instances for training the sign language captioning model.
10 . The method of claim 9 , wherein generating the given training instance, of the plurality of training instances, comprises:
obtaining sign language video content that includes a plurality of signs being performed by a user and a caption track for the plurality of signs being performed by the user;
segmenting the sign language video content into a plurality of sign language video content segments,
the plurality of sign language video content segments including the sign language video content segment for the given training instance, and
the sign language video content segment including a corresponding subset of the plurality of signs being performed by the user, in the sign language video content, and a corresponding subset of the caption track for the plurality of signs being performed by the user;
determining, based on processing the corresponding subset of the plurality of signs being performed by the user in the sign language video content segment for the given training instance, the corresponding sign language feature tokens for the sign language video content segment and for the given training instance; and
determining, based on processing the corresponding subset of the caption track for the plurality of signs being performed by the user in the sign language video content segment for the given training instance, the ground truth caption tokens associated with the ground truth sign language captions for the sign language video content segment and the ground truth timestamp tokens that align the ground truth sign language captions with respect to the sign language video content segment.
11 . The method of claim 10 , wherein the caption track for the plurality of signs being performed by the user comprises corresponding captions associated with the plurality of signs and corresponding caption timestamps for the corresponding captions associated with the plurality of signs.
12 . The method of claim 1 , wherein causing the sign language captioning model to be deployed is in response to determining that one or more conditions are satisfied.
13 . The method of claim 12 , wherein the one or more conditions comprise one or more of: whether the sign language captioning model has been updated based on a threshold quantity of training instances, or whether performance of the sign language captioning model satisfies a threshold quality of performance.
14 . The method of claim 1 , wherein causing the sign language captioning model to be deployed comprises:
identifying newly added sign language video content that is newly added to a repository of sign language video content;
processing, using the sign language captioning model, the newly added sign language video content to determine a timestamped caption track associated with the newly added sign language video content; and
storing, in association with the newly added sign language video content, the timestamped caption track.
15 . The method of claim 14 , further comprising:
in response to receiving a request for playback of the newly added sign language video content:
causing the timestamped caption track to be played back along with the playback of the newly added sign language video content.
16 . The method of claim 1 , wherein causing the sign language captioning model to be deployed comprises:
identifying an ongoing conversation between a given user of a client device and an automated assistant;
processing, using the sign language captioning model, vision data that captures a plurality of sign language signs of the given user that are directed to the automated assistant and a dialog history of the ongoing conversation between the given user and the automated assistant to determine a timestamped caption track associated with the ongoing conversation; and
causing the timestamped caption track to be visually rendered for presentation to the given user, via a display of the client device, throughout the ongoing dialog.
17 . The method of claim 1 , wherein causing the sign language captioning model to be deployed comprises:
identifying an ongoing conversation between a given user and an additional user;
processing, using the sign language captioning model, vision data that captures a plurality of sign language signs of a given user that are directed to the additional user and a dialog history of the ongoing conversation between the given user and the additional user to determine a timestamped caption track associated with the ongoing conversation; and
causing the timestamped caption track to be visually rendered for presentation to the given user and/or the additional user throughout the ongoing dialog.
18 . A system comprising:
at least one processor; and
memory storing instructions that, when executed by the at least one processor, cause the at least one processor to be operable to:
obtain a plurality of training instances for training a sign language captioning model,
each of the plurality of training instances including a corresponding training instance input and a corresponding training instance output,
the corresponding training instance inputs including at least corresponding sign language feature tokens for a sign language video content segment, and
corresponding alignment indicators that indicate whether ground truth sign language captions for the sign language video content segment are well-aligned or misaligned; and
the corresponding training instance outputs including ground truth caption tokens associated with the ground truth sign language captions for the sign language video content segment and ground truth timestamp tokens that align the ground truth sign language captions with respect to the sign language video content segment;
train, based on the plurality of training instances, the sign language captioning model, wherein the instructions to train the sign language captioning model based on a given training instance, of the plurality of training instances, comprise instructions to:
process, using the sign language captioning model, at least the corresponding sign language feature tokens for the sign language video content segment and the corresponding alignment indicator, included in the corresponding training instance input for the given training instance, to generate sign language captioning model output;
determine, based on the sign language captioning model output, predicted caption tokens associated with predicted sign language captions for the sign language video content segment and predicted timestamp tokens that are predicted to align the predicted sign language captions with respect to the sign language video content segment;
generate, based on a comparison of (i) the predicted caption tokens associated with the predicted sign language captions for the sign language video content segment and the ground truth caption tokens associated with the ground truth sign language captions for the sign language video content segment, and/or (ii) the predicted timestamp tokens that are predicted to align the predicted sign language captions with respect to the sign language video content segment and the ground truth timestamp tokens that align the ground truth sign language captions with respect to the sign language video content segment, one or more losses; and
update, based on one or more of the losses, the sign language captioning model; and
subsequent to training the sign language captioning model:
cause the sign language captioning model to be deployed.
19 . A non-transitory computer-readable storage medium storing instructions that, when executed by at least one processor, cause the at least one processor to be operable to perform operations, the operations comprising:
obtaining a plurality of training instances for training a sign language captioning model,
each of the plurality of training instances including a corresponding training instance input and a corresponding training instance output,
the corresponding training instance inputs including at least corresponding sign language feature tokens for a sign language video content segment, and
corresponding alignment indicators that indicate whether ground truth sign language captions for the sign language video content segment are well-aligned or misaligned; and
the corresponding training instance outputs including ground truth caption tokens associated with the ground truth sign language captions for the sign language video content segment and ground truth timestamp tokens that align the ground truth sign language captions with respect to the sign language video content segment;
training, based on the plurality of training instances, the sign language captioning model, wherein training the sign language captioning model based on a given training instance, of the plurality of training instances, comprises:
processing, using the sign language captioning model, at least the corresponding sign language feature tokens for the sign language video content segment and the corresponding alignment indicator, included in the corresponding training instance input for the given training instance, to generate sign language captioning model output;
determining, based on the sign language captioning model output, predicted caption tokens associated with predicted sign language captions for the sign language video content segment and predicted timestamp tokens that are predicted to align the predicted sign language captions with respect to the sign language video content segment;
generating, based on a comparison of (i) the predicted caption tokens associated with the predicted sign language captions for the sign language video content segment and the ground truth caption tokens associated with the ground truth sign language captions for the sign language video content segment, and/or (ii) the predicted timestamp tokens that are predicted to align the predicted sign language captions with respect to the sign language video content segment and the ground truth timestamp tokens that align the ground truth sign language captions with respect to the sign language video content segment, one or more losses; and
updating, based on one or more of the losses, the sign language captioning model; and
subsequent to training the sign language captioning model:
causing the sign language captioning model to be deployed.