Data augmentation and batch balancing for training multi-lingual model
A computer-implemented method includes: accessing a plurality of datasets, where each dataset of the plurality of datasets includes training examples; selecting datasets that include the training examples in a source language and a target language; and sampling, based on a sampling weight that is determined for each of the selected datasets, the training examples from the selected datasets to generate the training batches; training an ML model for performing at least a first task using the training examples of the training batches, by interleavingly inputting the training batches to the ML model; and outputting the trained ML model configured to perform the at least the first task on input utterances provided in at least one among the source language and the target language. The sampling weight is determined for each of the selected datasets based on one or more attributes common to the training examples of the selected dataset.
1 . A computer-implemented method comprising:
accessing a plurality of datasets, wherein each dataset of the plurality of datasets comprises a plurality of training examples;
generating training batches comprising at least some of the plurality of training examples, wherein the generating comprises:
selecting datasets among the plurality of datasets that comprise training examples in a source language and a target language, and
sampling, based on a sampling weight that is determined for each of the selected datasets, the training examples from the selected datasets to generate the training batches, wherein the sampling weight is determined for each of the selected datasets based on one or more attributes common to the training examples of the selected dataset, the one or more attributes including at least one among a language identifier, a translation source origin, and a presence of labels;
training a machine learning (ML) model for performing at least a first task using the training examples of the training batches, wherein the training comprises:
interleavingly inputting the training batches to the ML model,
evaluating, using a first loss, performance of at least the first task by the ML model on the training examples of the training batches, and
updating, using an optimization function, model parameters of the ML model based on the performance of the at least the first task; and
outputting the trained ML model configured to perform the at least the first task on input utterances provided in at least one among the source language and the target language,
wherein the plurality of datasets includes a labeled source language dataset comprising source language sentences in the source language and corresponding labels, as the training examples of the labeled source language dataset,
wherein the computer-implemented method further comprises:
prior to accessing the plurality of datasets, inputting into an ML translation model, the source language sentences and the corresponding labels, wherein the ML translation model is configured to translate the source language sentences from the source language into the target language, and
obtaining a labeled machine-translated target language dataset based on an output provided by the ML translation model, the labeled machine-translated target language dataset comprising machine-translated target language sentences in the target language and corresponding labels, as the training examples of the labeled machine-translated target language dataset, wherein the selecting the datasets further comprises:
selecting the labeled source language dataset,
selecting a labeled target language dataset comprising human-created target language sentences in the target language and corresponding labels, as the training examples of the labeled target language dataset, and
selecting the labeled machine-translated target language dataset, thereby augmenting the training examples of the labeled target language dataset with the training examples of the labeled machine-translated target language dataset to provide additional labeled training examples in the target language.
2 . The computer-implemented method of claim 1 , wherein:
the sampling the training examples comprises:
sampling the training examples from the labeled source language dataset, to generate one or more labeled source language training batches,
sampling the training examples from the labeled target language dataset, to generate one or more labeled target language training batches, and
sampling the training examples from the labeled machine-translated target language dataset, to generate one or more labeled machine-translated target language training batches, wherein the sampling weight for the labeled target language dataset is greater than the sampling weight for the labeled machine-translated target language dataset, which results in a greater number of the training examples sampled from the labeled target language dataset as compared to a number of the training examples sampled from the labeled machine-translated target language dataset, and
the ML model is trained using the training examples from the one or more labeled source language training batches, the one or more labeled target language training batches, and the one or more labeled machine-translated target language training batches.
3 . The computer-implemented method of claim 1 , wherein obtaining the labeled machine-translated target language dataset comprises:
identifying, in at least one of the source language sentences, a multi-word sequence in the source language;
identifying, in at least one of the machine-translated target language sentences that corresponds to the at least one of the source language sentences, respective words in the target language that correspond to the multi-word sequence;
determining whether the labels of the machine-translated target language sentences correctly identify the respective words in the target language; and
based on the determining that the labels of the machine-translated target language sentences do not correctly identify the respective words, correcting an alignment of the labels of the machine-translated target language sentences to correctly identify the respective words in the target language.
4 . The computer-implemented method of claim 1 , wherein:
the plurality of datasets further includes an unlabeled bilingual sentence pairs dataset comprising unlabeled bilingual sentence pairs in the source language and the target language, as the training examples of the unlabeled bilingual sentence pairs dataset,
the sampling the training examples comprises:
sampling the training examples from a set of labeled datasets comprising the labeled source language dataset, the labeled target language dataset, and the labeled machine-translated target language dataset, to generate one or more labeled training batches, each of the one or more labeled training batches comprising the training examples from one of the labeled source language dataset, the labeled target language dataset, or the labeled machine-translated target language dataset, and
sampling the training examples from the unlabeled bilingual sentence pairs dataset, to generate one or more unlabeled training batches, wherein the sampling weight for sampling the training examples from the set of labeled datasets is greater than the sampling weight for sampling the training examples from the unlabeled bilingual sentence pairs dataset, which results in a greater number of the training examples being sampled from the set of labeled datasets as compared to a number of the training examples sampled from the unlabeled bilingual sentence pairs dataset,
the ML model is trained using the training examples from at least one among (i) the one or more labeled training batches and (ii) the one or more unlabeled training batches, and
the computer-implemented method further comprises:
when the ML model is trained using the one or more of the labeled training batches, updating the model parameters by minimizing the first loss, which is a cross entropy loss function, and
when the ML model is trained using the one or more of the unlabeled training batches, updating the model parameters by minimizing a second loss, which is a Kullback-Leibler divergence function.
5 . The computer-implemented method of claim 1 , wherein:
the ML model is a multi-task model trained for performing the first task and a second task,
the first task is an aspect-based sentiment analysis (ABSA) task configured to predict a sentiment of the input utterances based on an aspect in the input utterances,
the second task is a sentence-level sentiment analysis (SLSA) task configured to predict a sentiment of the input utterances based on per sentence sentiment of the input utterances,
the sampling the training examples comprises:
sampling the training examples from the selected datasets to generate a set of ABSA training batches comprising ABSA training examples for training the ML model on the ABSA task, and
sampling the training examples from the selected datasets to generate a set of SLSA training batches comprising SLSA training examples for training the ML model on the SLSA task, and
the ML model is trained on the ABSA task using the ABSA training examples and on the SLSA task using the SLSA training examples, wherein the set of ABSA training batches and the set of SLSA training batches are interleavingly provided to the ML model.
6 . The computer-implemented method of claim 5 , further comprising:
when the ML model is trained using the training examples of the set of ABSA training batches, updating the model parameters by minimizing the first loss; and
when the ML model is trained using the training examples of the set of SLSA training batches, updating the model parameters by minimizing a second loss.
7 . A system comprising:
one or more processors; and
one or more computer-readable media storing instructions that, when executed by the one or more processors, cause the system to perform a method including:
accessing a plurality of datasets, wherein each dataset of the plurality of datasets comprises a plurality of training examples;
generating training batches comprising at least some of the plurality of training examples, wherein the generating includes:
selecting datasets among the plurality of datasets that comprise training examples in a source language and a target language, and
sampling, based on a sampling weight that is determined for each of the selected datasets, the training examples from the selected datasets to generate the training batches, wherein the sampling weight is determined for each of the selected datasets based on one or more attributes common to the training examples of the selected dataset, the one or more attributes including at least one among a language identifier, a translation source origin, and a presence of labels;
training a machine learning (ML) model for performing at least a first task using the training examples of the training batches, wherein the training includes:
interleavingly inputting the training batches to the ML model,
evaluating, using a first loss, performance of at least the first task by the ML model on the training examples of the training batches, and
updating, using an optimization function, model parameters of the ML model based on the performance of the at least the first task; and
outputting the trained ML model configured to perform the at least the first task on input utterances provided in at least one among the source language and the target language,
wherein the plurality of datasets includes a labeled source language dataset comprising source language sentences in the source language and corresponding labels, as the training examples of the labeled source language dataset,
wherein the method further includes:
prior to accessing the plurality of datasets, inputting into an ML translation model, the source language sentences and the corresponding labels, wherein the ML translation model is configured to translate the source language sentences from the source language into the target language, and
obtaining a labeled machine-translated target language dataset based on an output provided by the ML translation model, the labeled machine-translated target language dataset comprising machine-translated target language sentences in the target language and corresponding labels, as the training examples of the labeled machine-translated target language dataset,
wherein the selecting the datasets further includes:
selecting the labeled source language dataset,
selecting a labeled target language dataset comprising human-created target language sentences in the target language and corresponding labels, as the training examples of the labeled target language dataset, and
selecting the labeled machine-translated target language dataset, thereby augmenting the training examples of the labeled target language dataset with the training examples of the labeled machine-translated target language dataset to provide additional labeled training examples in the target language.
8 . The system of claim 7 , wherein:
the sampling the training examples includes:
sampling the training examples from the labeled source language dataset, to generate one or more labeled source language training batches,
sampling the training examples from the labeled target language dataset, to generate one or more labeled target language training batches, and
sampling the training examples from the labeled machine-translated target language dataset, to generate one or more labeled machine-translated target language training batches, wherein the sampling weight for the labeled target language dataset is greater than the sampling weight for the labeled machine-translated target language dataset, which results in a greater number of the training examples sampled from the labeled target language dataset as compared to a number of the training examples sampled from the labeled machine-translated target language dataset, and
the ML model is trained using the training examples from the one or more labeled source language training batches, the one or more labeled target language training batches, and the one or more labeled machine-translated target language training batches.
9 . The system of claim 7 , wherein obtaining the labeled machine-translated target language dataset includes:
identifying, in at least one of the source language sentences, a multi-word sequence in the source language;
identifying, in at least one of the machine-translated target language sentences that corresponds to the at least one of the source language sentences, respective words in the target language that correspond to the multi-word sequence;
determining whether the labels of the machine-translated target language sentences correctly identify the respective words in the target language; and
based on the determining that the labels of the machine-translated target language sentences do not correctly identify the respective words, correcting an alignment of the labels of the machine-translated target language sentences to correctly identify the respective words in the target language.
10 . The system of claim 7 , wherein:
the plurality of datasets further includes an unlabeled bilingual sentence pairs dataset comprising unlabeled bilingual sentence pairs in the source language and the target language, as the training examples of the unlabeled bilingual sentence pairs dataset,
the sampling the training examples includes:
sampling the training examples from a set of labeled datasets comprising the labeled source language dataset, the labeled target language dataset, and the labeled machine-translated target language dataset, to generate one or more labeled training batches, each of the one or more labeled training batches comprising the training examples from one of the labeled source language dataset, the labeled target language dataset, or the labeled machine-translated target language dataset, and
sampling the training examples from the unlabeled bilingual sentence pairs dataset, to generate one or more unlabeled training batches, wherein the sampling weight for sampling the training examples from the set of labeled datasets is greater than the sampling weight for sampling the training examples from the unlabeled bilingual sentence pairs dataset, which results in a greater number of the training examples being sampled from the set of labeled datasets as compared to a number of the training examples sampled from the unlabeled bilingual sentence pairs dataset,
the ML model is trained using the training examples from at least one among (i) the one or more labeled training batches and (ii) the one or more unlabeled training batches, and
the method further includes:
when the ML model is trained using the one or more of the labeled training batches, updating the model parameters by minimizing the first loss, which is a cross entropy loss function, and
when the ML model is trained using the one or more of the unlabeled training batches, updating the model parameters by minimizing a second loss, which is a Kullback-Leibler divergence function.
11 . The system of claim 7 , wherein:
the ML model is a multi-task model trained for performing the first task and a second task,
the first task is an aspect-based sentiment analysis (ABSA) task configured to predict a sentiment of the input utterances based on an aspect in the input utterances,
the second task is a sentence-level sentiment analysis (SLSA) task configured to predict a sentiment of the input utterances based on per sentence sentiment of the input utterances,
the sampling the training examples includes:
sampling the training examples from the selected datasets to generate a set of ABSA training batches comprising ABSA training examples for training the ML model on the ABSA task, and
sampling the training examples from the selected datasets to generate a set of SLSA training batches comprising SLSA training examples for training the ML model on the SLSA task, and
the ML model is trained on the ABSA task using the ABSA training examples and on the SLSA task using the SLSA training examples, wherein the set of ABSA training batches and the set of SLSA training batches are interleavingly provided to the ML model.
12 . The system of claim 11 , wherein the method further includes:
when the ML model is trained using the training examples of the set of ABSA training batches, updating the model parameters by minimizing the first loss; and
when the ML model is trained using the training examples of the set of SLSA training batches, updating the model parameters by minimizing a second loss.
13 . One or more non-transitory computer-readable media storing instructions that, when executed by one or more processors, cause the one or more processors to perform a method including:
accessing a plurality of datasets, wherein each dataset of the plurality of datasets comprises a plurality of training examples;
generating training batches comprising at least some of the plurality of training examples, wherein the generating includes:
selecting datasets among the plurality of datasets that comprise training examples in a source language and a target language, and
sampling, based on a sampling weight that is determined for each of the selected datasets, the training examples from the selected datasets to generate the training batches, wherein the sampling weight is determined for each of the selected datasets based on one or more attributes common to the training examples of the selected dataset, the one or more attributes including at least one among a language identifier, a translation source origin, and a presence of labels;
training a machine learning (ML) model for performing at least a first task using the training examples of the training batches, wherein the training includes:
interleavingly inputting the training batches to the ML model,
evaluating, using a first loss, performance of at least the first task by the ML model on the training examples of the training batches, and
updating, using an optimization function, model parameters of the ML model based on the performance of the at least the first task; and
outputting the trained ML model configured to perform the at least the first task on input utterances provided in at least one among the source language and the target language-,
wherein the plurality of datasets includes a labeled source language dataset comprising source language sentences in the source language and corresponding labels, as the training examples of the labeled source language dataset,
wherein the method further includes:
prior to accessing the plurality of datasets, inputting into an ML translation model, the source language sentences and the corresponding labels, wherein the ML translation model is configured to translate the source language sentences from the source language into the target language, and
obtaining a labeled machine-translated target language dataset based on an output provided by the ML translation model, the labeled machine-translated target language dataset comprising machine-translated target language sentences in the target language and corresponding labels, as the training examples of the labeled machine-translated target language dataset, wherein the selecting the datasets further includes:
selecting the labeled source language dataset,
selecting a labeled target language dataset comprising human-created target language sentences in the target language and corresponding labels, as the training examples of the labeled target language dataset, and
selecting the labeled machine-translated target language dataset, thereby augmenting the training examples of the labeled target language dataset with the training examples of the labeled machine-translated target language dataset to provide additional labeled training examples in the target language.
14 . The one or more non-transitory computer-readable media of claim 13 , wherein:
the sampling the training examples includes:
sampling the training examples from the labeled source language dataset, to generate one or more labeled source language training batches,
sampling the training examples from the labeled target language dataset, to generate one or more labeled target language training batches, and
sampling the training examples from the labeled machine-translated target language dataset, to generate one or more labeled machine-translated target language training batches, wherein the sampling weight for the labeled target language dataset is greater than the sampling weight for the labeled machine-translated target language dataset, which results in a greater number of the training examples sampled from the labeled target language dataset as compared to a number of the training examples sampled from the labeled machine-translated target language dataset, and
the ML model is trained using the training examples from the one or more labeled source language training batches, the one or more labeled target language training batches, and the one or more labeled machine-translated target language training batches.
15 . The one or more non-transitory computer-readable media of claim 13 , wherein obtaining the labeled machine-translated target language dataset includes:
identifying, in at least one of the source language sentences, a multi-word sequence in the source language;
identifying, in at least one of the machine-translated target language sentences that corresponds to the at least one of the source language sentences, respective words in the target language that correspond to the multi-word sequence;
determining whether the labels of the machine-translated target language sentences correctly identify the respective words in the target language; and
based on the determining that the labels of the machine-translated target language sentences do not correctly identify the respective words, correcting an alignment of the labels of the machine-translated target language sentences to correctly identify the respective words in the target language.
16 . The one or more non-transitory computer-readable media of claim 13 , wherein:
the plurality of datasets further includes an unlabeled bilingual sentence pairs dataset comprising unlabeled bilingual sentence pairs in the source language and the target language, as the training examples of the unlabeled bilingual sentence pairs dataset,
the sampling the training examples includes:
sampling the training examples from a set of labeled datasets comprising the labeled source language dataset, the labeled target language dataset, and the labeled machine-translated target language dataset, to generate one or more labeled training batches, each of the one or more labeled training batches comprising the training examples from one of the labeled source language dataset, the labeled target language dataset, or the labeled machine-translated target language dataset, and
sampling the training examples from the unlabeled bilingual sentence pairs dataset, to generate one or more unlabeled training batches, wherein the sampling weight for sampling the training examples from the set of labeled datasets is greater than the sampling weight for sampling the training examples from the unlabeled bilingual sentence pairs dataset, which results in a greater number of the training examples being sampled from the set of labeled datasets as compared to a number of the training examples sampled from the unlabeled bilingual sentence pairs dataset,
the ML model is trained using the training examples from at least one among (i) the one or more labeled training batches and (ii) the one or more unlabeled training batches, and
the method further includes:
when the ML model is trained using the one or more of the labeled training batches, updating the model parameters by minimizing the first loss, which is a cross entropy loss function, and
when the ML model is trained using the one or more of the unlabeled training batches, updating the model parameters by minimizing a second loss, which is a Kullback-Leibler divergence function.
17 . The one or more non-transitory computer-readable media of claim 13 , wherein:
the ML model is a multi-task model trained for performing the first task and a second task,
the first task is an aspect-based sentiment analysis (ABSA) task configured to predict a sentiment of the input utterances based on an aspect in the input utterances,
the second task is a sentence-level sentiment analysis (SLSA) task configured to predict a sentiment of the input utterances based on per sentence sentiment of the input utterances,
the sampling the training examples includes:
sampling the training examples from the selected datasets to generate a set of ABSA training batches comprising ABSA training examples for training the ML model on the ABSA task, and
sampling the training examples from the selected datasets to generate a set of SLSA training batches comprising SLSA training examples for training the ML model on the SLSA task, and
the ML model is trained on the ABSA task using the ABSA training examples and on the SLSA task using the SLSA training examples, wherein the set of ABSA training batches and the set of SLSA training batches are interleavingly provided to the ML model, and
the method further includes:
when the ML model is trained using the training examples of the set of ABSA training batches, updating the model parameters by minimizing the first loss, and
when the ML model is trained using the training examples of the set of SLSA training batches, updating the model parameters by minimizing a second loss.
18 . The computer-implemented method of claim 1 , wherein the generating the training batches comprises implementing a hierarchical weighted batch balancing method.