Detecting affective characteristics of text with gated convolutional encoder-decoder framework

1. A method for identifying an affect label of text with a gated convolutional encoder-decoder model, wherein the method includes one or more processing devices performing operations comprising:

receiving, at a gated convolutional encoder, input text;

encoding, by the gated convolutional encoder, the input text to generate a latent representation of the input text;

receiving, at a supervised classification engine, extracted linguistic features of the input text and the latent representation of the input text;

predicting, by the supervised classification engine, an affect characterization of the input text using the extracted linguistic features and the latent representation; and

identifying, by the gated convolutional encoder-decoder model, an affect label of the input text using the predicted affect characterization, wherein the gated convolutional encoder-decoder model is jointly trained using a weighted auto-encoder loss associated with a reconstruction engine and a weighted classification loss associated with the supervised classification engine.

2. The method of claim 1 , further comprising pre-training the reconstruction engine by:

encoding, by the gated convolutional encoder, training data to generate an initial latent representation of the training data; and

reducing an initial auto-encoder loss associated with decoding the initial latent representation of the training data, wherein the initial auto-encoder loss is based on a comparison between a decoded initial latent representation of the training data and the training data.

3. The method of claim 1 , further comprising jointly training the gated convolutional encoder-decoder model by:

encoding, by the gated convolutional encoder, training data to generate an initial latent representation of the training data;

training, by the reconstruction engine, the gated convolutional encoder-decoder model to reduce the weighted auto-encoder loss associated with decoding the initial latent representation of the training data; and

training, by the supervised classification engine, the gated convolutional encoder-decoder model to reduce the weighted classification loss associated with predicting an affect classification of the training data, wherein jointly training the gated convolutional encoder-decoder model comprises balancing a first weight attributed to the weighted auto-encoder loss and a second weight attributed to the weighted classification loss.

4. The method of claim 1 , wherein predicting the affect characterization of the input text using the extracted linguistic features and the latent representation comprises:

extracting the extracted linguistic features from the input text into a linguistic feature vector representation;

normalizing and concatenating the linguistic feature vector representation with the latent representation to generate an appended latent representation; and

providing the appended latent representation to a set of fully connected layers with a Softmax classifier loss layer to generate the predicted affect characterization of the input text.

5. The method of claim 4 , wherein the extracted linguistic features comprise lexical features, syntactic features, derived features, and psycholinguistic features of the input text.

6. The method of claim 1 , wherein the gated convolutional encoder-decoder model comprises a supervised component, a gated architecture, a pre-training component, a joint training component, and a linguistic feature component.

7. The method of claim 1 , further comprising:

generating, by the gated convolutional encoder, a training latent representation of training data, wherein jointly training the gated convolutional encoder-decoder model is performed using the training latent representation of the training data.

8. The method of claim 1 , wherein jointly training the gated convolutional encoder-decoder model using the weighted classification loss comprises:

extracting training linguistic features from training data into a training linguistic feature vector representation;

encoding the training data to generate a training latent representation of the training data;

normalizing and concatenating the training linguistic feature vector representation with the training latent representation to generate a training appended latent representation;

providing the training appended latent representation to a set of fully connected layers with a Softmax classifier loss layer to generate a predicted affect classification of the training data;

comparing the predicted affect classification of the training data to a ground-truth label of the training data; and

training the gated convolutional encoder-decoder model using the weighted classification loss based on a comparison between the predicted affect classification of the training data and the ground-truth label of the training data.

9. A computing system comprising:

means for receiving, at a gated convolutional encoder, input text;

means for encoding, by the gated convolutional encoder, the input text to generate a latent representation of the input text;

means for receiving, at a supervised classification engine, extracted linguistic features of the input text and the latent representation of the input text;

means for predicting, by the supervised classification engine, an affect characterization of the input text using the extracted linguistic features and the latent representation; and

means for identifying, by a gated convolutional encoder-decoder model, an affect label of the input text using the predicted affect characterization, wherein the gated convolutional encoder-decoder model is jointly trained using a weighted auto-encoder loss associated with a reconstruction engine and a weighted classification loss associated with the supervised classification engine.

10. The computing system of claim 9 , further comprising:

means for pre-training the reconstruction engine by:

encoding, by the gated convolutional encoder, training data to generate an initial latent representation of the training data; and

reducing an initial auto-encoder loss associated with decoding the initial latent representation of the training data, wherein the initial auto-encoder loss is based on a comparison between a decoded initial latent representation of the training data and the training data.

11. The computing system of claim 9 , further comprising:

means for jointly training the gated convolutional encoder-decoder model by:

encoding, by the gated convolutional encoder, training data to generate an initial latent representation of the training data;

training, by the reconstruction engine, the gated convolutional encoder-decoder model to reduce the weighted auto-encoder loss associated with decoding the initial latent representation of the training data; and

training, by the supervised classification engine, the gated convolutional encoder-decoder model to reduce the weighted classification loss associated with predicting an affect classification of the training data, wherein jointly training the gated convolutional encoder-decoder model comprises balancing a first weight attributed to the weighted auto-encoder loss and a second weight attributed to the weighted classification loss.

12. The computing system of claim 9 , wherein the means for predicting the affect characterization of the input text using the extracted linguistic features and the latent representation comprises:

means for extracting the extracted linguistic features from the input text into a linguistic feature vector representation;

means for normalizing and concatenating the linguistic feature vector representation with the latent representation to generate an appended latent representation; and

means for providing the appended latent representation to a set of fully connected layers with a Softmax classifier loss layer to generate the predicted affect characterization of the input text.

13. The computing system of claim 12 , wherein the extracted linguistic features comprise lexical features, syntactic features, derived features, and psycholinguistic features of the input text.

14. The computing system of claim 9 , wherein the gated convolutional encoder-decoder model comprises a supervised component, a gated architecture, a pre-training component, a joint training component, and a linguistic feature component.

15. The computing system of claim 9 , further comprising:

means for generating, by the gated convolutional encoder, a training latent representation of training data, wherein jointly training the gated convolutional encoder-decoder model is performed using the training latent representation of the training data.

16. A non-transitory computer-readable medium having instructions stored thereon, the instructions executable by a processing device to perform operations comprising:

receiving, at a gated convolutional encoder, input text;

encoding, by the gated convolutional encoder, the input text to generate a latent representation of the input text;

receiving, at a supervised classification engine, extracted linguistic features of the input text and the latent representation of the input text;

predicting, by the supervised classification engine, an affect characterization of the input text using the extracted linguistic features and the latent representation; and

identifying, by a gated convolutional encoder-decoder model, an affect label of the input text using the predicted affect characterization, wherein the gated convolutional encoder-decoder model is jointly trained using a weighted auto-encoder loss associated with a reconstruction engine and a weighted classification loss associated with the supervised classification engine.

17. The non-transitory computer-readable medium of claim 16 , wherein jointly training the gated convolutional encoder-decoder model using the weighted classification loss comprises:

extracting training linguistic features from training data into a training linguistic feature vector representation;

encoding the training data to generate a training latent representation of the training data;

normalizing and concatenating the training linguistic feature vector representation with the training latent representation to generate a training appended latent representation;

providing the training appended latent representation to a set of fully connected layers with a Softmax classifier loss layer to generate a predicted affect classification of the training data;

comparing the predicted affect classification of the training data to a ground-truth label of the training data; and

training the gated convolutional encoder-decoder model using the weighted classification loss based on a comparison between the predicted affect classification of the training data and the ground-truth label of the training data.

18. The non-transitory computer-readable medium of claim 16 , the instructions further executable by the processing device to perform operations comprising:

pre-training the reconstruction engine by:

encoding, by the gated convolutional encoder, training data to generate an initial latent representation of the training data; and

reducing an initial auto-encoder loss associated with decoding the initial latent representation of the training data, wherein the initial auto-encoder loss is based on a comparison between a decoded initial latent representation of the training data and the training data.

19. The non-transitory computer-readable medium of claim 16 , the instructions further executable by the processing device to perform operations comprising:

jointly training the gated convolutional encoder-decoder model by:

encoding, by the gated convolutional encoder, training data to generate an initial latent representation of the training data;

training, by the reconstruction engine, the gated convolutional encoder-decoder model to reduce the weighted auto-encoder loss associated with decoding the initial latent representation of the training data; and

training, by the supervised classification engine, the gated convolutional encoder-decoder model to reduce the weighted classification loss associated with predicting an affect classification of the training data, wherein jointly training the gated convolutional encoder-decoder model comprises balancing a first weight attributed to the weighted auto-encoder loss and a second weight attributed to the weighted classification loss.

20. The non-transitory computer-readable medium of claim 16 , wherein instructions executable by the processing device to perform operations comprising predicting the affect characterization of the input text using the extracted linguistic features and the latent representation comprise:

extracting the extracted linguistic features from the input text into a linguistic feature vector representation;

normalizing and concatenating the linguistic feature vector representation with the latent representation to generate an appended latent representation; and

providing the appended latent representation to a set of fully connected layers with a Softmax classifier loss layer to generate the predicted affect characterization of the input text.