System for fine-grained affective states understanding and prediction

1. A system configured to predict fine-grained affective states, the system comprising a processor configured to execute instructions to:

create training data comprising content conveying emotions;

combine all sentences from the content conveying emotions that express a same emotional state together into a plurality of emotional documents, each emotional document comprising sentences that express the same emotional state;

assign an identical emotion identifier to each of the sentences expressing the same emotional state in the plurality of emotional documents;

create a trained model by performing a training process using the training data to train a model using a feed forward neural network, wherein each of the emotional documents in the plurality of emotional documents forms an input layer of the feed forward neural network during the emotion vector space model training process and each unit in the input layer corresponds to a single emotion identifier, and wherein the training process converts discrete emotions into emotion vector representations and converts words found in the sentences within a window size of the same emotional state to word vectors, and wherein the training process co-trains the word vectors in an interleaved fashion while training the emotion representations to place both the emotion vectors and the word vectors into the same vector space; and

use the trained model to predict fine-grained affective states for text conveying an emotion.

2. The system of claim 1 , wherein the processor further executes instructions to apply dimension reduction and rotation techniques on the trained model to identify a principle dimension of a set of fine-grained affective states.

3. The system of claim 1 , wherein the processor further executes instructions to apply clustering techniques on the trained model to identify basic emotions of a set of fine-grained affective states.

4. The system of claim 1 , wherein the processor further executes instructions to generate domain-specific emotional taxonomy by aggregating predicted results.

5. The system of claim 1 , wherein the process further comprises projecting the emotion vector representations of the fine-grained affective states into a distributed vector space, wherein similar fine-grained affective states are close to each other in the distributed vector space.

6. The system of claim 1 , wherein creating the training data comprising content conveying emotions includes:

performing a keyword search to extract social media posts containing an emotion keyword for a particular timeframe;

apply filtering heuristics to remove duplicates and spam postings;

construct dependency relations between words appearing in a social media post;

apply soft labeling for creating an automatically labeled emotion dataset; and

delete all emotions with less than a predetermined number of appearances from the automatically labeled emotion dataset.

7. A computer program product for predicting fine-grained affective states, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processor to:

create training data comprising content conveying emotions;

combine all sentences from the content conveying emotions that express a same emotional state together into a plurality of emotional documents, each emotional document comprising sentences that express the same emotional state;

assign an identical emotion identifier to each of the sentences expressing the same emotional state in the plurality of emotional documents;

create a trained model by performing a process using the training data to train a model using a feed forward neural network, wherein each of the emotional documents in the plurality of emotional documents forms an input layer of the feed forward neural network during the process and each unit in the input layer corresponds to a single emotion identifier, and wherein the process converts discrete emotions into emotion vector representations and converts words found in the sentences within a window size of the same emotional state to word vectors, and wherein the training process co-trains the word vectors in an interleaved fashion while training the emotion representations to place both the emotion vectors and the word vectors into the same vector space; and

use the trained model to predict fine-grained affective states for text conveying an emotion.

8. The computer program product of claim 7 , wherein the processor further executes the program instructions to apply dimension reduction and rotation techniques on the trained model to identify a principle dimension of a set of fine-grained affective states.

9. The computer program product of claim 7 , wherein the processor further executes the program instructions to apply clustering techniques on the trained model to identify basic emotions of a set of fine-grained affective states.

10. The computer program product of claim 7 , wherein the processor further executes the program instructions to generate domain-specific emotional taxonomy by aggregating predicted results.

11. The computer program product of claim 7 , wherein the process further comprises projecting the emotion vector representations of the fine-grained affective states into a distributed vector space, wherein similar fine-grained affective states are close to each other in the distributed vector space.