Sentiment analysis of content using expression recognition

1. A computerized-method for providing a sentiment score by evaluating expressions of participants during a video meeting, said computerized-method comprising:

in a computerized system comprising a processor, a database of video meeting recordings;

a memory to store the database, said processor is configured to operate a Sentiment Analysis (SA) module, said SA module is configured to:

retrieving one or more recordings of a video meeting from the database of video meeting recordings of each participant in the video meeting and associating the one or more recordings with a participant;

dividing each retrieved recording into segments;

processing the segments in a Facial Expression Recognition (FER) system to associate each segment with a timestamped sequence of expressions for each participant in the video meeting; and

processing each segment in an Artificial Neural Network (ANN) having a dense layer, by applying a prebuilt and pretrained deep learning model, to yield the sentiment score for each statement for each participant; wherein the sentiment score for each statement for each participant in the video meeting is calculated by:

retrieving the timestamped sequence of expressions for the participant that is associated with the statement,

attributing a preconfigured sentiment score for the timestamped sequence of expressions for the participant to yield the sentiment score of a statement for the participant.

2. The computerized method of claim 1 , wherein the prebuilt deep learning model comprising a building stage comprising:

receiving one or more timestamped sequences of expressions;

tokenizing each timestamped sequence of expressions and then converting each tokenized timestamped sequence of expressions into a word vector;

embedding the word vector into a matrix of a preconfigured size vectors;

converting each expression in the word vector into an integer to yield a numerical vector;

padding the numerical vector into a preconfigured size;

embedding the numerical vector into a matrix of vectors, wherein each vector in the matrix of vectors, represents a number in the numerical vector, and wherein for each number in the numerical vector assigning a preconfigured number of random weights to yield a matrix of a predefined size vectors;

providing each vector of the embedded matrix to a deep learning system to yield a final vector in a preconfigure size;

providing the final vector to the dense layer to receive one or more neurons which indicate a sentiment score for each statement for each participant; and

converting each indicated sentiment score into ‘0’ or ‘1’ according to a predefined threshold.

3. The computerized method of claim 2 , wherein the dense layer comprising one or more neural layers,

wherein each neural layer is applying an activation function.

4. The computerized method of claim 2 , wherein a loss function and an optimizer to reduce the loss function to minimum by reducing an error between an actual value and a predicted value of the deep learning model are applied on the dense layer.

5. The computerized method of claim 2 , wherein the deep learning system is selected from: a Recurrent Neural Network (RNN) system or Gated Recurrent Unit (GRU) system or Long Short Term Memory (LSTM).

6. The computerized method of claim 2 , wherein the embedding of the word vector is performed by:

converting each expression in the word vector into an integer to yield a numerical vector;

padding the numerical vector by generating random numbers to fill up the numerical vector of a preconfigured size and constantly updating these numbers, by using backpropagation into a preconfigured size;

embedding the numerical vector into a matrix of vectors, wherein each vector in the matrix of vectors, represents a number in the numerical vector, and wherein for each number in the numerical vector assigning a preconfigured number of random weights to yield a matrix of a predefined size vectors.

7. The computerized method of claim 2 , wherein the pretrained deep learning model comprises a training stage comprising:

running the deep learning model with some proportionate values of batch size and epochs of video meeting recordings having an audio;

converting the audio of the video meeting into one or more text statements;

dividing each retrieved recording into segments;

processing the segments in a Facial Expression Recognition (FER) system to associate every segment with a timestamped sequence of expressions;

associating each of the one or more text statements with a segment of the video meeting recording;

processing each segment in a prebuilt deep learning model having a dense layer, to yield a sentiment score for each statement;

comparing the yielded sentiment score with a preconfigured sentiment score and adjusting weights of each neural layer of the ANN, accordingly.

8. The computerized method of claim 1 , wherein a sentiment score of a statement for all participants in the video meeting is calculated by:

adding the sentiment score of a statement for each participant to yield a total sentiment score for a statement,

dividing the total sentiment score for a statement by a number of all the participants in the video meeting, to yield a score, and

comparing the score with a predefined threshold to yield a final sentiment score for a statement for all participants.

9. The computerized method of claim 8 , wherein an overall sentiment score of a video meeting is calculated by:

adding the final sentiment score for a statement for all participants for each statement to yield a total sentiment score for all statements,

dividing the total sentiment score for all statement by a number of all statements in the video meeting to yield a score,

comparing the score with a predefined threshold to yield an overall sentiment score for a video meeting.

10. The computerized method of claim 1 , wherein the FER system is implemented in an open source computer vision and machine learning software library.

11. The computerized method of claim 10 , wherein the open source computer vision and machine learning software library is an Open Source Computer Vision (OpenCV) Library.

12. The computerized method of claim 1 , wherein the activation function is ‘sigmoid’ to yield a single neuron or ‘softmax’ to yield multiple neurons for more than two distinct sentiments.