System and method for quality assurance of media analysis

1. A system for determining quality level of labeling of a plurality of multimedia data items, wherein the labels relate to reputation polarity, wherein reputation polarity comprises determining an impact of a statement on a reputation of a brand, perception of a brand, or a combination thereof; wherein said brand comprises a person, company, organization or other identifiable entity; the system comprising a plurality of worker computational devices for labeling the items, each worker computational device comprising a processor for executing instructions for labeling the items and a memory for storing the instructions; an analysis computational device for analyzing the labels to determine accuracy by calculating cross polarity accuracy scores and neutral bias scores for a plurality of labels for each item, and for providing feedback through said worker computational devices to increase accuracy of labeling; wherein said cross polarity accuracy score is determined at least partially according to a number of correct labels determined at each worker computational device; wherein said neutral bias score is at least partially determined according to a number of correct reputation polarity labels determined at each worker computational device; wherein said analysis computational device comprises an analysis processor for executing instructions for analyzing the labels and for providing feedback, and an analysis memory for storing the instructions; and a computer network for communication between said computational devices;

the system further comprising a model computational device for being trained on labeled multimedia data items to train a model to automatically determine reputation polarity by classifying an unlabeled multimedia data item, wherein said model computational device comprises a model processor for executing a plurality of instructions for automatically training a model to automatically determine reputation polarity of an unlabeled multimedia data item, and a model memory for storing said instructions; wherein said model to automatically determine reputation polarity receives said plurality of labeled multimedia data items and classifies said multimedia data items according to reputation polarity; wherein an error in correctly classifying said multimedia data items by said model is calculated and is used to adjust said model; wherein said training is repeated until a sufficiently low error is reached.

2. The system of claim 1 , wherein said machine learning model is selected from the group consisting of recurrent and convolutional neural networks.

3. The system of claim 1 , wherein said instructions comprise instructions for determining an accuracy of said labels on a numerical scale, wherein said feedback comprises a number on said numerical scale.

4. The system of claim 1 , wherein said instructions for automatically training said model to form a trained model comprise instructions for training a model selected from the group consisting of recurrent and convolutional neural networks.

5. The system of claim 1 , wherein said trained model comprises a network of neurons and wherein a layer of the network contains a plurality of neurons with linear activation, each neuron relating to a specific class, wherein a softmax determines a final class selection.

6. The system of claim 5 , wherein said plurality of classes relates to a sentiment of a social media message, and wherein said plurality of classes comprises negative, neutral, and positive sentiment; wherein said sentiment comprises contribution of said social media message to polarity of reputation.

7. The system of claim 1 , wherein said processor of each computational device comprises a hardware processor configured to perform a predefined set of basic operations in response to receiving a corresponding basic instruction selected from a predefined native instruction set of codes, and wherein said analysis computational device comprises a first set of machine codes selected from the native instruction set for receiving said labels for each item, a second set of machine codes selected from the native instruction set for analyzing said labels to determine if said labels are correct, and a third set of machine codes selected from the native instruction set for providing feedback on correctness of said labels.

8. The system of claim 7 , wherein each worker computational device comprises a first set of machine codes selected from the native instruction set for receiving said items and a second set of machine codes selected from the native instruction set for supporting labeling of said items.

9. The system of claim 8 , wherein said supporting labeling of said items comprises providing a display to a user operating said worker computational device to display each item and an input device for receiving a label for each item.

10. The system of claim 8 , wherein said supporting labeling of said items comprises automatically labeling said items according to execution of a third set of machine codes.

11. The system of claim 1 , wherein each worker computational device further comprises a display and a plurality of instructions stored in said memory, for receiving and displaying at least one multimedia data item from the training set for labeling, and for providing a label for the at least one multimedia data item according to a reputation polarity of the at least one multimedia data item.

12. The system of claim 11 , further comprising a monitoring computational device, wherein said monitoring computational device is in communication with said computer network, wherein said monitoring computational device comprises a processor for executing instructions and a memory for storing said instructions, wherein said instructions comprise instructions for automatically analyzing said labels of said multimedia data items from said worker computational devices and for automatically determining a correct label for training a model; and wherein said instructions comprise instructions for automatically determining the reputation polarity of the analysis set according to a trained model.

13. The system of claim 1 , wherein said multimedia data item comprises a social media message.

14. The system of claim 1 , wherein a plurality of said multimedia data items is obtained from a plurality of data streams and wherein a selection of multimedia data items from one of said data streams is increased for training said model to automatically determine reputation polarity.

15. The system of claim 1 , wherein said model to automatically determine reputation polarity is retrained at least intermittently on new labeled multimedia data items.

16. The system of claim 15 , wherein said model to automatically determine reputation polarity is retrained continuously on a plurality of new labeled multimedia data items.

17. The system of claim 1 , further comprising a data store for storing models;

wherein after training, said model is stored as a full package, said full package comprising inferred weights, model structure, word mappings, and metadata about quality of said model on validation and test data sets.

18. The system of claim 1 , wherein said training further comprises calculating an output error for each batch of multimedia data items and backpropagating said error throughout said model, followed by adjusting said weights in a direction of said error.

19. The system of claim 18 , further comprising applying dropout, wherein a percentage of said weights are randomly frozen for each batch of multimedia data items, and are not updated based on backpropagated error.

20. The system of claim 1 , wherein said training is performed in a two step process, wherein a document comprises a plurality of segments, wherein said multimedia data items are labeled at an individual segment level and at an overall document level;

wherein said model is first optimized to predict classification of individual segments;

wherein a multi-segment classifier is constructed with said model, wherein said multi-segment classifier is trained according to labeled documents.

21. The system of claim 20 , wherein said document comprises a plurality of sentences and wherein said plurality of segments comprises said plurality of sentences.

22. The system of claim 1 , wherein said multimedia data item comprises a blog post or a news item.

23. The system of claim 1 , further comprising a brand reputation monitor for preparing an analytical breakdown for determining an item selected from the group consisting of a time series by sentiment for a data stream, top authors by mention count, top sites by mention count, a word cloud for a given sentiment related to reputation polarity, an emoji cloud for a given sentiment related to reputation polarity.

24. The system of claim 1 , wherein said cross polarity accuracy is calculated according to a sum of a number of incorrectly assigned polarity labels, multiplied by a total number of assigned labels by the worker as negative or positive, and a total number of labels that are negative or positive, and then divided by the total number of labels that are negative or positive.

25. The system of claim 1 , wherein said neutral bias score is calculated according to a sum of a number of worker assigned labels that are neutral, multiplied by a total number of labels that are negative or positive, and divided by a total number of labels that are negative or positive.