Method and system for validating media content
A method and a server for factually validating a content of a given snippet of a plurality of snippets are provided. The method comprises: identifying, within the plurality of snippets, at least one other snippet for the given snippet, the at least one other snippet providing context to the given snippet; generating, for the given snippet, a respective target vector; generating, for the at least one other snippet, a respective context vector; feeding both the respective target vector and the respective context vector to a Natural Language Processing (NLP) model to determine a probability value indicative of whether a content of the at least one other snippet confirms factual accuracy of that of the given snippet, in response to the probability value being equal to or greater than a threshold probability value, determining that the content of the given snippet is factually accurate.
1. A computer-implemented method for factually validating a content of a given snippet of a plurality of snippets, the plurality of snippets generated for a search object by a generative machine-learning model, the search object having been identified based on a respective search query, the method being executable by a server, the method comprising:
identifying, by the server, within the plurality of snippets, at least one other snippet for the given snippet,
the at least one other snippet providing context to the given snippet;
generating, by the server, for the given snippet, using a text embedding algorithm, a respective target vector representative of the given snippet;
generating, by the server, for the at least one other snippet, using the text embedding algorithm, a respective context vector representative of the context provided by the at least one other snippet;
feeding, by the server, both the respective target vector and the respective context vector to a Natural Language Processing (NLP) model to determine a probability value indicative of whether a content of the at least one other snippet confirms factual accuracy of a content of the given snippet,
the NLP model having been trained to determine whether in-use snippets confirm factual accuracy of the given snippet; and
in response to the probability value being equal to or greater than a threshold probability value, determining, by the server, that the content of the given snippet is factually accurate.
2. The method of claim 1 , wherein the generative model is configured to generate the plurality of snippets based on reference digital documents of a plurality of digital documents responsive to the respective search query.
3. The method of claim 2 , wherein:
a given digital document of the plurality of digital documents has been pre-assigned with a relevance parameter indicative of how relevant the given digital document to the respective search query is; and
the reference digital documents have been determined as being associated with top N highest respective relevance parameters.
4. The method of claim 1 , wherein the identifying the at least one other snippet for the given snippet comprises identifying at least one other snippet, content of which is semantically related to at least one fact in the given snippet.
5. The method of claim 1 , wherein the plurality of snippets generated by the generative machine-learning model form a cohesive text for the search object.
6. The method of claim 1 , wherein the generative machine-learning model is a Transformer-based machine-learning model.
7. The method of claim 1 , wherein the NLP model has been pre-trained to determine if two consecutive snippets of the plurality of snippets are semantically coherent therebetween.
8. The method of claim 7 , wherein the NLP model has further been fine-tuned to determine the probability value of whether the content of the at least one other snippet confirms factual accuracy of the content of the given snippet by applying to the NLP model an add-on machine-learning model, the add-on machine-learning model having been trained to generate vector representations for the NLP model, training the add-on machine-learning model comprising:
obtaining, by the server, a plurality of training objects, a given training object of the plurality of training objects including (i) a respective first training snippet; (ii) at least one training prompt having been generated based on the respective first training snippet; (iii) a respective second training snippet having been identified as being contextually related to the respective first training snippet; and (iv) a respective label indicative of whether a content of the respective second training snippet factually confirms a content of the respective first training snippet; and
feeding, by the server, the plurality of training objects to the add-on machine-learning model to train the add-on machine-learning model to generate (i) the respective target vector; and (ii) the respective context vector.
9. The method of claim 8 , wherein the at least one training prompt comprises a question which the respective first training snippet answers.
10. The method of claim 8 , wherein the at least one training prompt comprises a rephrasing of the respective first training snippet.
11. The method of claim 8 , wherein the respective label has been assigned by a human assessor.
12. The method of claim 8 , wherein the NLP model has been fine-tuned, by the server, with weights of nodes of the NLP model, determined after the NLP model was pre-trained, being frozen.
13. The method of claim 8 , wherein the add-on machine-learning model comprises a Long Short-Term Memory (LSTM) neural network.
14. The method of claim 8 , wherein fine-tuning of the NLP model comprises applying a P-tuning approach.
15. The method of claim 8 , wherein the NLP model comprises a Transformer-based machine-learning model.
16. A server for factually validating a content of a given snippet of a plurality of snippets, the plurality of snippets generated for a search object by a generative machine-learning model, the search object having been identified based on a respective search query, the server comprising a processor and a non-transitory computer-readable memory storing instructions, and the processor, upon executing the instructions, being configured to:
identify within the plurality of snippets, at least one other snippet for the given snippet,
the at least one other snippet providing context to the given snippet;
generate, for the given snippet, using a text embedding algorithm, a respective target vector representative of the given snippet;
generate, for the at least one other snippet, using the text embedding algorithm, a respective context vector representative of the context provided by the at least one other snippet;
feed both the respective target vector and the respective context vector to a Natural Language Processing (NLP) model to determine a probability value indicative of whether a content of the at least one other snippet confirms factual accuracy of a content of the given snippet,
the NLP model having been trained to determine whether in-use snippets confirm factual accuracy of the given snippet; and
in response to the probability value being equal to or greater than a threshold probability value, determine that the content of the given snippet is factually accurate.
17. The server of claim 16 , wherein to identify the at least one other snippet for the given snippet, the processor is configured to identify at least one other snippet, content of which is semantically related to at least one fact in the given snippet.
18. The server of claim 16 , wherein the NLP model has been pre-trained to determine if two consecutive snippets of the plurality of snippets are semantically coherent therebetween.
19. The server of claim 18 , wherein the NLP model has further been fine-tuned to determine the probability value of whether the content of the at least one other snippet confirms factual accuracy of the content of the given snippet by applying to the NLP model an add-on machine-learning model, and the processor is further configured to train the add-on machine-learning model to generate vector representations for the NLP model, by:
obtaining a plurality of training objects, a given training object of the plurality of training objects including (i) a respective first training snippet; (ii) at least one training prompt having been generated based on the respective first training snippet; (iii) a respective second training snippet having been identified as being contextually related to the respective first training snippet; and (iv) a respective label indicative of whether a content of the respective second training snippet factually confirms a content of the respective first training snippet; and
feeding the plurality of training objects to the add-on machine-learning model to train the add-on machine-learning model to generate (i) the respective target vector; and (ii) the respective context vector.
20. The server of claim 19 , wherein the at least one training prompt comprises a question which the respective first training snippet answers.