Performing semantic analyses of user-generated textual and voice content

1. A method for performing semantic analysis on a user-generated text string, the method comprising:

obtaining a plurality of known text strings from one or more web portals, wherein the plurality of known text strings are obtained by crawling the one or more web portals using one or more application program interfaces (APIs);

cleaning the plurality of known text strings, wherein the cleaning comprises removing generic and domain-specific stop words and uniform resource locator (URL) symbols and text from the plurality of known text strings;

assembling the cleaned plurality of known text strings into an input document;

training a neural network model with the input document to obtain a first distributed vector representation of the cleaned plurality of known text strings and a second distributed vector representation of a plurality of words in the cleaned plurality of known text strings;

computing a relevance matrix of the first and second distributed representations based on a cosine distance between each of the plurality of words and the plurality of known text strings;

performing a latent dirichlet allocation (LDA) operation using the relevance matrix as an input to obtain a distribution of topics associated with the plurality of known text strings;

monitoring the one or more web portals for new text strings using the one or more APIs;

determining that a quantity of the new text strings meets a threshold sufficient to train the neural network model; and

refining the distribution of topics based on adding the threshold quantity of new text strings to the plurality of known text strings to obtain a refined distribution of topics including the new text strings.

2. The method of claim 1 , wherein the cleaning further comprises one or more of: identifying and removing spurious textual patterns, correcting spelling, and filtering based on length.

3. The method of claim 1 , further comprising narrowing the distribution of topics by:

determining seed topics from the distribution of topics;

obtaining contextually similar words associated with the seed topics;

computing a second relevance matrix based on a cosine distance between each of the contextually similar words and the plurality of known text strings; and

performing the LDA operation using the second relevance matrix as an input to obtain a narrowed distribution of topics.

4. The method of claim 3 , wherein the seed topics are determined based on most-recurring topics from the distribution of topics.

5. The method of claim 3 , wherein the seed topics are determined based on user input.

6. The method of claim 3 , further comprising assigning a high non-zero value to each of the contextually similar words.

7. The method of claim 1 , wherein computing the relevance matrix comprises:

projecting the plurality of text strings and the words in the same dimension; and

computing the cosine distance between the plurality of text strings and the words.

8. The method of claim 7 , wherein each row of the relevance matrix indicates the cosine distance.

9. The method of claim 7 , further comprising filtering the plurality of words prior to generating the relevance matrix.

10. The method of claim 9 , wherein the filtering comprises identifying words associated with one or both of a very high or a very low cosine distance.

11. The method of claim 10 , further comprising performing the LDA operation using the relevance matrix with the filtered words.

12. The method of claim 1 , further comprising determining an actionable issue associated with the topic, identifying a remedial action, and notifying a service provider.

13. A method for performing semantic analysis on a user-generated text string, the method comprising:

obtaining a plurality of known text strings from one or more web portals, wherein the plurality of known text strings are obtained by crawling the one or more web portals using one or more application program interfaces (APIs);

cleaning the plurality of known text strings, wherein the cleaning comprises removing generic and domain-specific stop words and uniform resource locator (URL) symbols and text from the plurality of known text strings;

assembling the cleaned plurality of known text strings into an input document;

training a neural network model with the input document to obtain a first distributed vector representation of the cleaned plurality of known text strings and a second distributed vector representation of a plurality of words in the cleaned plurality of known text strings;

computing a first relevance matrix of the first and second distributed representations based on a cosine distance between each of the plurality of words and the plurality of known text strings;

performing a first latent dirichlet allocation (LDA) operation using the first relevance matrix as an input to obtain a plurality of most-recurring topics from the plurality of text strings;

labeling the plurality of most-recurring topics as seed topics;

extracting a plurality of contextually similar words associated with each of the seed topics, wherein the extracting comprises querying the neural network model;

computing a second relevance matrix based on a cosine distance between each of the plurality of contextually similar words and the plurality of text strings from the user-generated content; and

performing a second latent dirichlet allocation (LDA) operation using the second relevance matrix as an input to obtain a distribution of issues related to the seed topics.

14. The method of claim 13 , wherein querying the neural network model comprises querying a global neural network model and a specific neural network model with each of the one or more seed topics, wherein the specific neural network model is trained using known text strings corresponding to at least one of the one or more seed topics.

15. The method of claim 13 , further comprising clustering the distribution of issues.

16. The method of claim 13 , further comprising:

repeating the operations using new unseen user-generated content from the one or more data sources; and

computing an intra-cluster consistency at an interval.

17. A system for extracting actionable topics from user-generated content, the system comprising:

a first processing module communicably coupled to one or more web portals, the first processing module being configured to obtain a plurality of known text strings from one or more web portals by crawling the one or more web portals using one or more application program interfaces (APIs);

a second processing module communicably coupled to the first processing module, the second processing module configured to:

clean the plurality of known text strings, wherein the cleaning comprises removing generic and domain-specific stop words and uniform resource locator (URL) symbols and text from the plurality of known text strings;

assemble the cleaned plurality of known text strings into an input document;

train a neural network model with the input document to obtain a first distributed vector representation of the cleaned plurality of known text strings and a second distributed vector representation of a plurality of words in the cleaned plurality of known text strings;

compute a relevance matrix of the first and second distributed representations based on a cosine distance between each of the plurality of words and the plurality of known text strings;

perform a latent dirichlet allocation (LDA) operation using the relevance matrix as an input to obtain a distribution of topics associated with the plurality of known text strings;

monitor the one or more web portals for new text strings using the one or more APIs;

determine that a quantity of the new text strings meets a threshold sufficient to train the neural network model;

refine the distribution of topics based on adding the threshold quantity of new text strings to the plurality of known text strings to obtain a refined distribution of topics including the new text strings;

extract a location associated with the refined distribution of topics, a time associated with the refined distribution of topics, and an actionable topic associated with the refined distribution of topics;

identify a remedial action based on the actionable topic; and

transmit a recommendation to a service provider network associated with the actionable topic, wherein the recommendation comprises the identified remedial action, the location, and the time.

18. The system of claim 17 , wherein the one or more web portals comprises an online social network, and wherein the known or new text string comprises one or more of a post on the online social network or a reply on the online social network.

19. The system of claim 17 , wherein extracting the time or location further comprises:

extracting a time or location metadata from the input document, and

correlating the time or location metadata with a corresponding time or location extracted from the refined distribution of topics.