Input and Data agnostic topic-evolution system for R&D insights

We are inventing a data-agnostic and input-agnostic insights algorithm, adoptable to differing text sparsities and volumes.áThe core merit is a general purpose Markov Chain Monte Carlo algorithm(MCMC) for business insights.áOur algorithm is trained on unstructured text corpora for Patent Datasets (900 Intel Patents), and our goal would be to manually validate the results for the DoD on the datasets and insights needed to achieve business outcomes.The detailing of the innovation and significance of our technology lies as following:We analyze the ground-truth topic creations to develop the bestáBayesian topic generation algorithm, both qualitatively analyzing and quantitatively measuring topic coherence, diversity, and number of topics of which statistical topic modeling algorithms perform best according to the text sparsities and volumes of the data to inference. Our system runs multiple statistical topic modeling algorithms, including Mallet LDA, Tomotopy LDA, Gibbs Sampling LDA, NMF, and SVD algorithms in parallel, and routes the data to the model outputs that achieves the best metrics we optimize for determining the most robust topic creation performances.ááWe develop techniques to converge on the optimal number of topics and Bayesian hyperparameters, by approximating Bayesian hyperparameter values that optimize for a weighted multi-objective function to maximize coherence and diversity. We run multiple-fold validation procedures in order to converge the approximate number of topics per the dataset, and approximate to the upper bound number of topics. The purpose of these procedures is to make the develop the system to inference data-agnostically with robust results. Steps 1 and 2 in parallel allow our system to be ôdata-agnosticö.áWe run a Singular Value Decomposition algorithm to extract key-terms of the corpus, for the user to select key-terms to guide the posterior-distribution sampling topic modeling inference. This enables our system to be ôinput-agnosticö.áOnce we have this grounded, comprehensive, and granular and topic generation, we connect them temporally through the corpus ľ in which the user can interval their data into weeks, months, or quarters to get varying shades of trend emergences on varying intervals. We use varying heuristics to build topic convergences and divergences, and mine for deep insights by connecting temporal threads of information with n-gramming techniques to find connective affects of information for insights.áWe rank topic strands that have emerged through the corpus, and build a ranking algorithm factoring in NLP, coherence, similarity, and growth metrics to identify best topics to recommend to the user.áOnce we obtain the topics, we inference other companiesĺ patents, and vectorize many companiesĺ topics to identify which companies are most closely aligned, to the specific topic of analysis that the user has identified for the second-set of re-contextualizing topics across differing datasets.á