System and method for estimation of interlocutor intents and goals in turn-based electronic conversational flow
A system and method implemented on a computing device for analyzing a digital corpus of unstructured interlocutor conversations to discover intents, goals, or both intents and goals of one or more parties to the conversations, by grouping the conversation utterances according to semantic similarity clusters; selecting the best utterance(s) that mostly likely embody a party's stated goal or intent; creates a set of candidate intent names for each cluster based upon each intent utterance in each conversation in each cluster; rates each candidate intent (or goal) for each intent name; and selects the most likely candidate intent (or goal) name for the purposes of subsequent automation of future conversations such as, but not limited to, automated electronic responses using Artificial Intelligence and machine learning.
1. A method implemented by a computing device for discovering an intent or a goal or both an intent and a goal of a party in an interlocutor digital conversation, the method comprising:
grouping, by a computer system, a plurality of digitally-recorded conversations into clusters according to similarity;
selecting, by the computer system, one or more candidate utterances within the clusters which mostly likely embody a specific conversation party's stated goal or intent;
rating, by the computer system, each selected candidate utterance for each intent, for each goal, or for both intent and goal;
selecting, by the computer system, a most likely or highest rated candidate utterance; and
outputting, by the computer system, the selected candidate utterance into a digital model for use in artificial intelligence (AI) training data.
2. The method of claim 1 wherein the grouping is preceded by encoding, by the computer system, sentence embeddings contained in a corpus.
3. The method of claim 2 wherein the encoding sentence embeddings comprises performing Language-Agnostic Bidirectional Encoder Representations from Transformers Sentence Encoding (LABSE).
4. The method of claim 2 wherein the encoding sentence embeddings comprises performing Robustly Optimized Bidirectional Encoder Representations from Transformers Pretraining Approach (RoBERTa).
5. The method of claim 2 wherein the encoding sentence embeddings is followed by, prior to the grouping, performing, by the computer system, dimensionality reduction on the encoded sentence embeddings.
6. The method of claim 5 wherein the dimensionality reduction comprises performing Uniform Manifold Approximation and Projection (UMAP).
7. The method of claim 5 wherein the dimensionality reduction comprises performing t-Distributed Stochastic Neighbor Embedding (t-SNE).
8. The method of claim 1 wherein the grouping comprises performing clustering.
9. The method of claim 8 wherein the clustering comprises performing Kmeans clustering.
10. The method of claim 8 wherein the clustering comprises performing Consensus clustering.
11. The method of claim 1 wherein the selecting is preceded by performing cluster splitting.
12. The method of claim 11 wherein the cluster splitting comprises performing splitting clusters into clusters for label generation and clusters for label ranking.
13. The method of claim 1 further comprising creating, by the computer system, of a set of candidate intent names by performing label generation.
14. The method of claim 13 wherein the label generation comprises performing Generative Pre-trained Transformer 2 (GPT-2).
15. The method of claim 13 wherein the label generation comprises performing Bidirectional Encoder Representations from Transformers (BERT).
16. The method of claim 13 wherein the creating of the set of candidate intent names comprises performing simplification on the set of candidate intent names.
17. The method of claim 13 further comprising selecting a candidate intent name by ranking the set of candidate intent names according to a clusters split, and wherein the outputting further comprises outputting the selected candidate intent name.
18. A non-transitory computer program product for discovering an intent or a goal or both and intent and a goal of a party in an interlocutor digital conversation, comprising:
one or more tangible, non-transitory computer-readable memories which are not propagating signals per se; and
program instructions embodied by the one or more tangible, non-transitory computer-readable memories for causing one or more computer processors to, when executed by a processor:
group a plurality of digitally-recorded conversations into clusters according to similarity;
select one or more candidate utterances which mostly likely embody a specific conversation party's stated goal or intent;
rate each selected candidate utterance for each intent, for each goal, or for both intent and goal;
select a most likely or highest rated candidate utterance; and
output the selected most likely candidate intent utterance into a digital model for use in artificial intelligence (AI) training data.
19. A system for discovering an intent or a goal or both and intent and a goal of a party in an interlocutor digital conversation, comprising:
one or more computer processors for executing program instructions;
one or more tangible, non-transitory computer-readable memories which are not propagating signals per se, in communication with the one or more computer processors; and
program instructions embodied by the one or more tangible, non-transitory computer-readable memories for causing the one or more computer processors to, when executed by the one or more computer processors:
group a plurality of digitally-recorded conversations into clusters according to similarity;
select one or more candidate utterances which mostly likely embody a specific conversation party's stated goal or intent;
rate each selected candidate utterance for each intent, for each goal, or for both intent and goal;
select a most likely or highest rated candidate utterance; and
output the selected most likely candidate intent utterance into a digital model for use in artificial intelligence (AI) training data.