Semantic cluster formation in deep learning intelligent assistants

1. A method comprising:

receiving a user question directed to a subject having a documentation corpus;

determining a set of passages of the documentation corpus related to the user question;

ranking the set of passages according to relevance to the user question;

forming semantic clusters of sentences extracted from one or more top ranked passages of the set of passages according to sentence similarity;

determining a classification of whether the user question corresponds to a factoid response or a non-factoid response; and

providing a response to the user question based at least on a selected semantic cluster and the classification.

2. The method of claim 1 , further comprising:

contextualizing the user question against the documentation corpus based on at least one previous user question directed to the documentation corpus.

3. The method of claim 1 , further comprising:

processing similarity across consecutive sentences to determine an inter-sentence similarity metric;

determining the passages in the documentation corpus as groupings among the consecutive sentences based on the inter-sentence similarity metric exceeding a similarity threshold; and

determining break points between the passages based on the inter-sentence similarity metric falling below the similarity threshold among the consecutive sentences.

4. The method of claim 1 , further comprising:

forming a query comprising keywords based on the user question;

based on the query, processing the documentation corpus to determine local coherence to indicate at least the set of passages as relevant to the keywords along with relevance probabilities of the set of passages.

5. The method of claim 1 , further comprising, based on determining that the user question corresponds to the non-factoid response:

tokenizing the set of passages into the sentences; computing semantic similarity scores among the sentences of the set of passages; grouping the sentences into associated ones of the semantic clusters based on the semantic similarity scores; ranking the semantic clusters according to relevance to the user question; and selecting a top ranked semantic cluster as the selected semantic cluster to provide in the response.

6. The method of claim 1 , further comprising, based on determining that the user question corresponds to the factoid response: processing the selected semantic cluster with a neural network having inputs representing the selected semantic cluster and keywords determined from the user question to generate probabilities that sentences of the selected semantic cluster contain the factoid response; and based on the probabilities, selecting at least one among the sentences of the selected semantic cluster as the response.

7. The method of claim 1 , further comprising:

discovering images within the documentation corpus;

processing at least one among headings, captions, titles, and optical character recognition (OCR) results associated with the images to determine similarity scores to proximate ones of the passages;

selecting at least one of the images each having a corresponding one of the similarity scores within a threshold similarity to the selected semantic cluster; and

providing the at least one of the images with the response to the user question.

8. One or more computer storage media storing computer-useable instructions that, when used by one or more computing devices, cause the one or more computing devices to perform operations comprising:

receiving a user question associated with a documentation corpus;

determining a set of passages from the documentation corpus related to the user question;

employing a neural network to rank the set of passages according to relevance to the user question;

forming semantic clusters of sentences extracted from one or more top ranked passages of the set of passages according to sentence similarity;

determining a classification of whether the user question corresponds to a factoid response or a non-factoid response; and

producing a response to the user question based at least on a selected semantic cluster and the classification.

9. The one or more computer storage media of claim 8 , the operations further comprising contextualizing the user question against the documentation corpus based on at least one previous user question directed to the documentation corpus.

10. The one or more computer storage media of claim 8 , the operations further comprising:

processing similarity across consecutive sentences to determine an inter-sentence similarity metric;

determining the passages from the documentation corpus as groupings among the consecutive sentences based on the inter-sentence similarity metric exceeding a similarity threshold; and

determining break points between the passages based on the inter-sentence similarity metric falling below the similarity threshold among the consecutive sentences.

11. The one or more computer storage media of claim 8 , the operations further comprising processing the documentation corpus to determine local coherence to indicate at least the set of passages as relevant to keywords along with relevance probabilities of the set of passages.

12. The one or more computer storage media of claim 8 , the operations further comprising, based on the classification being that the user question correspond to the non-factoid response: tokenizing the set of passages into the sentences; computing semantic similarity scores among the sentences of the set of passages; grouping the sentences into associated ones of the semantic clusters based on the semantic similarity scores; ranking the semantic clusters according to relevance to the user question; and selecting a top ranked semantic cluster as the selected semantic cluster to provide in the response.

13. The one or more computer storage media apparatus of claim 8 , the operations further comprising, based on the classification being that the user question corresponds to the factoid response: processing the selected semantic cluster with the neural network having inputs representing the selected semantic cluster and keywords determined from the user question to generate probabilities that sentences of the selected semantic cluster contain the factoid response; and based on the probabilities, selecting at least one among the sentences of the selected semantic cluster as the response.

14. The one or more computer storage media of claim 8 , the operations further comprising:

discovering images within the documentation corpus;

processing at least one among headings, captions, titles, and optical character recognition (OCR) results associated with the images to determine similarity scores to proximate ones of the passages;

selecting at least one of the images each having a corresponding one of the similarity scores within a threshold similarity to the selected semantic cluster; and

providing the at least one of the images with the response to the user question.

15. A system comprising one or more processors and memory configured to provide computer program instructions to the one or more processors, the computer program instructions comprising:

an assistance interface configured to direct the one or more processors to receive a user question related to a subject and provide a response to the user question from among a documentation corpus describing the subject;

an assistance service configured to direct the one or more processors to:

determine a set of passages from the documentation corpus related to the user question;

employ a neural network to rank the set of passages according to probability of relevance to the user question;

produce semantic clusters of sentences extracted from one or more top ranked passages of the set of passages according to sentence similarity;

determine a classification of whether the user question corresponds to a factoid response or a non-factoid response; and

produce the response to the user question based at least on a selected semantic cluster and the classification.

16. The system of claim 15 , the assistance service further configured to direct the one or more processors to contextualize the user question against the documentation corpus based on at least one previous user question directed to the documentation corpus.

17. The system of claim 15 , the assistance service further configured to direct the one or more processors, based on the classification being that the user question corresponds to the non-factoid response, to:

compute semantic similarity scores among sentences of the set of passages;

group the sentences of the set of passages into associated ones of the semantic clusters based on the semantic similarity scores;

rank the semantic clusters according to relevance to the user question; and

select a top ranked semantic cluster to provide in the response.

18. The system of claim 15 , the assistance service further configured to direct the one or more processors to:

catalog images within the documentation corpus;

process at least one among headings, captions, titles, and optical character recognition (OCR) results associated with the images to determine similarity scores to proximate ones of the passages;

select at least one of the images each having a corresponding one of the similarity scores within a threshold similarity to the selected semantic cluster; and

produce the response to include the at least one of the images.