Scalable natural language processing for large and dynamic text environments

1. A system for retrieving a semantically matched knowledge structure, comprising:

a first device for refining at least a portion of an ontology;

a network for communicating data; and

a second device connected to the first device through the network, wherein the second device performs steps comprising:

constraint-matching a query semantic sequence received through the input component with a candidate semantic sequence stored on the memory component based on maximizing a plurality of constraints,

wherein at least one of the plurality of the constraints is a similarity score between a subsequence of the query semantic sequence and another subsequence of the candidate semantic sequence,

wherein the constraint matching generates a matched sequence and an associated matching score, and

wherein the constraint-matching preserves the ordering of the matched sequences;

finding the candidate semantic sequence for constraint-matching with the query semantic sequence, wherein the finding is based at least in part on a semantic comprised in the query semantic sequence;

determining a resulting semantic sequence chosen between the matched sequence and another matched sequence based at least in part on a comparison of the associated matching score and another associated matching score of the other matched sequence; and

refining the ontology based on the query semantic sequence and the candidate semantic sequence for use in another constraint-matching step.

2. The system of claim 1 , wherein the steps further comprise:

constraint-matching another generated query semantic sequence of a query sentence with another generated candidate semantic sequence of a candidate sentence to provide the other matched sequence and other associated matching score; and

determining the resulting semantic sequence between the matched sequence and the other matched sequence based at least in part on a comparison of the associated matching score and the other associated matching score using a cosine similarity function.

3. The system of claim 1 , wherein the steps further comprise:

generating a plurality of found sequences based on a search result list provided by a query-expansion search for, in part, a search term based on a provided semantic, wherein the query semantic sequence and the candidate semantic sequence are comprised in the plurality of found sequences;

if the matching score is above a threshold, determining a resulting semantic from the candidate sequence wherein the resulting semantic is associated with a slot-hint; and

including the resulting semantic in a semantically matched knowledge structure within the ontology.

4. The system of claim 1 , wherein the query semantic sequence and the candidate semantic sequence comprises a plurality of semantics, wherein at least one of the plurality of semantics comprises at least one of a grammatical type of a word, at least one ontological class of the word, a morphological stem of the word, or a slot-hint of the word, wherein the ontological class of the words matched is based on a least common subsumer of the words matched.

5. The system of claim 1 , wherein the similarity score is based on a function of at least a distance in ontological space between a semantic comprised in the subsequence and another semantic comprised in the other subsequence, a size of an intersection between attributes of the semantic and other attributes of the other semantic, or a morphological similarity between a word associated with the semantic and another word associated with the other semantic.

6. The system of claim 1 , wherein the refining comprises a neural network training of words or items in the ontology to create affinities between the words or items for measuring similarities in a vector space.

7. The system of claim 1 , wherein the matching score is based on a function of a plurality of similarity scores for matched subsequences.

8. The system of claim 1 , wherein the matched sequence comprises a matched subsequence that is matched against a query subsequence comprised in the query semantic, and wherein the matched subsequence is matched against a candidate subsequence comprised in the candidate semantic sequence.

9. The system of claim 1 , wherein the steps further comprise:

generating the query semantic sequence based at least in part on a sequence of semantic heads of sub-trees from a level of a natural language parse tree of a query sentence.

10. The system of claim 1 , wherein the steps further comprise:

generating the query semantic sequence based at least in part a portion of leaves of a natural language parse tree of a sentence.

11. A device for retrieving a semantically matched knowledge structure, comprising:

an input component for receiving an input;

a memory component for storing an ontology that includes a dense semantic vector indication of concepts;

a processor connected to the input component and the memory component, wherein the processor performs steps comprising:

constraint-matching a query semantic sequence received through the input component with a candidate semantic sequence stored on the memory component based on maximizing a plurality of constraints, wherein at least one of the plurality of the constraints is a similarity score between a subsequence of the query semantic sequence and another subsequence of the candidate semantic sequence based on a cosine similarity function applied to one or more dense semantic vectors from the ontology;

finding the candidate semantic sequence for constraint-matching with the query semantic sequence, wherein the finding is based at least in part on a semantic comprised in the query semantic sequence;

determining a resulting semantic sequence chosen between the matched sequence and another matched sequence based at least in part on a comparison of the associated matching score and another associated matching score of the other matched sequence; and

refining the ontology of semantic indications based on neural network training based on the query semantic sequence and the candidate semantic sequence for use in another constraint-matching step.

12. The device of claim 11 , wherein the query semantic sequence and the candidate semantic sequence comprises a plurality of semantics, wherein at least one of the plurality of semantics comprises at least one of a grammatical type of a word, at least one ontological class of the word, a morphological stem of the word, or a slot-hint of the word, wherein the ontological class of the words matched is the least common subsumer of the words matched.

13. The device of claim 11 , wherein the steps further comprise:

determining a slot-hint for a semantic in the query semantic sequence based at least in part on a named-entity-recognition function.

14. The device of claim 11 , wherein the steps further comprise:

providing the query semantic sequence based at least in part on a sequence of semantic nodes in a natural language parse tree of a query sentence; and

determining an answer slot-hint for the query sentence; and retrieving an answer semantic from the matched sequence based at least in part on a comparison of a slot-hint of the answer semantic and the answer slot-hint.

15. The device of claim 11 , wherein the steps further comprise:

selecting the candidate sequence from a search result list provided by a query-expansion search for, in part, a search term based on the semantic.

16. The device of claim 11 , wherein constraint-matching the query semantic sequence with the candidate semantic sequence comprises at least a common subsequence match, a Rabin-Karp pattern match, a Knuth-Morris-Pratt pattern match, Boyer-Moore pattern match, or a Diff pattern match.

17. The device of claim 11 , wherein the steps further comprise:

disambiguating an ontological class set of a semantic of the matched sequence based at least in part on a word-sense-disambiguation function.

18. The device of claim 11 , wherein the steps further comprise:

generating the candidate semantic sequence based at least in part on a sequence of nodes of a natural language parse of a candidate sentence; and

generating the query semantic sequence based at least in part on a sequence of nodes of a natural language parse of a candidate of query sentence.

19. The device of claim 11 , wherein the steps further comprise:

before constraint-matching the query semantic sequence with the candidate semantic, determining a slot-hint for a semantic in the candidate semantic sequence or the query semantic sequence based at least in part on a named-entity-recognition function.

20. A method for retrieving a semantically matched knowledge structure, comprising:

receiving a question and answer pair, wherein the answer is received from a query of a search engine;

constraint-matching a question with the answer based on maximizing a plurality of constraints, wherein at least one of the plurality of the constraints is a similarity score between question and answer, wherein the constraint matching generates a matched sequence;

finding, for one or more answer sequences, a subsequence that are not parsed as answer slots, wherein the finding includes finding a candidate semantic sequence from an ontology wherein the ontology includes a dense semantic vector indication of concepts;

obtaining query results from another search engine based on a combination of the answer or question, and the non-answer subsequence;

refining a KB based on the query results and the constraint matching and based on a neural network training, for a further subsequent semantic matching, wherein the KB includes a dense semantic vector indication of concepts, and wherein the refining further comprises refining at least one of a word2vec network, Hopfield network, or self-organizing map network.