Natural language semantic search system and method using weighted global semantic representations

• US 9,710,547 B2
• Filed: 11/21/2014
• Issued: 07/18/2017
• Est. Priority Date: 11/21/2014
• Status: Active Grant

• Current Assignee: Inbenta Holdings, Inc.
• Original Assignee: Inbenta Co.
• Inventors: Ma, Jordi Torras
• Primary Examiner(s): SYED, FARHAN M
• Application Number:
  US14/550,082
• Publication Number:
  US 20160147878A1
• Time in Patent Office: 970 Days
• Field of Search: None
• CPC Class Codes: G06F 16/3344 using natural language analysis

Abstract

Semantic Search Engine using Lexical Functions and Meaning-Text Criteria, that outputs a response (R) as the result of a semantic matching process consisting in comparing a natural language query (Q) with a plurality of contents (C), formed of phrases or expressions obtained from a contents'"'"' database (6), and selecting the response (R) as being the contents corresponding to the comparison having a best semantic matching degree. It involves the transformation of the contents (C) and the query in individual words or groups of tokenized words (W1, W2), which are transformed in its turn into semantic representations (LSC1, LSC2) thereof, by applying the rules of Meaning Text Theory and through Lexical Functions, the said semantic representations (LSC1, LSC2) consisting each of a couple formed of a lemma (L) plus a semantic category (SC).

Claims

1. A method for performing a semantic matching process, the method, with at least one computing device, comprising:

detecting one or more meanings of a query,

comparing the one or more meanings with one or more detected meanings of one or more pieces of content, and

outputting at least one response of the comparing;

wherein detecting one or more meanings of the query further comprises;

detecting and formalizing all meanings of the query into a global semantic representation, wherein the global semantic representation gives a full meaning of the query by transforming individual or groups of words of the query into semantic representations comprising pairs of lemma and a semantic category retrieved from a lexicon and lexical functions assignments and rules database, and

weighting the semantic representations in a basis of their category index and their frequency to generate a global weighted semantic representation of the query;

wherein detecting one or more meanings of the one or more pieces of content further comprises;

detecting and formalizing one or more meanings of the one or more pieces of content into a global semantic representation, wherein the global semantic representation gives a full meaning of the one or more pieces of content by transforming individual or groups of words of the one or more pieces of content into semantic representations comprising of pairs of lemma and a semantic category retrieved from the lexicon and lexical functions assignments and rules database; and

weighting the semantic representations in a basis of their category index and their frequency to generate a global weighted semantic representation of the one or more pieces of content;

wherein the comparing further comprises;

calculating a semantic matching degree and assigning a score between the global weighted semantic representation of the query and the global weighted semantic representation of the one or more pieces of content, and

retrieving at least one piece of content of the one or more pieces of content based on the at least one piece of content having the best assigned score and output the retrieved at least one piece of content as the response;

wherein the one or more pieces of content are formed of phrases or expressions obtained from a contents database.

2. The method of claim 1, wherein the detecting and formalizing all meanings of the query includes applying one or more lexical functions rules to change or contract the pairs of lemma and the semantic category retrieved from the lexicon in the lexicon and lexical functions assignments and rules database to generate one or more different versions of the global semantic representation of the query.

3. The method of claim 1, wherein the step of detecting and formalizing all the meanings of the one or more pieces of content includes applying lexical to the sequence representing the one or more pieces of content global meaning to transform, contract, or expand the pairs of lemma and the semantic category retrieved from the lexicon in the lexicon and lexical functions assignments and rules database.

4. The method of claim 1, further comprising, prior to calculating the semantic matching degree:

indexing, for each global weighted semantic representation, each single semantic representation, its frequency balanced semantic weight, its semantic approximation factor, and all its expansions allowed by the lexical functions rules, and for each expansion, its semantic approximation factor, of the one or more pieces of content in the contents database.

5. The method of claim 1, wherein the transformations of individual or groups of tokenized words of the query and the one or more pieces of content in semantic representations are performed by applying the rules of Meaning Text Theory and Lexical Functions rules and assignments, wherein the semantic representations comprise each of a couple formed of a lemma and a semantic category.

6. The method of claim 1, wherein the lexicon and lexical functions assignments and rules database comprises:

a database having multiple registers, each register comprising one or more fields, wherein each field comprises an entry word, a semantic category, and a lemma;

wherein the entries of each field are combined to represent the meaning of the entry word one or more other meanings of the entry word, contraries, superlatives, adjectives, and verbs; and

wherein the one or more other meanings are determined through lexical functions comprising one or more synonyms.

7. The method of claim 1, wherein the lexicon and lexical functions assignments and rules database is regularly updated on a time-basis or on a project-basis.

8. The method of claim 1, wherein the matching process between the query and the one or more pieces of content comprises:

for each semantic representation of the global semantic representation of the query;

assigning a category index that is proportional to importance of the semantic category;

normalizing the category index assignment to determine a semantic weight based on category wherein all category indexes for the global semantic representation of the query equal 1, and dividing the category index by the sum of category indexes of all semantic representation of the global semantic representation of the query;

assigning a frequency index to each semantic representation through a precalculated meaning-frequency table; and

calculating and normalizing a frequency balanced semantic weight by dividing the category by a log function of the meaning-frequency value for each meaning of the global semantic representation in the query and normalizing them in order that all frequencies of the global semantic representation of the query add up to 1;

for each of the semantic representations of the global semantic representation of the one or more pieces of content;

assigning a category index that is proportional to importance of the semantic category;

normalizing the category index to determine a semantic weight based on category wherein all category indexes for a global semantic representation of the one or more pieces of content add up to 1, and dividing the category index by the sum of category indexes of all semantic representations of the global semantic representation of the one or more pieces of content;

assigning a frequency index to each semantic representation through a precalculated meaning-frequency table, wherein each meaning has a computed frequency that represents the number of times it appears in different pieces of content and a semantic approximation factor that defines the quality of each appearance using a frequency index application, wherein the frequency index application is based on that as information decreases as probability of a meaning increases, based on a logarithmic proportion; and

calculating and normalizing a frequency balanced semantic weight by dividing the category by a log function of the meaning-frequency value for each meaning of the global semantic representation of the one or more pieces of content and normalizing them in order that all the frequency balanced semantic weights of the global semantic representation of the one or more pieces of content add up to 1.

9. The method of claim 1, wherein, for each semantic representation of the global semantic representation of the query:

if its lemma and semantic category combination match a semantic representation of the global weighted semantic representation of the one or more pieces of content or a semantic representation assigned to the semantic representation of the global semantic representation of the one or more pieces of content through a lexical function in a register of the lexicon and lexical functions assignments and rules, then calculate a partial positive similarity factor between 0 and 1, accounting for a semantic distance between each semantic representation and the semantic representation of the global semantic representation of the one or more pieces of content; and

if the semantic representation does not match any semantic representation of the global weighted semantic representation of the one or more pieces of content, then calculate a partial positive similarity factor as 0.

10. The method of claim 9, wherein a total positive similarity is calculated as the sum of all partial positive similarities of the global weighted semantic representation of the query.

11. The method of claim 10, wherein, for every semantic representation of the global weighted semantic representation of the one or more pieces of content not contributing to the total positive similarity, calculate a partial negative similarity as a frequency balanced semantic weight of the semantic representation of the global weighted semantic representation of the one or more pieces of content with no correspondence with any semantic representation in the global weighted semantic representations of the query.

12. The method of claim 11, wherein a total negative similarity is calculated as the sum of all partial negative similarities of the global weighted semantic representation of the one or more pieces content.

13. The method of claim 12, wherein for each of the one or more pieces of content, calculate a coincidence score as the difference between the total positive similarity and the total negative similarity when the linguistic type for the one or more pieces of content is a phrase;

wherein the total positive similarity value is the coincidence score when the linguistic type for the one or more pieces of content is free text.

14. The method of claim 13, wherein a semantic matching degree is determined based on the query and the one or more pieces of content for each coincidence between the global weighted semantic representation of the query and the global semantic representation of the one or more pieces of content as the coincidence for the reliability of matching of the one or more pieces of content.

15. The method of claim 14, wherein the response to the query is selected as the one or more pieces of content having a highest semantic matching degree.