# US 8,543,565 B2 · System and method using a discriminative learning approach for question answering > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 8,543,565 B2 - Document ID: US8543565B2 - Status: Granted - Kind: B2 - Application number: 11/851,817 - Filed: 2007-09-07 - Published: 2009-03-12 - Granted: 2013-09-24 - Earliest claimed: 2007-09-07 - Assignees: AT&T Intellectual Property II, L.P. - Applicants: Junlan Feng - Inventors: Junlan Feng - Web page: https://patentgrep.com/patent/US8543565B2 ## CPC - G06F 16/3329: Natural language query formulation - G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N) ## Abstract Disclosed are systems, methods, and computer readable media for answers to natural language questions. The method embodiment comprises training a lexical association model between a question and a first set of one or more possible answers, training a semantic association model between a question and a second set of one or more possible answers, receiving a user question containing at least one query word, parsing the user question syntactically and semantically, formulating a query from the parsed user question containing at least one query word, expanding the query based on the lexical association model and the semantic association model, weighting the at least one query word according to its importance when answering the user question, and returning an answer based on the weighted at least one query word, the lexical association model, and the semantic association model. Other features include using question-answer pairs mined to train the models and returning a plurality of answers in an order based on the lexical association model and the semantic association model. ## Claims ### Claim 1 A method comprising: training a lexical association model between a question and a first set of possible answers; training a semantic association model between the question and a second set of possible answers; receiving a user question containing query words; parsing the user question syntactically and semantically, to yield a parsed user question; expanding the query words based on the lexical association model and the semantic association model, to yield expanded query words; weighting the expanded query words according to importance when answering the user question, to yield a weighted expanded query words; and returning an answer based on a score calculated based on the following equation: score ⁡ ( q , a ) = ∑ w ∈ q , a ⁢ docf ⁡ ( w , q , a ) * qf ⁡ ( w ) + ∑ ne ∈ NE ⁡ ( qp ) , qp ∈ q ⁢ docf ⁡ ( ne , q , a ) * λ sem ⁡ ( qp , ne ) + ∑ v ∈ Exp ⁡ ( w ) , w ∈ q , v ∈ a ⁢ docf ⁡ ( v , q , a ) * λ lex ⁡ ( w , v ) wherein docf represents a normalized frequency distribution function, w represents individual query words, q represents individual questions, a represents individual answers, NE represents a set of named entity tags, ne represents individual named entity tags, qp represents a question phrase, v represents an answer word, w represents a query word, λsem represents a semantic association parameter, λlex represents a lexical association parameter, and qf represents the weighted expanded query words. ### Claim 2 (depends on 1) The method of claim 1, wherein question-answer pairs are used to train the lexical and semantic association models. ### Claim 3 (depends on 2) The method of claim 2, wherein the question-answer pairs are mined from the Internet. ### Claim 4 (depends on 1) The method of claim 1, wherein a plurality of questions are used to train the semantic and lexical association models. ### Claim 5 (depends on 1) The method of claim 1, wherein a plurality of answers are returned. ### Claim 6 (depends on 5) The method of claim 5, wherein the plurality of answers is returned in an order based on the lexical association model and the semantic association model. ### Claim 7 (depends on 1) The method of claim 1, wherein training the lexical association model further comprises finding associated answer words for the query word and building an expansion answer set for the query word, wherein the expansion answer set comprises an answer word. ### Claim 8 (depends on 7) The method of claim 7, wherein expanding the query words further comprises finding a question phrase that the question begins with, locating nouns in the question phrase, and replacing the question phrase with associated named entity tags and the nouns appearing in the question phrase. ### Claim 9 (depends on 1) The method of claim 1, wherein the first set and the second set of possible answers comprise different sets. ### Claim 10 (depends on 1) The method of claim 1, wherein the first set and the second set of possible answers comprise the same set. ### Claim 11 A system comprising: a processor; and a computer-readable storage medium having instructions stored which, when executed by the processor, cause the processor to perform operations comprising: training a lexical association model between a question and a first set of possible answers; training a semantic association model between the question and a second set of possible answers; receiving a user question containing query words; parsing the user question syntactically and semantically, to yield a parsed user question; expanding the query words based on the lexical association model and the semantic association model, to yield expanded query words; weighting the expanded query words according to importance when answering the user question, to yield a weighted expanded query words; and returning an answer based on a score calculated based on the following equation: score ⁡ ( q , a ) = ∑ w ∈ q , a ⁢ docf ⁡ ( w , q , a ) * qf ⁡ ( w ) + ∑ ne ∈ NE ⁡ ( qp ) , qp ∈ q ⁢ docf ⁡ ( ne , q , a ) * λ sem ⁡ ( qp , ne ) + ∑ v ∈ Exp ⁡ ( w ) , w ∈ q , v ∈ a ⁢ docf ⁡ ( v , q , a ) * λ lex ⁡ ( w , v ) wherein docf represents a normalized frequency distribution function, w represents individual query words, q represents individual questions, a represents individual answers, NE represents a set of named entity tags, ne represents individual named entity tags, qp represents a question phrase, v represents an answer word, w represents a query word, λsem represents a semantic association parameter, λlex represents a lexical association parameter, and qf represents the weighted expanded query words. ### Claim 12 (depends on 11) The system of claim 11, wherein question-answer pairs are used to train the lexical and semantic association models. ### Claim 13 (depends on 11) The system of claim 11, wherein a plurality of questions are used to train the semantic and lexical association models. ### Claim 14 (depends on 11) The system of claim 11, wherein a plurality of answers are returned. ### Claim 15 (depends on 14) The system of claim 14, wherein a plurality of answers is returned in an order based on the lexical association model and the semantic association model. ### Claim 16 (depends on 11) The system of claim 11, wherein training the lexical association model further comprises finding associated answer words for the query word and building an expansion set for the query word, wherein the expansion set comprises an answer word. ### Claim 17 (depends on 16) The system of claim 16, wherein expanding the query words further comprises finding a question phrase that the question begins with, locating nouns in the question phrase, and replacing the question phrase with associated named entity tags and the nouns appearing in the question phrase. ### Claim 18 (depends on 11) The system of claim 11, wherein the first set and the second set of possible answers are different sets. ### Claim 19 (depends on 11) The system of claim 11, wherein the first set and the second set of possible answers are the same set. ### Claim 20 A computer-readable storage memory having instructions stored which, when executed by a computing device, cause the computing device to perform operations comprising: training a lexical association model between a question and a first set of possible answers; training a semantic association model between the question and a second set of possible answers; receiving a user question containing query words; parsing the user question syntactically and semantically, to yield a parsed user question; expanding the query words based on the lexical association model and the semantic association model, to yield expanded query words; weighting the expanded query words according to importance when answering the user question, to yield a weighted expanded query words; and returning an answer based on a score calculated based on the following equation: score ⁡ ( q , a ) = ∑ w ∈ q , a ⁢ docf ⁡ ( w , q , a ) * qf ⁡ ( w ) + ∑ ne ∈ NE ⁡ ( qp ) , qp ∈ q ⁢ docf ⁡ ( ne , q , a ) * λ sem ⁡ ( qp , ne ) + ∑ v ∈ Exp ⁡ ( w ) , w ∈ q , v ∈ a ⁢ docf ⁡ ( v , q , a ) * λ lex ⁡ ( w , v ) wherein docf represents a normalized frequency distribution function, w represents individual query words, q represents individual questions, a represents individual answers, NE represents a set of named entity tags, ne represents individual named entity tags, qp represents a question phrase, v represents an answer word, w represents a query word, λsem represents a semantic association parameter, λlex represents a lexical association parameter, and qf represents the weighted expanded query words.