Generation for a QAsystem CRGFUJI XEROX Tomoko Ohkuma

Issues • Typical QA system often ranks wrong answers higher than correct answers. •

Typical QAsystem Question Analysis Source corpus Document Retrieval Information Extraction Answer Selection Answer ©

Typical QAsystem Question Analysis Specify the question type Construct a query Document Retrieval Retrieve

Basic idea • Examine answer candidates on the Internet (like Google) • Making queries

Examination System QAsystem question Query(sentence) XLE(PARSER) XLE(GENERATOR) XLEto. XMLto. XLE SVM candidates Transfer Anaphora

Two strategies • Rank correct answers higher – find possible answers from the internet

Treasure Hunt – Find possible answers • Q. Who died in 1963? • A.

Garbage Sweep – Remove impossible answers • Q. Who died in 1963? • A.

Why is generation needed? • Bag of words (Boolean) queries might hit something. •

Current status of Japanese grammar for generation • Coverage: 30%-> 98% – Bug fix

Generation for making queries • Ambiguities are welcome (if they are wellformed) • Some

Transfer for making queries: condense Who played Brigit in the movie? PRED play<who, Bridget>

Transfer for making queries: Main clause -> relative clause Where may I buy i.

Transfer for making queries: relative clause -> main clause Which company is the company

Future Work • Cut down costs for searching • Transfer rule writing • Paraphrase

Slides: 16

Download presentation

Generation for a QAsystem • CRG/FUJI XEROX • Tomoko Ohkuma Hiroshi MASUICHI

Issues • Typical QA system often ranks wrong answers higher than correct answers. • High ranked wrong answers convey the impression that the system is ‘useless’ than its actual accuracy. © 2003 -4 Fuji Xerox Co. , Ltd. All Rights Reserved.

Typical QAsystem Question Analysis Specify the question type Construct a query Document Retrieval Retrieve passages and rank them based on the query Information Extract named entities according to the question type Answer Selection Rank the answers based on the density of query words © 2003 -4 Fuji Xerox Co. , Ltd. All Rights Reserved.

Basic idea • Examine answer candidates on the Internet (like Google) • Making queries for the examination from the candidates and questions by using generation grammar on XLE © 2003 -4 Fuji Xerox Co. , Ltd. All Rights Reserved.

Treasure Hunt – Find possible answers • Q. Who died in 1963? • A. John Lennon, Robert Kennedy , John. F. Kennedy – The query should be strict in order to avoid hunting wrong answers. © 2003 -4 Fuji Xerox Co. , Ltd. All Rights Reserved.

Garbage Sweep – Remove impossible answers • Q. Who died in 1963? • A. Okra, Beatles, broccoli – The query should be simple in order to avoid sweeping correct answers. © 2003 -4 Fuji Xerox Co. , Ltd. All Rights Reserved.

Current status of Japanese grammar for generation • Coverage: 30%-> 98% – Bug fix – Add Gen. OT to lexical entries – Add negation rules to grammar rules © 2003 -4 Fuji Xerox Co. , Ltd. All Rights Reserved.

Transfer for making queries: condense Who played Brigit in the movie? PRED play<who, Bridget> Zellweger SUBJ PRED who int=+ OBJ PRED Bridget ADJUNCT PRED in<OBJ> OBJ PRED movie • Remove Adjunct(or MOD) • Change ‘int’ to decl • Replace interrogative pronoun with Answers Clause-type Interrogative declerative Renee Zellweger played Bridget. ゼルウィガーがブリジットを演じる。ブリジットをゼルウィガーが演じる。 © 2003 -4 Fuji Xerox Co. , Ltd. All Rights Reserved.

Transfer for making queries: Main clause -> relative clause Where may I buy i. Pod? PRED buy<null, ipod> where PRED Apple store SUBJ PRED null PRED buy<null, i. POD> OBJ PRED ipod. PRED null SUBJ ADJUNCT int=+ PRED where ADJUNCT OBJ int PRED + i. Pod ADJUNCT pron-type rel Apple store I can buy i. Pod{が|を}買えるアップルストア © 2003 -4 Fuji Xerox Co. , Ltd. All Rights Reserved. • Move PRED to ADJUNCT • Move interrogative pronoun to PRED • Replace interrogative pronoun with Answers

Transfer for making queries: relative clause -> main clause Which company is the company that sells the hybrid car? PRED Be< which-company, company> PREDwhich-company<pro> sell<Toyota , CAR> PRED SUBJ Toyota what SUBJPRED company int=+ XCOMP OBJ ADJUNCT PRED CAR sell<pro, OBJ> PRED pro MODSUBJ PRED hyprid Pron-type rel OBJ PRED car MOD PRED Hyprid Toyota sells the hybrid car トヨタはハイブリッドカーを発売するハイブリッドカーをトヨタは発売する © 2003 -4 Fuji Xerox Co. , Ltd. All Rights Reserved. • Extract ADJUNCT in XCOMP • Replace interrogative pronoun with Answers