The Latest Progress of KB-QA. Ming Zhou Microsoft Research Asia NLPCC 2018 Workshop Hohhot, 2018

Transcription

1 The Latest Progress of KB-QA Ming Zhou Microsoft Research Asia NLPCC 2018 Workshop Hohhot, 2018

2 Query CommunityQA KBQA TableQA PassageQA VQA Edited Response Question Generation Knowledge Table Doc Image Automatic Response Automatic Response Automatic Response Automatic Response Human-in-the-Loop (HI)

3 Who is Michelle Obama married to in 1992? λx. Marriage_Spouse(x, Michelle Obama) Marriage_StartDate(x, 1992) Barack Obama

4 KBQA in Search Engine

5 Semantic Parsing-based Method KBQA Information Answer Ranking-based Retrieval-based KBQA Method CCG SCFG DCS NN x Place_of_Birth Barack Obama Feature NLG Sub-graph Embedding Barack Obama Place_of_Birth Honolulu Barack Obama Place_of_Birth Honolulu

6

7 expression expression constant variable function application

8 China, Bill Gates, Mount Everest, 2017,

9 x, y, z,

10 λx. Place_Of_Birth(Barack Obama, x) the argument of the function the definition of the function

11 Deduction of two expressions λxλy. Place_Of_Birth x, y λx. (x = Barack Obama) λy. Place_Of_Birth Barack Obama, y

12 Natural Language Semantic Parsing Grammar-based Semantic Parsing Neural Network-based Semantic Parsing Logical Form

13 Logical Form Ranking Semantic Rule-Driven CYK Parsing Natural Language Question

14 syntactic semantic A CCG Rule Example border (S\NP)/NP Match natural language input natural language syntax semantics Syntactic symbols: S, N, NP, ADJ and PP Syntactic combinator: / and \ Slashes specify combination orders and directions λ-calculus expression Sematic types are the logical forms of the natural language parts

15 State borders New Mexico NP (S\NP)/NP NP S\NP > S <

16 Texas borders New Mexico borders(texas, new_mexico) use rules to extract all possible <Q, LF> pairs Texas := NP : texas borders := (S \ NP) / NP : λx.λy.borders(y, x) New Mexico := NP : new_mexico 1. maximize the likelihood: 2. keep CCG rules that occur in the highest scoring derivations of training data

17 lexical semantic Category X α β a sequence of terminal and non-terminal symbols the logical form of α Non-terminal symbols in α and β should have 1-to-1 correspondence. [Person] TomHanks Tom Hanks [Film] the moive starred by Person 1 λx. Film_Actor_Film( Person 1, x) [Film] the moive starred by Tom Hanks λx. Film_Actor_Film(Tom Hanks, x)

18 λxλy. Film_Film_Director(y, x) Film_Actor_Film(Tom Hanks, y) director of [Film] SCFG rule matching λy. Film_Actor_Film(Tom Hanks, y) the movie starred by [Person] SCFG rule matching Lots of semantic derivations will be generated during this procedure. Tom Hanks entity linking director of the movie starred by Tom Hanks

19 Paired Entities of a given KB Predicate Film.Film.Director <Forrest Gump, Robert Zemeckis> <Titanic, James Cameron> <Rain Man, Barry Levinson> Passage Retrieval from Raw Text Robert Zemeckis is director of Forrest Gump Titanic was a movie directed by James Cameron Barry Levinson was famous as the director of Rain Man Relation Patterns Film.Film.Director [Director] is director of [Film] 0.84 [Film] was a movie directed by [Director] 0.81 [Director] was famous as the director of [Film] 0.77 Director is director of Film 1 λx. Film_Director_Film(x, Film 1 )

20 Encoder Decoder

21

22 Decoder column value SQL SELECT WHERE COUNT MIN MAX AND > < = column value SQL column value t = 0 t = 2 t = 6 SQL Wilfrid Laurier York York <S> SELECT COUNT CFL Team WHERE College = "York" SQL SQL column SQL column SQL value Output Encoder Attention Pick # CFL Team Player Position College 27 Hamilton Tiger-Cats Connor Healy DB Wilfrid Laurier 28 Calgary Stampeders Anthony Forgone OL York 29 Toronto Argonauts Frank Hoffman DL York SELECT, WHERE, COUNT, MIN, MAX, AND, >, <, =. Question Table SQL

23

24 Learning-to-Rank Model Question Features Answer Features Question Feature Extraction Answer Feature Extraction Natural Language Question KB Answer Entity Candidate

25 what is the name of Justin Bieber brother attr what is det nsubj name prep_of London place_of_birth Justin Bieber person.sibling_s sibling CVT type gender sibling Person Male the nn Justin brother nn Bieber gender Jazmyn Bieber type gender Jaxon Bieber type Male Person Male Person h i (QG, TG; A) Question Graph (QG) Each feature is a pairwise concatenation of a question graph feature and a topic graph feature of a specific answer node candidate. Topic Graph (TG)

26 String Similarity Model Original Question Generated Question Answer-aware Question Generation Natural Language Question KB Answer Entity Candidate

27 which city was Obama born? Ranking Model (question-generated question) Predicate POS Question Generation Pattern NP What TYPE is the NP of ENTITY? NP VP What NP is VP by ENTITY? date of birth Barack Obama city place of birth Barack Obama DataOfBirth.BarackObama Type.City PlaceOfBirth.BarackObama

28 Distance on Embedding Space Question Embedding Answer Embedding Question Encoder (word embedding, CNN, RNN, ) Answer Encoder (word embedding, CNN, RNN, ) Natural Language Question KB Answer Entity Candidate

29 score(q, a) embedding of q embedding of a embedding matrix W dot product embedding matrix W binary encoding of q binary encoding of a 1987 Who did Clooney marry in 1987? G. Clooney K. Preston answer candidate Honolulu Model J. Travolta

30 Ranking using a set of Question-Subgraph features Original Question Answer Subgraph Answer Subgraph Extraction Natural Language Question KB Answer Entity Candidate

31

32

33 SP SP AR AR AR SP AR SP SP AR AR SP AR: Answer Ranking-based KBQA 30 SP: Semantic Parsing-based KBQA AR-based methods perform better than SP-based methods 0 F1 Score Berant 2013 Bao 2014 Bordes 2014 Berant 2014 Dong 2015 Yang 2014 Yao 2015 Bao+Yang2015 Berant 2015 Yih 2015 Xu 2016 Cheng2017

34

35

36

37

38

39