细致介绍了一种基于知识图谱的问答系统，构建了一个推理模型，在建模问题回答过程中是十分有效的。
Why QA 知識 工場 QA application One of the most natural human-computer 吴)8 interaction Key components of Chatbot, which attracts wide research interests from industries · QA for A ? One of most important tasks to evaluate the machine intelligence: Turing test Important testbed of many al techniques such as machine learning, natural C language processing, machine cognition Turing test kw.fudan. edu. cn/ga Why KBQA? 知識 工場 More and more knowledge bases are created Google knowledge graph yago wordNet free base, probase, nell, cyc, dbpedia Large scale, clean data rue knowle?e 4 WolframAlpha name name Barack Obama Honolulu dob pob population 1961 390K eNam线 category category Politician City SUM可 ReadThe web eNTE:=:PubMEd category name CY线aG Michelle 酸忘, Person Obama / freebase UniProt BabelNet date person dob 1992 1964 WordNet I Text Runner/ 三 category category Reverb Event Person tSetse AWakiTakonomyi WiniNet A piece of knowledge base, which consist of The boost of knowledge bases triples such as (d, population, 390k) kw. fudan. edu. cn/ga How KB-based QA works? 知識 工場 Convert natural language questions into structured queries over knowledge bases How many people live in Honolulu? SPARQL name name Barack obama Honolulu Select ?number dob population 1961 390K Where i category Res Honolulu Politician category City dbo: population ?num category name Person Michelle Obama person date dob 1992 1964 category category SQL Event Person Select value Key: predicate inference From KB Where subject=d and predicate=population kw.fudan. edu. cn/qa Two challenges for predicate inference D 知識 工場 Question Representation Identify questions with the same semantics Distinguish questions with different intents Semantic matching Map the question representation to the predicate in the KB ° Vocabulary gap Question in Natural language Predicate in KB a How many people are there in Honolulu? population b What is the population of Honolulu? population CWhat is the total number of people in Honolulu? population d) When was Barack Obama born? dob e who is the wife of barack obama? marriage->person-name O When was Barack Obama's wife born? marriage→ >person→name kw.fudan. edu. cn/ga Weakness of previous solutions OE 知識 工場 Template/rule based approaches Neural network based approaches Questions are strings Questions are numeric Represent questions by string based Represent questions by numeric templates, such as regular expression embeddings By human labeling By learning from corpus PROS PROS: User-controllable Feasible to understand diverse Applicable to industry use questions CONs: CONS: Costly human efforts Poor interpretability Not good at handling the diversity of Not controllable. Unfriendly to industrial questions application. How to retain advantages from both approaches? kw.fudan. edu. cn/qa Our approach 知識 工場 Representation: concept based templates Questions are asking about entities Interpretable User-controllable Question What is the population How many people are What is the number of of honolulu? there in honolulu? people in Honolulu? Template What is the population How many people are hat is the total numbel of $city? there in $City of people in city Learn templates from QA corpus, instead of manfully construction 27 million templates, 2782 intents Understand diverse questions kw. fudan. edu. cn/ga System Architecture 知識 工場 Offline procedure value Learn the mapping from templates to user Online Procedure predicates: P(p t Question Input: qa corpora, large scale taxonomy KB Complex Question Parsing tit ● Output:P(P|T) 合/ predicate Probabilistic Inferencing -f-eeee(Template Repository: P(PD------- Online procedure Template Extraction for Offline Predicates Parsing, predicate inference and answer rocedure Predicate Entity value retrieval Expansion Identification Input: binary factoid questions ( BFQs) Knowledge QAc。rpus Output: answers in KG Problem model 知識 工場 Given a question g, our goal is to find an answer y with maximal probability ( v is a simple value) arg m axP(v=vh=g arg m ax P(v/l, e, tp t, p e: entity t: template; p: predicat Basic idea: We proposed a generative model to explain how a value is found for a given question, Rationality of probabilistic inference uncertainty(e.g. some questions intents are vague) Incompleteness (e.g. the knowledge base is almost always incomplete) noisy(e.g. answers in the Qa corpus could be wrong