Answering aggregation question over knowledge base

Presentation on theme: "Answering aggregation question over knowledge base"— Presentation transcript:

1 Answering aggregation question over knowledge base
ywsun

2 Examples Which cities have more than three universities?
Greece held its last Summer Olympics in which year? How many daughters does Obama have? Who was the first president of United States after 2000? What ‘s the total number of songs originally performed by anna nalick?

3 A template-based parser
Which cities have more than three universities? SELECT ?y where{ ?x rdf:type ?c. ?x ?p ?y } HAVING (COUNT(?x) > ?num) Slots: <?c, class, onto:University> <?p, property, onto:city> <?num, numeric, 3> SELECT ?y where{ ?x rdf:type onto:University. ?x onto:city ?y } HAVING (COUNT(?x) >3) Template-based question answering over RDF data. WWW 2012:

4 A logical-form parser Greece held its last Summer Olympics in which year?  R[λx Year.Date.x ].argmax(Country.Greece, Index)  2004 Task: complex questions on semi-structured tables Input: a table t and a question x. Output: an answer y deduction rules Compositional Semantic Parsing on Semi-Structured Tables. ACL (1) 2015:

5 A transition-based parser
Buffer: How, many, daughters, does, Obama, have Stack Action NT choice TER choice NT answer answer( count answer(count( relatives.daughter answer(count(relatives.daughter( TER Obama answer(count(relatives.daughter(Obama RED answer(count(relatives.daughter(Obama) answer(count(relatives.daughter(Obama)) answer(count(relatives.daughter(Obama))) How many daughters does Obama have? answer(count(relatives.daughter(Obama))) Learning Structured Natural Language Representations for Semantic Parsing. ACL (1) 2017: 44-55

6 A multi-constraint query graph generator
Who was the first president of United States after 2000? 𝐶 1 𝐶 2 President 2000 Equal < 𝐶 3 MaxAtN 𝑦 1 1 𝑦 2 title from 𝐵 United States 𝑦 0 𝑥 officials holder Constraint-Based Question Answering with Knowledge Graph. COLING 2016:

7 A sequence-to-SQL parser
该任务的输入是一张web table或一个关系数据库表以及一个关于这张表的自然语言问句，输出是表达该问句语义的SQL语句。这个SQL语句可以在输入的表上被执行，从而得到问题的答案 目前，做生成任务比较流行的方法是基于序列到序列（sequence to sequence）架构的神经模型，这类模型一般由一个编码器（encoder）和一个解码器（decoder）组成。编码器负责建模句子表示，解码器则根据编码器得到的问句表示来逐个从词表中挑选出一个个符号进行生成。 我们提出了一个融合SQL语法的生成式语义解析模型，其整体结构如图2所示。这是一个序列到序列的模型，其编码器由双向的RNN组成，双向RNN的最终状态向量在首尾相连后作为解码器的初始状态。解码器则由三个频道和一个门单元组成。其中三个频道分别为Column、value、SQL频道，在每个频道中分别预测表中列名称、表中单元格名称和SQL语法关键字。而门单元则预测在每个时间节点应该选择哪个频道的预测结果作为输出。解码器在t时刻生成目标ytyt的概率如公式2所示，其中ztzt代表由门单元选择的频道，pzpz(·)是选择频道的概率，而pwpw(·)类似于公式1，它是各自频道的概率输出。 Semantic Parsing with Syntax- and Table-Aware SQL Generation. ACL (1) 2018: