比较问句的解析与求解 任务 定义 评价方法 常用问题数据集中比较问句模式

Presentation on theme: "比较问句的解析与求解 任务 定义 评价方法 常用问题数据集中比较问句模式"— Presentation transcript:

1 比较问句的解析与求解 任务 定义 评价方法 常用问题数据集中比较问句模式
解析: 解析比较问句的结构(如PAS, 五元组); 识别其类别. 拟解析方法: 独立于知识库; 依赖于知识库. 求解: 求解比较问句的答案. 通过把比较结构转化为知识库查询得到比较问句答案. 拟求解方法: EL/CL + PM+ Operation. 定义 什么是比较问句: 表达不同的objects(E|S-E|S|V)在某一方面比较(可量化)的问句 比较问句分哪些类别: 按语法: Adjectival Comparative; Verbal Comparative; Nominal Comparative; Adverbial Comparative. 按功能: >, >=, <, <= 比较问句的结构是什么: PAS 结点: Predicate(Comparative morphemes); Argument(E/C/V/P/Event) 边: Figure, Ground, Scale, Differential 评价方法 数据集 已有工作(6个baselines) 常用问题数据集中比较问句模式 挑战 Any statement comparing two or more entities , expressing some kind of measurement on a scale, or indicating some degree of having a measurable property.

2 Adjectival Comparatives #A VB (no) JJR than #B 6 20 --- 3 45 1.4w
类别 模式 GraphQ (135/2608; 134/2558) =269 ComplexWQ (209/3476; 182/3480;1572/27735)=1963 WikiQ (170) Yahoo (13w/448w) Train Test Dev Adjectival Comparatives #A VB (no) JJR than #B 6 20 --- 3 45 1.4w Adverbial Comparatives #A VB RBR than #B 8 Nominal Comparatives NP 122 84 910 78 93 32 Verbal Comparatives #A VB more than #B 1 Others 26(PP, exceed) 849(PP, exceed) 98(PP) 112(PP) 3, 38 32(#A or #B) 2.8w(#A or #B) 名词短语 1. PP[IN NP]: PP[after NP],PP[below NP],PP[from NP],PP[before NP],PP[in NP], PP[by NP], PP[under NP], PP[since NP], PP[over NP] 2. Gradable Adjectives 3. Comparative morphemes 4. 图是怎么构造; 子图怎么转成自然语言问题 5. GraphQ: 构图 随机的选择compatible node（值域是float,int,double）, 加约束。 转换成NL，通过人的转换。 6. ComplexWQ: 构图： 答案集合>1; 寻找numerical property common to all answers; 加a triplet and restrictor to r; 转换时候先机器转换成人可理解的,然后由人修正 7. 最后结论是什么? 8. 什么语言现象可以构造成一条边? 9. NP: at most NP ; at least NP ; NP[NP PP] ; NP[(JJR|RBR) than #A] ; NP[NP ADJP]

3 ▪ Types of comparatives
工作 任务 动机 定义 技术贡献 方法 实验 例子 2006-SIGIR 识别是不是比较句子 JJR,RBR 并非全是比较句; 没有的这些的也可能是比较句 比较句子 比较分类(4分类) entity, features 比较结构(5元组) 构造Pattern Class Sequential rules + NB classifier 数据集: 论坛, 评论, 新闻 Baselines: POS tags; CSR, SVM, keywords 结果: F1:75 Canon’s optics are better than those of Sony and Nikon 2006-AAAI (扩展版) 识别是不是比较句子, 同时识别比较成分(5元组) Class Sequential rules (CSR);label sequential rules(LSR); manual rules, 数据集:论坛, 评论, 新闻 Baselines: CRF 结果: 识别F1:80; 抽取F1:72; 2010-ACL /2013-TKDE 从比较问句中挖掘比较实体(识别与抽取同时做) 查找可对比实体 比较问句 比较元素 学Pattern (bootstrapping) 基于Sequential patterns来识别比较问句和抽取比较元素 数据集: yahoo子集 Baselines: 2006 结果: 识别F1:82.5; 抽取F1:83.3 Which is better #A or #B? 2015-EMNLP 预测比较结构 深层次理解比较句子 比较结构PAS 分类模型+ILP 全局约束 Multi-class classifier(P) LR classifier(R) LR classifier(S) Joint inference 数据集: 531句子(OntoNotes子集) Baselines: POS tags 结果: P-F1:79; R-F1:54 Mary is taller than Susan. 2016-CoNLL(扩展版) 联合预测比较和省略结构 深层次理解比较句子,同时考虑省略 概率估计; 联合模型 Log-linear model估算概率分布; Joint inference 数据集: 2800句子(评论) Baselines: rule-based; 结果: P-F1:76; A-F1:43 Mazda drove faster than Hyundai 归纳: 任务越来越难; 技术贡献偏学习; 2006-定义: A comparative sentence is a sen-tence that expresses a relation based on similarities or differences of more than one object. A comparison can be between two or more objects, groups of objects, one object and the rest of the objects. It can also be between an object and its previous or future versions 2006-定义 ▪ Types of comparatives - Gradable comparatives: The gradable types are defined based on the relationships of greater or less than, equal to, and greater or less than all others ◦ Non-Equal Gradable (比较对象之间有顺序上的差异): Relations of the type greater or less than that express an ordering of some objects with regard to certain features. ◦ Equative: Relations of the type equal to that state two objects as equal with respect to some features ◦ Superlative: Relations of the type greater or less than all others that rank one object over all others - Non gradable comparatives: Sentences which compare features of two or more objects, but do not grade them. ◦ Object A is similar to or different from Object B with regard to some features(只是说明两个事 物有差异，但没有高低、优劣之分) 例: 途安和毕加索的风格特点、细致程度存在差异 ◦ Object A has feature F1, Object B has feature F2 (F1 and F2 are usually substitutable) ◦ Object A has feature F, but object B does not have 2006: 将比较句的识别看作一个分类问题，利用 关键词 (收集了83个关键词) 和 CSR 挖掘 POS序列得到的模板 作为特征，用朴素贝叶斯进行分类(比较类、非比较类) 文献[5]在文献[4]的基础上作了一些扩展工作: • Identify comparative sentences and classify the identified comparative sentences into different types • Extract comparative relations: the extraction of entities and their features that are being compared, and comparative keywords Definition (comparative relation): A comparative relation captures the essence of a comparative sentence and is represented with the following: ( relationWord, features, entityS1, entityS2, type) relationWord: The keyword used to express a comparative relation in a sentence features: a set of features being compared entityS1 and entityS2: Sets of entities being compared. Entities in entityS1 appear to the left of the relation word and entities in entityS2 appear to the right of the relation word type: non-equal gradable, equative or superlative 例: Canon’s optics is better than those of Sony and Nikon = (better, {optics}, {Canon}, {Sony, Nikon}, non-equal gradable) 本文提出semantic framework来表示the meaning of comparative structure. 他把comparisons表示成predicate-argument structures inter-connected with semantic roles. Types of Comparison : Comparatives, superlatives, equatives, excessive, and assetives. John is taller than Sam. The table is longer than the sofa is wide.