LarKC推理与决策

LarKC: The Large Knowledge Collider ( 大规模知识加速器 )

LarKC Website: http://www.larkc.eu

LarKC中文主页: http://cn.larkc.eu

•欧盟第7框架计划(FP7)的LarKC项目的目标是开发大规模知识加速器,LarKC被设计为一个大规模分布式不完备推理平台。按照LarKC (The Large Knowledge Collider) 项目的起源,其中文可译为大规模知识对撞机,原因是这个名字的由来是受到了欧盟原子能研究组织开发的大规模强子对撞机(Large Hadron Collider)的启发。
•项目开始时间:2008年4月1日
项目结束时间:2011年9月30日
项目时长:42个月
项目经费:约1千万欧元
欧盟委员会资助:约700万欧元
成员单位:13个(来自11个国家) 单位, 80人左右
•如果用简短的话来描述LarKC项目的目的,那便是:基于海量数据的分布式、不完全搜索、推理平台。
•分布式体现在数据集( 目前主要处理的是RDF格式的数据 )分布在万维网、本地等不同的来源。
•不完全体现在“在有限时间内,基于海量数据的确定性推理几乎是不可能的”,因此只能“在不完全数据上进行令用户足够满意的推理”。
•平台则体现在LarKC将与基于语义Web的问题求解组件都以插件的形式组织在一起,通过一个管道(Pipe line)进行调用。
•决策器(/plugins/decider ):

简单范本恒时决策器:一个简单范例,在模板插件集上反复执行工作流,在每个循环结束时返回中间结果(在返回最终结果之前)。

模板决策器:一个用来创建新的决策器插件的简单模板。

•„变形器(/plugins/transform ):

面向三元组SPARQL 查询的变形器:返回一系列包含在给定输入SPARQL 查询中的三元组。

模板变形器:一个用来创建新的变形器插件的简单模板。

•识别器(/plugins/identify ) :

Sindice三元组识别器 :为了三元组集合创建一系列RDF 图

模板识别器:一个用来创建新的识别器插件的简单模板。

•选择器(/plugins/select ):

增长数据选择器:为给定系列声明提供完整选择

模板选择器:一个用来创建新的选择器插件的简单模板。

•推理机(/plugins/reason):

Sparql 查询估算推理机:包含 Jena 推理机并能针对给定查询执行SPARQL 选择、组合、描述与问题询查。

模板推理机:一个用来创建新的推理机插件的简单模板。

Jena

Jena由 HP Labs 开发的Java开发工具包, 用于Semantic Web(语义网)中的应用程序开发;Jana是开源的,在下载的文档中有Jena的完整代码。Jena框架主要包括:

a). 以RDF/XML、三元组形式读写RDF;

b). RDFS,OWL,DAML+OIL等本体的操作;(对基于XML语法的文件进行解析)

c). 利用数据库保存数据;(RDF Model)

d). 查询模型;(实现SPARQL查询语言,用于信息查询搜索)

e). 基于规则的推理. (用于检索过程中的推理,这种推理是基于规则的)

Jena的主页:http://openjena.org/

Jena推理机

推理机一般实现两个功能:

①检查本体的一致性, 即保证本体中已获得的类和个体的各条知识之间在关系逻辑上的一致性, 不能互相矛盾。

②得到隐含的知识, 在本体创建过程中, 一般都要遵循一条原则: 尽量简化本体的同时使得本体尽量包含足够多的信息。这时就需要本体推理机来获取本体中隐含的信息。

RDF Model可以直接用于信息检索,但通常情况下要结合推理子系统和Ontology子系统生成具有语义推理能力的InfModel或OntModel.

Jena还提供了ARQ查询引擎,它实现SPARQL查询语言和RDQL,从而支持对模型的查询。另外,查询引擎与关系数据库相关联,这使得查询存储在关系数据库中的本体时能够达到更高的效率。

•Jena自身包含了一系列的推理规则, 主要是针对本体的特点定义的一些规则, 用于检查概念的可满足性,不同类之间的关系, 属性的传递、互逆、不相交等, 例如以下两条规则:
•subClassOf(?x rdfs:subClassOf ?y),(?y rdfssubClassOf ?z) → (?x subClassOf ?z)
•disjointwith(?x owl:disjointWith ?y),(?a rdf:type ?x),(?b rdf:type ?y → (?a owl:differentFrom ?b)
•这些规则可以称为通用规则,但对于具体领域内的一些具体信息的检索还不能够满足要求
Jena自定义规则
•可以定制自己的规则创建特定的推理机以满足需要,自定义规则是对通用规则的补充,也是实际应用领域的个性化需求。例如:
•Rule1:(?x guarantee ?y),(?y subclassof ?z) → (?x guarantee ?z)
•Rule2:(?x workin ?y),(?y guarantee ?z) → (?x guarantee ?z)
•Rule1说明如果人员x具有保障装备y的能力,而y是z的子类,则x具有装备z的保障能力。
•Rule2说明如果人员x属于某单位y,单位y是保障装备z的,那么x具有装备z的保障能力。
基于语义的信息查询流程
(1)提出查询请求:用户通过查询界面提出查询请求,系统据此生成针对全局本体的查询语句;(2)查询扩展:通过本体推理机进行推理,检索出全局本体中与用户要查询的数据语义相同或相似的概念,并重新构造查询语句;(3)查询分解:根据查询控制返回的结果以及全局-局部映射表将全局查询语句分解为对各局部本体的查询,分解后的查询语句是用局部本体的术语描述的;(4)查询转换:通过局部本体到关系数据库的映射,将对局部本体的查询转换为对关系数据库的查询;

(5)生成查询结果:得到查询结果并返回给用户.

查询扩展实现

•传统的查询扩展是基于关键词的,如同义词林.
•基于知识数据的查询, 是实现数据库的语义查询的关键。如自定义规则进行查询扩展.
•下面结合装备保障应用实例, 说明具体实现过程:
•首先, 通过对装备保障数据库的分析,建立装备保障本体, 下图为部分本体结构图。
装备保障领域部分本体结构图

如果在关系数据库中进行查找具有火炮保障能力的人员,则难以实现。通过本体结构图可以看到,王某具有保障“自行火炮”的能力,通过本体推理王某可以作为候选人员。同时,如果李某工作于榴弹炮保障分队则可以作为候选人员输出。上述过程可以使用Jena推荐的本体查询语言SPARQL, 利用之前自定义的两条推理规则实现.

LarKC相关资料来源

•阶段性报告:http://www.larkc.eu/deliverables/

•WP1: 概念性框架及评估;
•WP2: 语义空间信息检索和选择(语义空间模型);
•WP3: 语义网中的统计机器学习;
•WP4: 逻辑推理和决策技术报告;
•WP5: 各样模式的分析设计实现和时间测试;
•… …

WP1: overview of relevant work in other areas

1. Introduction

•The objective of WP1 is the conceptual integration of the different models of reasoning offered by many of the disciplines previously mentioned. In order to tackle this integration work package WP1 will:
•1. Build bridges to relevant work in these disciplines that may provide either single means or patterns of integration of these means.
•2. Define algorithmic schemata as operational instruments for integrating complementary techniques.
•3. Reflect our integration of inductive and deductive techniques into a coherent framework at the meta-scientific level.
•4. Derive a paradigm of heuristic problem solving from this together with means for evaluation and measurement of utility.

2. Approximation Theory

Approximation theory relies on so-called “approximate algorithms”, which aim to find a nearly optimal solution to a specific problem that cannot be solved exactly within a reasonable time.
•Many of these algorithms, including the probabilistic, fuzzy, non-monotonic, case-based, qualitative or granular reasoning, are fundamental to human thinking and behavior.
2.1 Case-based Reasoning
The basic idea of Case-based Reasoning(CBR) is to tackle new problems by referring to similar problems that have already been solved in the past.
•”similar problems have similar solutions“
•CBR uses different forms of approximate reasoning and reasoning under uncertainty, notably probabilistic and fuzzy set-based techniques.
2.2 Probabilistic Reasoning

•(probability reasoning) 推理的一种形式。人们根据不确定的信息做出决定时进行的推理。人们在日常生活中经常会遇到许多不确定的信息,即具有概率性质的信息,若据以推理,便是概率推理。如天阴并不一定意味着要下雨,肚子痛并不一定是得了胃病。

•The probably approximately correct (PAC) learning model is fundamental to computational learning theory. It provides a probabilistic framework for the study of learning and generalization.
2.3 Granular Reasoning
•人类在处理大量复杂信息时.由于人类认知能力有限,往往会把大量复杂信息按其各自特征和性能将其划分为若干较为简单的块.每个被分出来的块就被看成是一个粒。实际上,粒就是指一些个体通过不分明关系、相似关系、邻近关系或功能关系等所形成的块。这种处理信息的过程称信息粒化.
•Granular computing deals with reasoning and information processing with multi-level hierarchical structures. Data, information, and knowledge are arranged in multiple levels according to their granularity. A higher level contains more abstract or general knowledge, whilst a lower level contains more detailed or specific knowledge. Reasoning can be performed on various levels.
2.4 Relevance to LarKC

Reasoning in LarKC must consider multiple strategies. Depending on special requirements or constraints, one may choose a particular strategy as given above. The use of multiple strategies may speed up the reasoning process. Web reasoning also needs to combine rule-based reasoning with case-based reasoning. Web information sources cannot be captured as knowledge (rules), but only in form of cases, due to the complexity and variety of Web information and Web service provisioning. The experience gained from the cases can also be used to improve the modeling and implicitly the rules. Based on this observation, we can develop hybrid reasoning methods on the Web information at multiple levels.

3. Quantum Logics

•Quantum logics are non-classical logics arising from the mathematical formalism of quantum theory. 冯.诺依曼 1936
•量子逻辑属于异常逻辑, 即修改基本公理或推理规则.
•An important application of quantum logics is the “Theories of Quasisets”, first developed by Takeutu in 1981. The basic aim there was to provide a mathematical description for a collection of micro-objects, which can avoid the counter-intuitive properties of the classical identity relation, such as the problem of “extensional equality” with the sense that “two quantum sets that are extensionally equal do not necessarily share all the same properties”.
3.1 Relevance to LarKC
•The semantic space model at the core of work package WP2 (Retrieval and Selection) shares some commonalities with quantum mechanics, i.e., those of quantum logics, in which vector spaces stand in for the neuro-physics and mechanisms such as Hilbert spaces and uncertainty are investigated as vehicles for computation.
4. Meta-Reasoning
Fig4.1: Reasoning as the Component of Problem-solving Process
Fig4.2: Problem-solving Process using Meta-Reasoning
4.1 Relevance to LarKC
•When is it a “good” moment in time to refocus our current (unsuccessful) reasoning process on different parts of the knowledge base or to use different methods to perform a certain inference (subproblem)?” are in fact meta-reasoning problems.
•Meta-reasoning seems to be a promising and conceptually elegant tool for scaling up the reasoning capabilities of existing high-performance reasoners to Web-scale knowledge bases.
5. Cognitive Architectures

•SOAR (State Operator And Result)
•是一个通用的问题求解程序,具有从经验中学习的功能,即能够记住自己是如何解决问题的,并把这种经验和知识用于以后的问题求解过程之中.

(艾伦·纽厄尔(Allen Newell)生前参与的最后一个AI项目)

http://en.wikipedia.org/wiki/Soar_(cognitive_architecture)

•ACT-R (Adaptive Control Thought-Rational)

http://en.wikipedia.org/wiki/ACT-R

5.1 Relevance to LarKC
•The cognitive architectures in question have never been used with amounts of knowledge anywhere close to what LarKC will use. One could tweak the assumptions of the architectures to work with RDF data and investigate how far one can get or which changes or adaptations are required in order to deal with scalability issues
6. Similarity Theories
Approximation theory is concerned with how functions can best be approximated with simpler functions.

参考: http://en.wikipedia.org/wiki/Approximation_theory

Similarity Theory, the science of the conditions under which physical phenomena are similar.

参考:http://encyclopedia2.thefreedictionary.com/Similarity+Theory

6.1 The Trouble with Similarity

What does it mean to say that two objects a and b are similar? One intuitive answer is to say that they have many properties in common. But this intuition does not take us very far, because all objects have infinite sets of properties in common. For example, a plum and a lawnmower both share the properties of weighing less than 100 pounds (and less than 101 pounds, etc). That would imply that all objects are similar to all others (and vice versa, if we consider that they are different in an infinite set of features too).

6.1 The Trouble with Similarity

Goodman proposed that similarity is thus a meaningful concept when defined with respect to a particular criterion. Instead of considering similarity as a binary relation s(a,b), we should think of it as a ternary relation s(a,b,representation), where similarity is defined with respect to a particular representation of the objects. Now the problem becomes what is the representation. There are four psychological theories that tried to solve the problem.

6.2 Metric Models(度量模型)

6.3 Feature Models(特征模型)

6.4 Structural Mapping Models(结构映射模型)

6.5 Transformation Distance Models(转换距离模型)

6.6 Relevance to LarKC

•In a sense, psychologists and semantic web practitioners face similar problems: trying to model the world with a formalism. Knowing what humans consider similar may ultimately help build better systems.

•Both XML and RDF are structured data, which can mapped into metric space, considered as sets of features, are amenable to structure-mapping, and can be compared through transformational distance. Thus, these psychological theories of similarity may be directly relevant to the semantic web.
The Rest Chapters
7. Brain Science/Neuroscience
8. Economics
9. Bounded Rationality – Herbert Simon
10. Patterns and Pattern Languages
11. Distributed and Parallel Computing
12. Software Engineering

WP4: survey of web scale reasoning

1. Introduction
•General background: 自1999年左右起, 语义网中知识的表示和推理成为一项重要挑战. Semantic Web vision的目标是在world-wide规模级把信息结构化. 但通常学者开发的推理方法都是rather small, closed, trustworthy, consistent and static domains. (受限的一阶逻辑子集)
•The goals of LarKC: 开发Reasoning plus-in, 调研现有的reasoning技术方法, 进行扩展使其能够应用到web scale reasoning中. More exactly, 需要调研如何使用这些方法能够实现如下goals:
•1). Scalability; 2).Heterogeneity; 3). Dynamics; 4).Inconsistency; 5).Parallelism.

•要实现这些goals就要采取如下的一些方法:

1. Approximate Reasoning; trade-offs between computation time and answer quality.

2. Resource Bounded Reasoning; Heuristics are simple decision strategies that exploit informational structures in the environment.

3. Rule-based Reasoning for dynamic and incomplete knowledge;

4. Contextual and Modular Reasoning;

5. Distributed and Parallel Reasoning. include P2P architectures and various approaches such as a Web Service.

2. Approximate Reasoning

Fig. Typical knowledge-based architecture

Approximation of the reasoning method

Approximation of the request

Approximation of the knowledge base

3. Resource Bounded Reasoning

•Herbert A. Simon
How rational are people, given their limited computational capabilities and their incomplete knowledge?
•In psychology, two research programs have worked toward answering this question. One program is the heuristics and biases program; the other is the program on fast and frugal heuristics.
•The program on fast and frugal heuristics does not uncritically accept the normative standard of logic, statistics, and probability theory.
4. Rule-based Reasoning for dynamic and incomplete knowledge
NonmonotonicReasoning

“all birds that are not penguins and not ostriches and . . . Fly”

•Default Logic

缺省逻辑致力于形式化上面的推理规则,而不需要明确提及所有的例外。

•Circumscription

它假定除非特殊指定否则事物同预期的一样

•Autoepistemic Logic

自动认识逻辑是致力于形式化关于知识的表示和推理的形式逻辑。命题逻辑只能表达事实,而自动认识逻辑可以表达关于事实的知识和知识的缺乏。

(参 维基百科)

5. Contextual and Modular Reasoning
•Contextual Reasoning

大多数的认知过程都依赖于环境,或称为上下文。

•Modular Reasoning

•Stuckenschmidtand and Klein propose a framework of ontology modularization. They state the motivation of ontology modularization as follows:

Distributed Systems; Large Ontologies; Efficient Reasoning.

6. Distributed Reasoning and Parallel Reasoning

7. Use Cases and Reasoning Tasks

•Previous chapters are all about very general approaches to scalability, but these approaches can in principle be applied to many different reasoning tasks.
•We first identify which reasoning tasks are relevant for Semantic Web reasoning (points 1-3), then we discuss which scalability approaches have been applied to which Semantic Web reasoning tasks (point 4).

1. make categories of use-cases;

2. survey the most prominent Semantic Web system applications and prototypes of the past few years, and classify each into one or more of these use-case categories;

3. link each of the use-case categories to different basic reasoning tasks;

4. discuss the relation between different reasoning tasks and the large scale reasoning approaches from Chapters 2-6 .

• approximate reasoning

• resource-bounded reasoning

• rule-based reasoning for dynamic and incomplete knowledge

• modularity and contextual reasoning

• distributed reasoning and parallel reasoning

7.2 Categorization of Semantic Web Use-cases

the requests the user makes;
the answer the system gives;
the knowledge that it uses for computing this answer.
•A terminology is set of class-definitions (a.k.a. concepts) organized in a subsumption hierarchy; instances are member of such concepts. Ontologies consist of terminologies and instance sets.
7.2 Categorization of Semantic Web Use-cases

We will consider an ontology as a set of triples:

O={< s,p,o >: s∈Subject,p∈Predicate,o∈Object},where{∈,⊆}⊆Predicate

Namely, we consider two specific predicates:⊆ for the subsumption relation, and ∈ for the membership relation.

We use < C1,⊆,C2 >to denote that a class C1 is subsumed by a class C2; We use < i,∈,C >to denote that an individual i is a member of a class C. Thus, a terminology T is a set of triples which predicate

is the subsumption ⊆ only, and an instant set I is a set of triples which predicate is the membership relation ∈only.

•Ontologies may have different versions and different authorship. They may be context dependent. Thus, it is useful to introduce an additional parameter on triple sets to represent additional information of ontologies, like provenance, version, reliability, preference, context, module, etc.
•We define semantic web data as a quadruple set.

WebData={< s,p,o,a >: s∈Subject,p∈Predicate,o∈Object,a∈Thing}

•For example, we use{< s,p,o,O1>,< O1,hasVersion,1.0.0, _>}to represent that triple< s,p,o >belongs to the ontology O1 with the version 1.0.0.
7.2 Categorization: Search
•Perhaps the most prototypical Semantic Web use-case is search.

request: concept c

knowledge:

•Terminology T: set of < ci,⊆,cj >

•Instance set A: set of < i,∈,cj >

answer: members of the instance set:

search(c,(T,A)) ={< i,∈,c >|∃cj(< c,⊆,cj >∈T∧< i,∈,cj >∈A)}

7.2 Categorization: Browse
•Browsing is very similar to searching, but has as crucial difference that its output can either be a set of instances, or a set of concepts, that can be used for repeating the same action (i.e. further browsing)

request: concept c

knowledge:

•Terminology T: set of < ci,⊆,cj >

•Instance set A: set of < i,∈,cj >

answer: members of the instance set or concepts of the ontology:

brows(c,(T,A)) = search(c,(T,A))∪

{< cj,⊆,c >| < cj,⊆,c >∈T}∪{< c,⊆,cj >| < c,⊆,cj >∈T}

7.2 Categorization: Data integration

•The goal of data-integration is to take multiple instance sets, each organized in their own ontology, and to construct a single, merged instance set, organized in a single, merged ontology.

request: multiple ontologies with their instance sets:

{(T,A)| < i,∈,c >∈A∧(< c,⊆,cj >∈T}

knowledge:

answer: a single ontology and instance set

dataintegration({(T1,A1>,..,< Tn,An>}) = (Tck,Ack).

7.2 Categorization: Personalization and recommending

•Personalization consists of taking a (typically large) data set plus a personal profile, and to return a (typically much smaller) data set based on this user profile.

request:

•instance set A: set of < i,∈,cj >

•profileP⊆Ontologies×Ontologies

knowledge: terminology T: set of < ci,⊆,cj >

answer:

reduced instance set:personalisation(P,A,T)⊆A

7.2 Categorization: Web-service selection

•Rather than only searching for static material such as text and images, the aim of semantic web services is to allow searching for active components, using semantic descriptions of web-services.

request: functionality f

knowledge:

•ontology Of: set of < fi,p,fj >

•instance set Oi: set of < servicei,”implement”,fj >

answer: members of the instance set:

service_selection(f,(Of,Oi)) =

{< si,”implement”,f >| < f,p,fj >∈Of∨< fj,p,f >∈Of}

7.2 Categorization: Web-service composition

•An even more ambitious goal than web-service selection is to compose a given number of candidate services into a single composite service with a specific functionality:

request: functionality f

knowledge:

•ontology Of: set of < fi,p,fj >

•instance set Oi: set of < servicei,”implement”,fj >

answer: composition of service instances:

service_composition(f,(Of,Oi)) =

control flow of si’s that satisfy f, where for all si: < si,”implement”,fj >∈Oi

7.2 Categorization: Question answering
•This use-case could also have been called “deducing implicit information:

request: concept c

knowledge:

•terminology T: set of < ci,⊆,cj >

•instance set A: set of < i,∈,cj >

•background knowledge BK: rules about the concepts in T

answer:answer a

Question_answering(c,(T,A)) =

{< i,∈,c >∈A|∃cj(< c,⊆,cj >∈T∧< cj,⊆,c >∈T}∪BK|-a)}

7.2 Categorization: Semantic Enrichment

•This use case concerns about annotation of objects, which can be for instance images, documents, etc. Based on these added meta-data other use-case like search, browse increase in their quality of answers.

request:

•instance i

•meta data of an instance: set of < i,p,c >

knowledge: ontology O: {< s,p,o >:

s∈Subject, p∈Predicate, o∈Object}

answer:

Semantic_enrichment(i,O)⊇O

7.3 Coverage of use-case categorization

The Semantic Web Challenge” is an annual event of the Semantic Web community. http://challenge.semanticweb.org/

•Most applications can be classified in our category of use-cases.
•One application belongs often to more than one use-case.
•In the web challenges from 2005-2007 there were no applications implementing web-service selection
•There is very little work on ”question answering” based on semantic web techniques. SMART is the only question answering application
•Search and browse are use-cases that are often together in an application.
7.4 Semantic Web reasoning tasks

The following table shows how each of the use-cases from section 7.2 can be implemented using the reasoning tasks described in figure.

Search: the description of the Search use-case above shows that it is a combination of classification (to locate the query-concept in the ontology in order to find its direct sub- or super-concepts) followed by retrieval (to determine the instances of those concepts, which form the answers to the query).

7.5 Using the approaches to web-scalability for the reasoning tasks

In principle, each of the approaches (approximate, resource-bounded, rule-based reasoning, modularity and contextual, distributed and parallel) from Chapters 2-6 can be applied to each of the reasoning tasks we identified above, in order to achieve web-scalability.

This would produce the cross-product of all approaches in Chapters 2-6 with the four reasoning tasks.

Such as. approximate subsumption, parallelised realisation, etc

WP4: Implemented plug-ins for interleaving reasoning and selection of axioms

The general scenario of interleaving reasoning and selection consists of the following three steps:

Selection: Use a selector to select part of data.

Reasoning: Use a reasoner to reason over the selected data;

Deciding: Use a decider to decide whether or not the procedure should be stopped and return an answer or go back to the selection step to continue the interleaving processing.

• Query-based selection.

• User-interest-based selection.

• Language-based selection.

Implementation of the DIG Reasoner Plug-in

1. Data Translation. Because the data set imported to a reasoner plug-in in the LarKC platform is designed to be a set of statements. The first step of the DIG reasoner plug-in is to translate a set of statements (ontology data) into a DIG data, so that it can be posted to the external DIG reasoner. If it is an OWL-DL data, the system will use the OWL2DIG library to translate it into a DIG data.

2. Query Translation. Since the query to a reasoner plug-in in the LarKC platform is designed to be a SPARQL query, that query should be translated into DIG queries, so that they can be posted into the external DIG reasoner.

3. Query and answer processing. The DIG reasoner plug-in may have to make several DIG queries to get the complete answer to a given SPARQL query. For example, we cannot express a single DIG query which involves two variables such as SubClassof(x,y). However, SPARQL is expressive to provide a query which involves multiple variables. The reasoning result of a SPARQL query can be obtained by multiple DIG query steps, in which one step is used to obtain variable binding of a single variable, then another step is used to obtain variable binding of another variable by instantiating a variable of the corresponding the DIG query.

4. Translate DIG answers into SPARQL answers. Since the output of a reasoned plug-in is designed to be a SPARQL answer (say, a variable binding for a SPARQL select query), the system have to translate the DIG answers into their SPARQL answers.

关于吴老师问过我的规则和同义词的问题

在我看过的几份技术报告后没有能够让我来明确回答这个问题,不过根据这几篇阅读过的内容比较明确的是系统要追求数据库尽可能的精简不要冗余,同时处理时间又要尽可能的短. 这就存在一个两难问题,数据库要精简,就会必定会不完整,需要大量的规则,而规则又会把查询问题扩展,拖延运行时间。可以考虑在后期对对数据库进行处理,对于特定域来说可能好处理一些,对于开放域问题非完全同义的词汇太多,离线处理又会造成较大的歧义.对于一个完全开放领域去做一个折中考虑也是一个比较难的问题。

参考:LarKC 阶段性报告

.

发表评论

电子邮件地址不会被公开。 必填项已用 * 标注

*

您可以使用这些 HTML 标签和属性: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <strike> <strong>