知识库复杂问答综述（很棒的一篇综述）_适合写综述的选题资源-CSDN文库

需积分: 50 28 浏览量 2021-09-11 23:25:48 上传评论收藏 1.04MB PDF 举报

该长文沿用短文的叙述角度，将复杂知识图谱问答方法总结为基于语义解析 (Semantic Parsing-based; SP-based) 的方法和基于信息检索 (Information Retrival-based; IR-based) 的方法。从挑战和解决策略的角度讨论现有的复杂知识图谱问答工作。和原短文相比，该长文将内容从扩展到20页。该长文归纳了更多的论文，主要板块加入了更深入讨论和更多实例细节以及图示表达，另增加了背景知识介绍和技术总结板块。知识库复杂问答（KBQA）是人工智能领域的一个关键任务，旨在从知识库中获取答案来回应用户的问题。早期的研究主要集中在处理简单的查询，而在面对涉及多个主题、复合关系或数值运算的复杂问题时，表现力则显得不足。近年来，研究人员提出了一系列创新方法，专门针对这些问题进行研究。本综述首先介绍了复杂KBQA任务的概念及其相关的背景知识。知识库通常由一系列三元组（subject, relation, object）构成，存储了大量的事实信息。大型知识库如Freebase、DBPedia、Wikidata和YAGO等，已经成为 KBQA 研究的重要资源。在详细介绍部分，文章区分了两种主流的复杂KBQA方法：基于语义解析（SP-based）的方法和基于信息检索（IR-based）的方法。SP-based方法通常通过解析自然语言问题来生成逻辑形式，然后在知识库中执行这些逻辑表达式以找到答案。而IR-based方法则依赖于对问题和知识库中的实体进行匹配，通过搜索和排名策略来找到最相关的答案。文章进一步通过流程设计图解阐述了这两种方法的工作原理，并对比了它们的主要差异和相似性。例如，SP-based方法可能更适用于理解复杂的逻辑结构，但可能对语言理解和解析的准确性要求较高；而IR-based方法可能在效率上占据优势，但在处理复杂逻辑和计算问题时可能显得不足。接下来，综述分析了这两种方法在应对复杂问题时面临的挑战，包括语言理解和解析的复杂性、知识表示的局限性、以及数值计算的准确性和效率问题。文中详述了现有研究中采用的先进解决方案和技术，如深度学习模型用于增强语言理解、多跳推理技术处理多步逻辑操作、以及半监督和无监督学习策略来弥补标注数据的不足。文章总结了复杂KBQA领域的未来研究方向，如提高对自然语言的语义理解能力，发展更高效和精确的推理机制，以及如何结合不同的方法来提升整体性能。此外，还探讨了如何更好地集成和利用未结构化信息以及实时更新的知识库，以适应不断变化的知识需求。这篇综述为理解复杂KBQA任务的最新进展提供了一个全面的视角，不仅总结了相关研究，还深入讨论了技术和方法，对于研究人员和从业者来说，是一份极有价值的信息资源。

资源推荐

资源详情

资源评论

JOURNAL OF L

X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 1

Complex Knowledge Base Question Answering:

A Survey

Yunshi Lan*, Gaole He*, Jinhao Jiang, Jing Jiang, Wayne Xin Zhao Member, IEEE,

and Ji-Rong Wen Senior Member, IEEE,

Abstract— Knowledge base question answering (KBQA) aims to answer a question over a knowledge base (KB). Early studies mainly

focused on answering simple questions over KBs and achieved great success. However, their performance on complex questions is

still far from satisfactory. Therefore, in recent years, researchers propose a large number of novel methods, which looked into the

challenges of answering complex questions. In this survey, we review recent advances on KBQA with the focus on solving complex

questions, which usually contain multiple subjects, express compound relations, or involve numerical operations. In detail, we begin

with introducing the complex KBQA task and relevant background. Then, we describe benchmark datasets for complex KBQA task and

introduce the construction process of these datasets. Next, we present two mainstream categories of methods for complex KBQA,

namely semantic parsing-based (SP-based) methods and information retrieval-based (IR-based) methods. Speciﬁcally, we illustrate

their procedures with ﬂow designs and discuss their major differences and similarities. After that, we summarize the challenges that

these two categories of methods encounter when answering complex questions, and explicate advanced solutions and techniques

used in existing work. Finally, we conclude and discuss several promising directions related to complex KBQA for future research.

Index Terms—Knowledge base question answering, knowledge base, question answering, natural language processing, survey.

1 INTRODUCTION

nowledge base (KB) is a structured database that con-

tains a collection of facts (alias triples) in the form (sub-

ject, relation, object). Large-scale KBs, such as Freebase [1],

DBPedia [2], Wikidata [3] and YAGO [4], have been con-

structed to serve many downstream tasks. Among them,

knowledge base question answering (KBQA) is a task that

aims to answer natural language questions with KBs acting

as its knowledge source. Nowadays, KBQA has attracted

intensive attention from researchers as it plays an important

role in many intelligent applications. For example, Amazon

Alexa, Apple Siri, and Microsoft Cortana are integrated

with the function of answering factoid questions from users.

Chatbots such as Microsoft Xiaoice and Zo have also demon-

strated a high degree of conversational capability, where

factoid question answering frequently occurs [5], [6].

Early work on KBQA focused on answering a simple

question, where only a single fact is involved. For example,

“Who was the nominee of The Jeff Probst Show?” is a simple

question which includes the subject “The Jeff Probst Show”,

the relation “nominee” and queries about the object entity

“Jeff Probst” of fact “(The Jeff Probst Show, nominee, Jeff Probst)”

• * Y. Lan and G. He contribute equally to this work.

• This work is done when Y. Lan was with School of Computing and

Information System, Singapore Management University. E-mail: ys-

lan.2015@phdcs.smu.edu.sg

• Jing Jiang is with School of Computing and Information System, Singa-

pore Management University. E-mail: jingjiang@smu.edu.sg.

• G.He is with the School of Information, Renmin University of China, and

Beijing Key Laboratory of Big Data Management and Analysis Methods.

E-mail: hegaole@ruc.edu.cn

• W.X. Zhao (corresponding author) , Jinhao Jiang and J. Wen are with

Gaoling School of Artiﬁcial Intelligence, Renmin University of China, and

Beijing Key Laboratory of Big Data Management and Analysis Methods.

E-mail: batmanﬂy@gmail.com.

Manuscript revised xxx.

The Jeff

Probst Show

producer

spouse

is_a

married_date

nominee

Jeff Probst

Shelley

Wright

1996

Lisa Ann

Russell

married_date

2011

spouse

is_a

Talk show

Survivor

host

CBS Television

Distribution

distributed by

m ulti-hop num erical

constrained

FIGURE 1: An example of complex KBQA for the question

“Who is the ﬁrst wife of TV producer that was nominated for The

Jeff Probst Show?”. We present the related KB subgraph for

this question. The ground truth path heading to the answer

is annotated with colored borders. The topic entity and the

answer entity are shown in the bold font and shaded box

respectively. “multi-hop” reasoning, “constrained” relations

and “numerical” operation are highlighted in black dotted

box. We use different colors to indicate different reasoning

hops to reach each entity from the topic entity.

in KBs. Similarly, simple question “Who is the wife of Jeff

Probst?” can be answered using just the facts “(Jeff Probst,

spouse, Shelley Wright)” and “(Jeff Probst, spouse, Lisa Ann

Russell)”. It is not trivial to retrieve the correct entity from

the large-scale KBs, which consists of millions or even bil-

lions of facts. Therefore, researchers have spent much effort

in proposing different models to answer simple questions

over KBs [7], [8], [9], [10], [11].

Recently, researchers started paying more attention to

arXiv:2108.06688v1 [cs.CL] 15 Aug 2021

JOURNAL OF L

X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 2

User

SP-based Methods

𝑞 → 𝑙

→ 𝒜

Jeff Probst

Question

Who is the wife of Jeff

Probst?

SPARQL

SELECT ?name

WHERE {

ns:m.02pbp9 ns:people.marriage.spouse ?e1.

}

Question-specific Graph

IR-based Methods

𝑞& → 𝒢

→ 𝒜

Answers

Shelley Wright

Lisa Ann Russell

Shelley

Wright

Lisa Ann Russell

nationality

spouse

Agent

Data

Module

Answer

Prediction

Entity

Linking

FIGURE 2: Architecture of KBQA systems. The entity linking procedure is shown in red color.

answering complex questions over KBs, i.e., the complex

KBQA task [12], [13]. Complex questions usually contain

multiple subjects, express compound relations, or include

numerical operations. Take the question in Figure 1 as an

example. This example question starts with the subject “The

Jeff Probst Show”. Instead of querying a single fact, the

question requires the composition of two relations, namely,

“nominee” and “spouse”. This query is also associated with

an entity type constraint “(Jeff Probst, is a, TV producer)”. The

ﬁnal answer should be further aggregated by selecting the

possible candidates with the earliest marriage date. Gener-

ally, complex questions are questions involving multi-hop

reasoning, constrained relations or numerical operations.

Tracing back to the solutions for simple KBQA task,

a number of studies from two mainstream approaches

have been proposed. We show the overall architecture of

simple KBQA systems in Figure 2. These two approaches

ﬁrst recognize the subject in a question and link it to an

entity in the KB (referred to as the topic entity). Then

they derive the answers within the neighborhood of the

topic entity by either executing a parsed logic form or

reasoning in a question-speciﬁc graph extracted from the

KB. The two categories of methods are commonly known as

semantic parsing-based (SP-based) methods and informa-

tion retrieval-based (IR-based) methods in prior work [7],

[8], [11], [14]. They design different working mechanisms

to solve the KBQA task. The former approach represents

a question by a symbolic logic form, and then executes

it against the KB to obtain the ﬁnal answers. The latter

approach constructs a question-speciﬁc graph delivering

the comprehensive information related to the question, and

ranks all entities in the extracted graph based on their

relevance to the question. For example, the SP-based meth-

ods would ﬁrst transfer the question “Who is the wife of

Jeff Probst?” into an executable SPARQL query (shown in

Figure 2). The entities “Shelly Wright” and “Lisa Ann Russell”

are retrieved by executing the SPARQL query in the KB.

Alternatively, the IR-based methods ﬁrst identify a question-

speciﬁc graph (see Figure 2) containing all the candidate

entities that are directly connected to “Jeff Probst” such as

“TV producer”, “US”, “Shelley Wright” and so on. The answer

entities are selected based on the ranking of these candidate

entities.

However, when applying the two mainstream ap-

proaches to the complex KBQA task, complex questions

bring in challenges on different parts of the approaches. We

identify the main challenges as follows:

• Parsers used in existing SP-based methods are difﬁ-

cult to cover diverse complex queries (e.g., multi-hop

reasoning, constrained relations, and numerical opera-

tions). Similarly, previous IR-based methods may fail to

answer a complex query, as their ranking is performed

over small-scale entities without traceable reasoning.

To cover as many complex questions as possible, the

development of an expressive logic form for SP-based

methods or strong reasoning ability for IR-based meth-

ods is highly demanded.

• More relations and subjects in complex questions in-

dicate a larger search space of potential logic forms

for parsing, which will dramatically increase the com-

putational cost. Considering the complex question in

Figure 1, all logic forms with 2-hop relations involv-

ing “The Jeff Probst Show” should be enumerated and

different combinations with the constraint object “TV

producer” should also be considered. Meanwhile, more

relations and subjects prevent IR-based methods from

retrieving all relevant entities for ranking, which makes

the common incomplete KB issue become severer.

• Both approaches involve question understanding as

a primary step. When questions become complicated

from both semantic and syntactic aspects, models are

required to have strong capabilities of natural lan-

guage understanding and generalization. Comparing

the question “Who is the ﬁrst wife of TV producer that

was nominated for The Jeff Probst Show?” with another

question “Who is the wife of the ﬁrst TV producer that

was nominated for The Jeff Probst Show?”, the models are

supposed to understand that the ordinal number “ﬁrst”

is used to constrain “wife” or the phrase “TV producer”.

• It is expensive to label the ground truth paths heading

to the answers (see the annotated ground truth path

in Figure 1) for KBQA task. Generally, only question-

answer pairs are provided. This indicates SP-based

methods and IR-based methods have to be trained

without the annotation of correct logic forms and rea-

soning paths. Such weak supervision signals bring dif-

ﬁculties to both approaches due to the lack of necessary

guidance during the intermediate reasoning process.

Excessive spurious reasoning paths may head to the

correct answers but semantically mismatch the ques-

tions.

Regarding the related surveys, we observe Wu et al. [15]

and Chakraborty et al. [16] reviewed the existing work on

simple KBQA. Furthermore, Fu et al. [17] investigated the

current advances on complex KBQA. They provided a gen-

eral view of advanced methods only from the perspective

JOURNAL OF L

X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 3

of techniques and more focused on application scenarios

in e-commerce domain. Different from these surveys, our

work tries to identify the challenges encountered in previ-

ous studies, and extensively discuss existing solutions in a

comprehensive and well-organized manner. Speciﬁcally, we

categorize the methods for complex KBQA into two main-

stream approaches based on their working mechanisms. We

decompose the overall procedure of the two approaches into

a series of functional modules, and analyze the challenges in

each module. Such a way is particularly helpful for readers

to understand the potential challenges and solutions for

complex KBQA. It is worth noting that this survey is an

extended version of the short survey [18]. As a comparison,

this survey has several main differences: (1) This paper

introduces traditional approaches, preliminary knowledge,

and evaluation protocol of KBQA in multiple aspects and a

more concrete way, which goes far beyond the scope of the

short survey; (2) Based on the two mainstream categories of

KBQA methods, this paper provides a more comprehensive

discussion about their similarity and difference regarding

core modules and working mechanisms; (3) To have a

better view of existing solutions for various challenges, this

paper includes additional subsections to summarize tech-

nical contributions about complex KBQA methods; (4) This

paper gives a more thorough outlook on several promising

research directions and unaddressed challenges.

The remainder of this survey is organized as follows.

We will ﬁrst introduce the task formulation and prelim-

inary knowledge about the task in Section 2. After that,

we introduce multiple available datasets and how they

are constructed in Section 3. Next, we introduce the two

mainstream categories of methods for complex KBQA in

Section 4. Then following the categorization, we ﬁgure out

typical challenges and solutions to these challenges in Sec-

tion 5. Then, we discuss several recent trends in Section 6.

Finally, we conclude and summarize the contribution of this

survey in Section 7.

2 B ACKGROUND

In this section, we ﬁrst brieﬂy introduce KBs and task

formulation of KBQA, then we talk about the traditional

approaches for KBQA systems and the general evaluation

protocol of the systems.

2.1 Knowledge Base

As mentioned earlier, KB is usually in the format of triples.

They are designed to support modeling relationships be-

tween entities. Take Freebase [19] as an example for KB.

Each entity in Freebase has a unique ID (refereed to as

mid), one or more types, and uses properties from these

types in order to provide facts [3]. For example, the Free-

base entity for person Jeff Probst has the mid “m.02pbp9”

and the type “people.marriage” that allows the entity to

have a fact with “people.mar riage.spouse” as the property

and “m.0j6d0bg” (psychotherapist Shelley Wright) as the

value. Freebase incorporates compound value types (CVTs)

to represent n-ary (n > 2) relational facts [1] like “Jeff

Probst was married to Shelley Wright in 1996”, where three

1. PREFIX: < http : //rdf.freebase.com/ns/ >

entities, namely “Jeff Probst”, “Shelley Wright” and “1996”,

are involved in a single statement. Different from entities

which can be aligned with real world objects or concepts,

CVTs are artiﬁcially created for such n-ary facts.

In practice, large-scale open KBs (e.g., Freebase and

DBPedia) are published under Resource Description Frame-

work (RDF) to support structured query language [3], [20].

To facilitate access to large-scale KBs, the query language

SPARQL is frequently used to retrieve and manipulate data

stored in KBs [3]. In Figure 2, we have shown an executable

SPARQL to obtain the spouses of entity “Jeff Probst”.

Different KBs are designed with different purposes, and

have varying properties under different schema design.

For example, Freebase is created mainly by community

members and harvested from many resources including

Wikipedia. YAGO [21] takes Wikipedia and WordNet [22]

as the knowledge resources and covers taxonomy of more

general concepts. WikiData [3] is a multilingual KB which

integrates multiple resources of KBs with high coverage and

quality. A more comprehensive comparison between open

KBs is available at [23].

2.2 Task Formulation

Formally, we denote a KB as G = {he, r, e

i|e, e

∈ E, r ∈ R},

where he, r, e

i denotes that relation r exists between subject

e and object e

, E and R denote the entity set and relation

set, respectively.

Given the available KB G, this task aims to answer natu-

ral language questions q = {w

, w

, ..., w

} in the format of

a sequence of tokens and we denote the predicted answers

. Specially, existing studies assume the correct answers

derive from the entity set E of the KB. Unlike answers

of simple KBQA task, which are entities directly connected

to the topic entity and A

⊂ E , the answers of the complex

KBQA task are entities multiple hops away from the topic

entities or even the aggregation of them. Generally, a KBQA

system is trained using a dataset D = {(q, A

)}.

2.3 Traditional Approaches

General KBQA systems for simple questions have a pipeline

framework as displayed in Figure 2. The preliminary step

is to identify topic entity e

of the question q, which aims

at linking a question to its related entities in the KBs. In

this step, named entity recognition, disambiguation and

linking are performed. It is usually done using some off-

the-shelf entity linking tools, such as S-MART [24], DBpedia

Spotlight [25], and AIDA [26]. Subsequently, an answer

prediction module is leveraged to predict the answer

taking q as the input.

For simple KBQA task, the predicted answers are usu-

ally located within the neighborhood of the topic entities.

Different features, as well as methods, are proposed to rank

these candidate entities. Since the answer prediction module

usually takes questions with the detected topic entities as

the input and outputs predicted answers, existing relation

extraction tools like OpenNRE [27] can also be directly

utilized as the answer prediction module but may encounter

domain inconsistency issue and result in poor performance.

Therefore, this module is usually designed and trained

using the question and the ground truth. Most of the effort

剩余19页未读，继续阅读

评论收藏

内容反馈

syp_net

粉丝: 158
资源: 1187

知识库复杂问答综述（很棒的一篇综述）

基于知识图谱的问答系统综述

问答系统综述论文5篇论文，硕士论文和期刊论文

问答系统研究综述

《知识图谱:构建到应用》综述论文（2020年）

知识图谱综述—中文版

2017年，知识图谱综述文章

知识图谱研究进展综述 (2).docx

知识图谱前沿跟进，看这篇就够了，Philip S. Yu 团队发布权威综述，六大开放问题函待解决！

关于知识表示的6篇论文

知识图谱80篇精选论文集.rar

基于卷积神经网络的自动问答.pdf

神经自然语言处理进展综述：建模，学习，推理

开放式关系抽取综述 - review on Open IE.zip

A Survey of Large Language Model 中文版

nlp2014-2017年重要论文合集

博客中聚类算法（K-means、FCM、DBSCAN、DPC）的数据集（免积分）

机器学习期末复习题及答案

神经网络回归预测--气温数据集

Mathwork+Matlab+编程手册

Ollama软件windows安装包(版本0.3.10)

中文短信数据集-带标签

时间序列预测模型实战案例(Xgboost)(Python)(机器学习)包括时间序列预测和时间序列分类，点击即可运行！

亚博K210模型训练部署

Plecs电力电子仿真PLECS41.64 电力系统仿真软件免安装版本

shape_predictor_68_face_landmarks.zip

hugging face的models-openai-clip-vit-large-patch14文件夹

改进版的yolov5+双目测距

最新资源