信息检索：相关性的生成理论

信息检索

需积分: 3 102 浏览量更新于2024-08-01 收藏 2.14MB PDF 举报

身份认证购VIP最低享 7 折!

领优惠券(最高得80元）

"《信息检索：相关性的生成理论》是由Victor Lavrenko所著的一本书，属于《信息检索系列》第26卷。该书深入探讨了在大型、松散结构或无结构的集合中搜索有用信息的科学领域——信息检索（Information Retrieval, IR）。书中可能包含了对信息检索理论的生成模型的详细阐述，特别是关于文本检索和相关性判断的方面。此书的部分内容可能引用了1998年美国广播公司、有线新闻网络和宾夕法尼亚大学信托人的版权材料，并且注明了相关的 ACM 计算分类。该书由Springer-Verlag Berlin Heidelberg出版社于2009年出版，遵循德国版权法的规定，复制或部分复制需获得许可。" 《信息检索：相关性的生成理论》这本书是信息检索领域的经典之作，作者Victor Lavrenko来自英国爱丁堡大学的信息学院。信息检索作为一门科学，其核心在于从大量的、非结构化的数据中寻找并提供相关信息。在20世纪70年代，这个领域刚刚起步时，主要关注的是图书馆中的文本信息。书中可能涵盖了信息检索的基本概念、模型和技术，包括查询分析、文档索引、相似度计算以及用户行为分析等。生成理论部分可能涉及如何建立一个能够模拟用户查询意图和文档内容的相关性模型，这在现代搜索引擎中至关重要。作者可能讨论了语言模型、概率检索模型以及基于机器学习的方法，这些方法能够根据用户的反馈不断优化检索结果的准确性。此外，书中提到的图片可能涉及信息检索系统的工作流程示例、检索结果的可视化展示，或者是关于信息检索系统的架构设计。ACM计算分类（1998）将这本书归类为H.3（信息存储与检索）、G.3（一般系统理论）和F.2（信息处理系统），这表明它涵盖了信息技术的多个层面。这本书对于理解信息检索的理论基础、相关性评估和实际应用具有重要意义，无论是对于研究人员、学生还是在信息管理和搜索技术领域工作的专业人士，都是一本有价值的参考资料。通过阅读本书，读者可以深入理解信息检索的复杂性和挑战，以及如何构建更智能、更精准的信息检索系统来满足用户的需求。

资源详情

资源推荐

List of Tables

3.1 Notation: frequently used symbols andexpressions. ........... 44

5.1 Examples of mono-lingual relevance models for the TDT2

topics. Probabilities are estimated using techniques from

section 5.1.1, starting with just the topic title. For each topic

we show 10 words with highest estimated probabilities under

therelevancemodel. ......................................107

5.2 Information for the corpora used in ad-hoc retrieval

experiments. dl denotes average document length, cf stands

for average collection frequency of a word, and ql represents

averagenumber ofwordsperquery. ........................109

5.3 Comparison of Relevance Models (RM) to the baseline

systems: (tf.idf) Okapi / InQuery weighted sum, (LCA)

tf.idf with Local Context Analysis, (LM) language model

with Dirichlet smoothing, (LM+X) language model with

Ponte expansion. Relevance Models noticeably outperform

all baseline systems. Stars indicate statistically signiﬁcant

diﬀerences in performance with a 95% conﬁdence according to

the Wilcoxon test. Signiﬁcance tests are performed against the

tf.idf baseline.............................................112

5.4 Comparison of Relevance Models (RM) to the InQuery

(tf.idf) and language-modeling (LM) systems. Relevance

Model signiﬁcantly outperforms both baselines. Stars indicate

statistically signiﬁcant diﬀerences in performance with a 95%

conﬁdence according to the Wilcoxon test. Signiﬁcance is

against the tf.idf baseline. . . . . .............................114

XX List of Tables

5.5 Comparison of Relevance Models (RM) to the InQuery

(tf.idf) and language-modeling (LM) systems. Relevance

Model signiﬁcantly outperforms both baselines. Stars indicate

statistically signiﬁcant diﬀerences in performance with a 95%

conﬁdence according to the Wilcoxon test. Signiﬁcance is

against the tf.idf baseline. . ................................115

5.6 Contrasting performance of probability-ratio and cross-

entropy ranking. Cross-entropy results in substantially better

performanceonallﬁvedatasets involved.....................116

5.7 Composition of the BBN bilingual lexicon . . . . . . .............126

5.8 Composition of the parallel corpus used in our experiments. . . . 126

5.9 Average Precision on the TREC9 cross-language retrieval

task. Cross-lingual relevance models perform around 95% of

thestrong mono-lingual baseline ...........................128

5.10 Initial precision on the TREC9 CLIR task. Cross-lingual

Relevance Models noticeably outperform the mono-lingual

baselines. ...............................................129

5.11 Parallel corpus size has a very signiﬁcant eﬀect on the quality

of Cross-lingual Relevance Models. Adding the pseudo-parallel

TDTcorpusmorethan doubledtheaverageprecision. ........131

5.12 Comparing recall and precision of the four models on the

task of automatic image annotation. CRM substantially

outperforms all other models. Percent improvements are over

the second-best model (CMRM). Both CRM and CMRM

signiﬁcantly outperform the state-of-the-art translation model. . 145

5.13 Comparing CRM and CMRM on the task of image retrieval.

Our model outperforms the CMRM model by a wide margin

on all query sets. Boldface ﬁgures mark improvements that are

statistically signiﬁcant according to sign test with a conﬁdence

of 99% (p-value < 0.01)....................................146

5.14 Performance of CRM on the TREC video retrieval task. . . .....150

5.15 Summary statistics for the ﬁve NSDL ﬁelds used in our

retrievalexperiments. .....................................155

5.16 Performance of the 63 test queries on the evaluation data. Bold

ﬁgures show statistically signiﬁcant diﬀerences. Evaluation is

based on retrieving 1000 documents. Across all 63 queries,

there a total of 1253 relevant documents. ....................158

5.17 Performance of diﬀerent algorithms on the Link Detection

task. Comparing in native language (4db) is a deﬁnite win.

Relevance models substantially outperforms the tf·idf baseline.

Rule of Interpretation (ROI[2]) has a marginal eﬀect. Boldface

marksminimum costineachcolumn. .......................172

Introduction

Information Retrieval (IR) is a ﬁeld of science concerned with searching for

useful information in large, loosely structured or unstructured collections. In

the 1970s, when the ﬁeld was in its infancy, the word information primarily

meant text, speciﬁcally the kinds of text one might ﬁnd in a library. Today

information exists in many more forms, often quite diﬀerent from books or

scientiﬁc articles which were the focus of early retrieval systems. Web search-

ing is perhaps the most famed application of information retrieval today, and

on the web we may ﬁnd relevant information in the form of text, but also as

images, as audio ﬁles, as video segments, even as hyper-links between various

pages. A modern search system must have the capability to ﬁnd, organize and

present to the user all of these very diﬀerent manifestations of information,

because all of them may have some relevance to the user’s information need.

The notion of relevance plays an extremely important part in Information

Retrieval, some went so far as to say that it deﬁnes the entire ﬁeld and serves

as a distinguishing feature from database theory and library science [120]. The

concept of relevance, while seemingly intuitive, is nevertheless hard to deﬁne,

and even harder to model in a formal fashion, as evidenced by over 160 publi-

cations attempting to deal with the issue. Relevance is also the subject of this

book. We will not attempt to bring forth a new deﬁnition of relevance, nor

will we provide arguments as to why one deﬁnition might be theoretically su-

perior or more complete than another. Instead, we will take a widely accepted,

albeit somewhat conservative deﬁnition, make several assumptions, and from

them develop a new probabilistic model that explicitly captures that notion of

relevance. In our development we will make a determined eﬀort to avoid the

heuristic steps that accompanied the development of popular retrieval models.

To this end, a substantial portion of our narrative will be devoted to a formal

discussion of event spaces and assumptions underlying our model.

2 1 Introduction

1.1 Contributions

There are two major contributions in this work. The ﬁrst contribution is a

new way to look at topical relevance within the ﬁeld of Information Retrieval.

The second contribution is a new generative model for collections of discrete

data.

1.1.1 A new model of relevance

Our ﬁrst contribution will be of primary interest to a researcher in Information

Retrieval, or to anyone concerned with modeling topical content. This book

will propose a new way to look at topical relevance. The new approach will

be summarized as a hypothesis (deﬁned in section 3.1), and will lead to a new

formal model of Information Retrieval. Some of the main advantages of the

model are:

• It represents a natural complement of the two dominant models of retrieval:

the classical probabilistic model and the language modeling approach.

• The model explicitly combines documents, queries and relevance in a single

formalism. This has been somewhat of a challenge in previous formulations.

• It requires no heuristics to compute the probabilities associated with the

relevant class.

• The retrieval model has relevance feedback and query expansion as integral

parts.

• It captures a wide array of IR tasks. As far as we know, this is the

ﬁrst model capable of performing ad-hoc search, image and video anno-

tation, handwriting retrieval, semi-structured search and Topic Detection

all within the same formalism.

• It outperforms state-of-the-art benchmarks. A few examples: 10-25% im-

provement in ad-hoc search, 5-30% in topic detection, 15-60% in image

annotation and retrieval.

• It allows queries to be issued in any form: keywords, passages, whole doc-

uments, sketches, portions of images, video frames – as long as we have

enough data to estimate the parameters.

• It can recover missing portions of incomplete or degraded information

representations and provides a particularly eﬀective method of search-

ing semi-structured databases which achieves a 30% improvement over

a language-modeling baseline.

1.1.2 A new generative model

The second contribution of our work may be useful to a language-modeling

practitioner, or to a reader with a more general interest in statistical modeling

(not necessarily of language). In chapter 4 we will propose a new method

for modeling exchangeable sequences of discrete random variables. The main

distinguishing qualities of our formalism are:

1.2 Overview 3

• It can successfully handle rare events and does not ignore outliers in the

training data.

• It does not make any structural assumptions, allowing the data to speak

for itself.

• The model is easy to train, either generatively, or with a discriminative

objective.

• The model is eﬀective: 15-20% reduction in predictive perplexity over the

baselines, plus successful applications to retrieval tasks mentioned in the

previous section.

1.1.3 Minor contributions

In addition to the two major contributions, this book contains a number

of smaller results. These results mostly represent new insights into existing

models, or serve as explanations for curious experimental observations. Some

of these minor contributions are:

• We will show that, contrary to popular belief, the classical probabilistic

model need not be predicated on the assumption of word independence,

or even on the weaker assumption of linked dependence suggested by [29].

Instead, it requires a much more plausible assumption, which we call pro-

portional interdependence.

• We will try to explain why explicit models of word dependence have never

resulted in consistent improvements in retrieval performance, either in the

classical or the language-modeling framework.

• We will assert that documents and queries can be viewed as samples from

the same underlying distribution, even though they may look very diﬀer-

ent.

• We will explain why relative entropy (KL divergence) has empirically out-

performed probability ratio as a criterion for ranking documents in the

language-modeling framework.

• We will conclude that a Probabilistic LSI model with k aspects is equiva-

lent to a regular, document-level mixture model with N>>kcomponents.

• We will argue that Latent Dirichlet Allocation is not a word-level mixture,

it is a regular document-level mixture restricted to the k-topic sub-simplex.

• We will demonstrate that a form of KL-divergence ranking may be inter-

preted as a statistical signiﬁcance test for a hypothesis that a document

and a query originated from the same relevant population.

1.2 Overview

This book will be structured in the following manner. Chapter 2 will be

devoted to the notion of relevance and to previously proposed models of rel-

evance. We will start with a simple deﬁnition and then brieﬂy outline several

剩余210页未读，继续阅读

wenbranger

粉丝: 35
资源: 11

信息检索：相关性的生成理论

A generative vision model

AT-GAN:A Generative Attack Model译文

Structured Relation Discovery using Generative Models

a paper named Generative Time Series Forecasting with Diffusion, Denoise, and Disentanglement

Generative Pre-trained Transformer

Generative Adversarial zero-shot

rewriting a deep generative model

deep learning based generative adversarial network history of development

怎么用WEKA软件做Generative Topographic Mapping

手动编写一个generative model用来二分类

GANprintR: Improved Fakes and Evaluation of the State of the Art in Face Manipulation Detection

a style-based generator architecture for generative adversarial networks主要研究了什么

Improving Language Understanding by Generative Pre-Training

Generative Pre-Training

generative ai都包括哪些

Generative Pre-Training是什么

guided diffusion

generative adversarial network history

score-based generative models代码

最新资源