상세정보

In this thesis, we investigate the construction of open domain knowledge base which
consists of thousands of domains and its relevant verbs. We define a domain to be an
atomic unit representing the semantics of texts, such as topics, themes, etc. The relevant
verbs are a set of verbs semantically related to each domain. In particular, they
are essential to clarifying domains according to its context. To this end, we firstly build
a taxonomy of domains from a domain knowledge, i.e., Open Directory Project (ODP).
We then identify domain-related verbs from different information sources. From the experimental
results, we confirm that the proposed approach significantly outperforms an
existing verb identification approach.
Based on the ODP-based open domain knowledge base, we present a couple of knowledgeempowered
applications to verify its effectiveness. We describe knowledge-empowered
contextual advertising as the first application. In the application, we infer domains of
webpages (i.e., what a user wants) and models verbs associated with the inferred domains
in the action perspective (i.e., what a user wants do). This way enables us to model
the intent of a user visiting a webpage as pairs of the domain and its associated verbs.
Subsequently, we incorporate such two features into and ads ranking framework.
In addition to contextual advertising, we utilize the ODP-based open domain knowledge
base in a traditional natural language processing task, i.e., keyphrase extraction from
short texts. In this application, we extract keyphrases relevant to domains of texts. In
particular, we identify domains of a short text and extract quality keyphrases from terms
that contribute considerably to the identified domains. We also extract representative
verb words as verb keyphrases which represent the domains of a short text in the action
perspective.
We show that our knowledge-empowered applications ourperform state-of-the-art techniques
on real-world datasets. In knowledge-empowered contextual advertising, the proposed
methodology delivers the right ads satisfying a user’s information needs. The
proposed keyphrase extraction methodology effectively extracts both quality keyphrases
and verb keyphrases. From in-depth analysis, we verify that our ODP-based open domain
knowledge base is indeed effective to text understanding, including contextual advertising
and keyphrase extraction.

Abstract
Contents i
List of Figures iv
List of Tables vi
1 Introduction 1
 1.1 Background and Motivation  2
 1.2 Problem Statement  4
  1.2.1 Building Taxonomy of Domains  5
  1.2.2 Identifying Domain-related Verbs  6
 1.3 Knowledge-empowered Applications  7
  1.3.1 Contextual Advertising  7
  1.3.2 Keyphrase Extraction  8
 1.4 Contributions of Thesis  9
  1.4.1 Organization of Thesis  10
2 Background 12
 2.1 Open Directory Project 13
 2.2 ODP-based Semantic Classification  14
 2.3 ODP-based Semantic Ranking  20
 2.4 ODP-based Intelligent Services on Smartphones  22
  2.4.1 Content Curation Service 22
  2.4.2 Personalized Intelligent Interface 27
  2.4.3 Conversational Photo Sharing Service  30
  2.4.4 Others  33
3 Constructing Open Domain Knowledge Base 37
 3.1 Building Taxonomy of Domains  38
 3.2 Preliminary 39
 3.3 Identifying Relevant Verbs to Domains  40
  3.3.1 Searching Relevant Documents  41
  3.3.2 Extracting Relevant Verbs  43
 3.4 Performance Evaluation  44
 3.5 Summary  48
4 Knowledge-empowered Contextual Advertising 50
 4.1 Introduction 50
 4.2 Methodology 53
  4.2.1 Verbal Intent Modeling 53
  4.2.2 Ads Ranking  54
 4.3 Contextual Advertising Engine  55
  4.3.1 Topic Classifier  56
  4.3.2 Verbal Feature Generator  57
  4.3.3 Ads Ranker 57
 4.4 Evaluation 57
  4.4.1 Settings and Dataset  57
  4.4.2 Evaluation of Ranking Ads  59
  4.4.3 Qualitative Analysis 63
 4.5 Related Works  65
  4.5.1 Contextual Advertising  65
  4.5.2 Search Task, User Intent Identification and Verb Representation  66
 4.6 Summary  67
5 Knowledge-empowered Keyphrase Extraction 68
 5.1 Introduction  69
 5.2 Methodology  71
  5.2.1 Identifying Domains  72
  5.2.2 Selecting Keywords  72
  5.2.3 Extracting Quality Keyphrases 74
  5.2.4 Extracting Verb Keyphrases  76
 5.3 Experimental Study  77
  5.3.1 Datasets and Settings  77
  5.3.2 Keyphrase Extraction Results  81
  5.3.3 Ablation Study 87
  5.3.4 Keyword Extraction Results  91
  5.3.5 Domain Identification Results  92
 5.4 Related Works  95
  5.4.1 Keyphrase Extraction  95
  5.4.2 Semantic Classification 96
 5.5 Summary  97
6 Conclusion 98
 6.1 Summary of Thesis  98
 6.2 Future Work 100
Bibliography 103