A comparative study of TFIDF LSI and multiwords

A comparative study of TF*IDF, LSI and multi-words for text classification Presenter : JIAN-REN CHEN Authors : Wen Zhang, Taketoshi Yoshida, Xijin Tang 2011. ESWA Intelligent Database Systems Lab

Outlines n Motivation n Objectives n Methodology n Experiments n Conclusions n Comments Intelligent Database Systems Lab

Motivation Although TF*IDF, LSI and multi-word have been proposed for a long time, there is no comparative study on these indexing methods, and no results are reported concerning their classification performances. Intelligent Database Systems Lab

Objectives • A comparative study of TF*IDF, LSI and multi-words for text classification. - information retrieval - text categorization • indexing term: ① semantic quality ② statistical quality Intelligent Database Systems Lab

Methodology - TF*IDF 1) 2) 3) 4) wi, j : the weight for term i in document j N : the number of documents in the collection tfi, j : is the term frequency of term i in document j dfi : is the document frequency of term i in the collection Terms (keywords) of the document collection documents Intelligent Database Systems Lab

Methodology - LSI Given a term-document matrix X = [x 1 , x 2 , . . . , xn ] є Rm and suppose the rank of X is r, LSI decomposes the X using SVD as follows: 1. 2. Xk=Uk’Σk. Vk. T ’ Terms (keywords) of the document collection documents Intelligent Database Systems Lab

Methodology - Multi-word its occurrence frequency should be at least twice in a document. the length of the multi-word should be between 2 and 6 Intelligent Database Systems Lab

Experiments - Datasets Chinese corpus： Tan. Corp. V 1. 0 14150 documents 20 categories Select 1200 documents agriculture 5, 468, 301 individual words 219, 115 sentences history politics economy English corpus： Reuters-22173 distribution 1. 0 22173 documents 135 categories Select 2032 documents Crude (520) 50, 837 sentences agriculture (574) Trade (514) 281, 111 individual words Interest (424) Intelligent Database Systems Lab

Experiments - Evaluation Intelligent Database Systems Lab

Experiments - Chinese Intelligent Database Systems Lab

Experiments - English Intelligent Database Systems Lab

Experiments – t-test Intelligent Database Systems Lab

Comparison information retrieval TF*IDF Chinese LSI English multi-word text categorization computation complexity O(n m) best O(n 2 r 3) O(ms 2) Intelligent Database Systems Lab

Conclusions • LSI can produce better indexing in discriminative power. • LSI and multi-word have better semantic quality than TF*IDF, and TF*IDF has better statistical quality than the other two methods. • The number of dimension is still a decisive factor for indexing when we use different indexing methods for classification. Intelligent Database Systems Lab

Comments • Advantages - Compare with TF*IDF, LSI and multi-words • Disadvantage - semantic quality and statistical quality are considered merely by our intuition instead of theory • Applications - text mining Intelligent Database Systems Lab