Statistical Identification of Key Phrases for Text Classification

Abstract

Abstract.Algorithms for text classification generally involve two stages, the first of which aims to identify textual elements (words and/or phrases) that may be relevant to the classification process. This stage often involves an analysis of the text that is both language-specific and possibly domain-specific, and may also be computationally costly. In this paper we examine a number of alternative keyword-generation methods and phrase-construction strategies that identify key words and phrases by simple, language-independent statistical properties. We present results that demonstrate that these methods can produce good classification accuracy, with the best results being obtained using a phrase-based approach.

A Hybrid Statistical Data Pre-processing Approach for Language-Independent Text Classification

“…Accuracy figures, describing the proportion of correctly classified "unseen" documents, were obtained using the Ten-fold Cross Validation (TCV). A support threshold value of 0.1% and a Lower Noise Threshold (LNT) value of 0.2% were used, as suggested in [6]. A confidence threshold value of 50% was used (as proposed in the published evaluations of a number of associative classification studies [5,15,28]).…”

“…the first k words for each predefined class) that are selected from the ordered list of potential significant words (in a descending manner based on their contribution value) are defined to be significant words. In [6] the authors (based on the above definitions) propose a statistical "bag of phrases" (DR) approach for TC, namely DelSNcontGO: phrases are Delimited by stop marks (S) and/or noise words (N), and (as phrase contents) made up of sequences of one or more significant words (G) and ordinary words (O); sequences of ordinary words delimited by stop marks and/or noise words that do not include at least one significant word (in the contents) are ignored. The experimental results presented in [6] show that DelSNcontGO performs well with respect to the accuracy of classification.…”

“…In [6] the authors (based on the above definitions) propose a statistical "bag of phrases" (DR) approach for TC, namely DelSNcontGO: phrases are Delimited by stop marks (S) and/or noise words (N), and (as phrase contents) made up of sequences of one or more significant words (G) and ordinary words (O); sequences of ordinary words delimited by stop marks and/or noise words that do not include at least one significant word (in the contents) are ignored. The experimental results presented in [6] show that DelSNcontGO performs well with respect to the accuracy of classification. In this paper, this statistical "bag of phrases" DR approach will be further concerned in the section of experimental results.…”

“…; (3) W GLO " read ! to create a global word set, where the word document-base support supp GLO is associated with each word u h in W GLO ; (4) for each C i ∈ C do (5) W LOC " read documents that reference C i to create a local word set, where the local support supp LOC is associated with each word u h in W LOC ; (6) for each word u h ∈ W LOC do (7) contribution " (u h .supp LOC / u h .supp GLO …”

“…[6,29]). In [3] Antonie and Zaïane argue: an associative text classifier "is fast during both training and categorization phases", especially when handling large document-bases; and such classifiers "can be read, understood and modified by humans".…”