The Semantic Structure of the FLOB Corpus : Using Probabilistic Topic Models for Corpus Description

FLOB　コーパス　ノ　イミ　コウゾウ　カクリツロンテキ　トピック　モデル　ニ　ヨル　ゲンゴ　シヨウイキ　ノ　トクチョウ　ヅケ

タバタ, トモジ

Tabata, Tomoji

Graduate School of Language and Culture, Osaka University

オオサカダイガク　ダイガクイン　ゲンゴブンカ　ケンキュウカ

Topic modelling is a machine learning method for uncovering hidden semantic structures in a corpus of texts. Based on a probabilistic inference algorithm called latent Di richlet allocation, the technique makes it possible to identify sets of frequently co-occurring words, or topics, that characterize a text as well as classify texts into groups defined by inferred sets of strongly associated topics. One of the major advantages topic modellin g has over traditional key-word detection techniques employed in many stylometric or corpus linguistic studies is that topic models do not simply help to classify texts, but enable us to visualize complex yet meaningful interrelationships between vocabulary items, topics, and more importantly, association between topics and texts in the form of network diagrams and heatmaps. The present study applies topic modelling to the FLOB corpus with a view to analyzing latent semantic structures underlying in the corpus and mapping its subcorpora (or,registers) in the network of words, topics, and texts. What is of special interest is that by means of this approach it is now possible to shed new lighto n thematic structures composed by a large number of infrequent words, which would otherwise escape the net of key-word statistics due to infrequency of occurrence. Emerging results from this research are expected to open up a new avenue of inquiry into key semantic patterns in a large collection of texts, thereby demonstrating a possibility of building a bridge between findings from machine learning text mining and stylistics, distant reading and close reading,with an empirical interplay of insights that will benefit interdisciplinary stylistics.