In this paper, we propose an automatic method for detecting discourse structure using a variety of clues existing in the surface information of sentences. We have considered three types of clue information: clue expressions, occurrence of identical/synonymous words/phrases, and similarity between two sentences. Experimental results have shown that, in the case of scientific and technical texts, considerable part of the discourse structure can be estimated by incorporating the three types of clue information, without performing sentence understanding processes which requires giving knowledge to computers.