In this study, we examine the effect of background knowledge and local cohesion on learning from texts. The study is based on construction-integration model. Participants were 176 undergraduate students who read a Computer Science text. Half of the participants read a text of maximum local cohesion and the other a text of minimum local cohesion. Afterwards, they answered open-ended and multiple-choice versions of text-based, bridging-inference and elaborative-inference questions. The results showed that students with high background knowledge, reading the low-cohesion text, performed better in bridging-inference and in elaborative-inference questions, than those who read the high-cohesion text. Students with low background knowledge, reading the high-cohesion text, performed better in all types of questions than students reading the low-cohesion text only in elaborative-inference questions. The performance with open-ended and multiple-choice questions was similar, indicating that this type of question is more difficult to answer, regardless of the question format.