The self-paced moving-window reading paradigm was used in two experiments to examine the possibility of retaining ambiguity in comprehending Japanese temporarily ambiguous sentences, consisting of “subject-NP / VP (object-NP and transitive verb) / head-NP of relative-clause”. The biases for the main-clause or the relative-clause interpretations were manipulated by defining the strength of co-occurrence between the subject-NPs or head-NPs and the VPs, using the rating of typicality as an agent of the VP. In Experiment 1, the proper namess in Japanese were used for the weak biases. The rating values of co-occurrence for them were significantly lower than that for the strong biases, although there was neither semantic anomaly nor implausibility. In reading the ambiguous strings having weak biases for main-clause, it was predicted that readers retained both of the possible interpretations because of no trigger about whether the main-clause interpretation or the relative-clause interpretation was assumed, and consequently that the garden path effects at the head-NPs were reduced. The experiment did not, however, show any reductive effect, but the reading time data suggested that the use of the proper names could increased readers' processing load. In Experiment 2, the proper names were replaced to the common nouns which also had weak biases. The results showed that the clear garden path effects were obtained only in the condition with strong main-clause and weak relative-clause biases, while the effects were eliminated in the other conditions with weak main-clause biases or strong relative-clause biases. These results suggested that the interaction between lexical connectivity and processing load decides whether the structural ambiguity should be resolved or retained.