The conventional analysis estimates both the locations and strengths of neural source activations from event-related magnetoencephalography data that are averaged across about a hundred trials. In the present report, we propose a new method based on a minimum modified-l 1-norm to obtain the dependence of strengths on the presented stimuli from a limited number of trial data. It estimates the strengths from 10-trial average data and the locations from 100-trial average data. The method can be applied to neural activations whose strengths, but not locations, depend on the presented stimuli. For instance, it can be used in experiments in which the activation in the anterior temporal area (aT) is measured by varying semantic relatedness between stimuli in linguistic experiments. We conducted a realistic simulation, following previous experiments on lexico-semantic processing, in which five neural sources were simultaneously activated. The results showed that when the signal-to-noise ratio was one for non-averaged data, the proposed method had standard deviations of 13 % for the normalized strengths in the aT. It is inferred on the basis of the general linear model in which the strength has a linear dependence on the stimulus parameters that the proposed method can detect the dependence at a significance level of 1 % if the peak-to-peak change in normalized strength is more than 49 %. It is smaller than 66 % for the conventional method, which estimated locations and strengths from 10-trial data for each point. Thus, the proposed method can plot an activation-strength versus stimulus-parameter curve with better sensitivity.