Therefore, all following analyses focus on gender errors.
Clearly, across sessions, participants mostly made gender errors, whereas other errors occurred at a lesser rate.
Figure 3 depicts the percentage of gender errors out of all responses for all eight groups in each task.
Likewise, a significant interaction between grammaticality and emotionality in the participants' performance revealed that the presence of emotional adjectives did facilitate the detection of gender errors.
In order to assess the effects of session, training method, and training mode on the number of gender errors, we coded the errors as a binary variable (with all gender errors scoring 1 and all correct responses and all other errors scoring 0) and assessed the effects of training method and training mode for each session (see Fig. 3 and Table 8).
Gender errors were dominant during all sessions and did not vanish towards the later sessions, indicating that, as we had anticipated, choosing the correct gender-like category in the production tasks constituted the most prominent problem for the learners.
As of session 3, there was an effect of training method, with participants trained with the random training method producing significantly more gender errors than those trained with the blocked training method.
The modes Melody and Rhyme&Melody were associated with considerably fewer gender errors than the Prose mode, and this effect intensified over sessions (Session:Melody, χ 2 (1) = 5.29, p < 0.05, Session:Rhyme&Melody, χ 2 (1) = 24.39, p < 0.001).
In order to assess the interaction of time, training mode, and training method, we carried out an overall analysis of the number of gender errors observed from Sessions 2 to 4. For reasons to do with the way the data were coded, we were unable to include items as a random variable.
With respect to the effects of training mode, we observed a significant advantage of mode Rhyme&Melody as of Session 3. From the start, this effect interacted with the training method, with the number of gender errors decreasing most clearly in the blocked Rhyme&Melody group (see Fig. 3).
It is also noteworthy that the analysis of gender errors showed that the emerging benefit of Rhyme&Melody over Prose was temporally correlated with the emerging benefit of blocking and the interaction of blocking and Rhyme&Melody, so clearly it was the combination of input optimization techniques that did the trick.
