The present treatment gaps in the in-place mental health care network could be filled by a dialog-based clinical decision support system (CDSS). In this CDSS, patient statements are input. A transformer-based token classification network XLM-R (XLM-Roberta), is used to identify the desired features present in these statements. In this work, we have trained a model on several corpora. The training and testing sets for the first series of experiments were taken from the same dataset. The second series of experiments evaluates these models against non-native test sets; these non-native test sets belong to corpora other than the training datasets. We also trained a model on mixed samples. By mixed samples, we mean a set that contains sentences from corpora and sentences from corpus dedicated to mental health care. In the first series of experiments; the training and test sets were drawn from the same dataset, and it was found that the performance was remarkable in this particular situation. This happened because training and test set came from the same distributions. Further, the results of the second set of studies were very identical to those of the first. It occurred because, regardless of the sources, the distribution of features in online corpora is almost the same. We discovered that there are no dialect-specific samples in the web resources that are available. Additionally, samples in colloquial Hindi are absent from these corpora. Due to these reasons, model performance was not up to the mark. The model picked dialect-specific features when dialect-specific examples were added to the training set. Using only online resources, a CDSS can be created, but its use and practical application will be limited. For real-time deployment, CDSS created with corpus dedicated to metal health, will be more beneficial. A mixed-corpora approach is the best approach, because it provides rich vocabulary, versatile samples and also picks all the desired features.