Generally, hunting for themes depends on specific research questions (theory-driven) or on their emergence out of data (data-driven). According to the six phases put forward by Braun and Clarke (2006), and to address the instance you mentioned in your question, you could have three initial themes in the theme-searching stage (Phase 3: Searching for themes): positive talk, neutral talk, and negative talk. Later, you can review and refine those themes in the reviewing stage (Phase 4: Reviewing themes), wherein you should decide whether some of the initial themes might stay as they are, be merged or disregarded based on whether they are consistent, lining with data homogeneity, or inconsistent, stemming from data heterogeneity (diversity). In the last-but-one stage (Phase 5: Defining and naming themes), you could define and refine your themes based on an account consistency or a central organizing concept that should serve the purpose of your analysis. In short, consistency and logic will be the determining drive in the inclusion/exclusion decision making throughout the proposed phases. You could see Braun and Clarke (2019) and Terry et al. (2017) for updated insights.

Braun, V., & Clarke, V. (2006). Using thematic analysis in psychology. Qualitative Research in Psychology, 3(2), 77–101. https://doi.org/10.1191/1478088706qp063oa

Braun, V., & Clarke, V. (2019). Reflecting on reflexive thematic analysis. Qualitative Research in Sport, Exercise and Health, 11(4), 589–597. https://doi.org/10.1080/2159676X.2019.1628806

Terry, G., Hayfield, N., Clarke , V., & Braun, V. (2017). Thematic analysis. In C. Willig & W. S. Rogers (Eds.), The Sage handbook of qualitative research in psychology (2nd ed., pp. 17-37). SAGE Publications Ltd. https://dx.doi.org/10.4135/9781526405555

Best,

If these three responses all address the same basic content, then you might well have three sub-themes under a single more general theme.

Be careful with trying to 'quantify' your qualitative data. Even a theme that is only mentioned by a small number of respondents can be a valuable contribution to your data analysis - it may show a contrasting or additional view point which you lose if you simply use percentages to include or exclude themes. Some of the most interesting material in qualitative research emerges from 'maverick' comments which provide a new direction of thought for the researcher, something that perhaps not many people have thought of. Don't lose these voices by over relying on how many people have said something, try to gauge the importance instead.

It would be nice to try machine learning to do the job instead of you. Specifically, it would be really interesting truly to create a decision tree that describes all possible answers. Decision trees are splitting the original data according to some criteria. When gini index is used during splitting, it usually gives the best split.

The beauty of decision trees is that they are nicely interpretable unlike many other machine learning methods.

Some clustering methods can give interesting results too. Clustering methods put together similar answer into groups automatically.

It would be interesting to try several ML methods on the same data. This can bring new insights into your data and its interdependence.