There are many ways to quantify a qualitative variable. One way is to use five point scale, i.e. very good, good, average, bad, very bad.

You cannot quantify qualitative data. You can, however, do the reverse.

Assign each observation a numerical score for the qualitative variable using the scale of your choice. Use logistic regression rather than OLS because the scale you pick will influence the latter but not the former.

What Prof Deng is suggesting is called "integer encoding". This is okay for some ordered categories, but you have to handle a bias with encoding. Maybe it makes more sense to say for example, low quality is 1, medium quality is 3, high quality is 10? Or would it be better if it was -1, 0, 1? Check what gives you better predictions. Another way to go is "one-hot encoding", where each category label is a variable, and the value is boolean (e.g., low quality: 1, medium quality: 0, high quality: 0, for the same building). Both methods assume categorical data. With natural language data, you have to do more tricky encoding yourself to get category labels (or have an NLP model do it for you). Before you do anything, I recommend you get familiar with encoding. Machine learning blogs can offer you a primer if you search for "encoding categorical data" or "encoding qualitative data".

This could be very serious. Suppose you start with three groups of data. You code them 0, 1, and 2. Then somebody runs a linear regression and discovers the following: the 0 group is essentially unrelated to the dependent variable, the 1 group is doing about half of what the 2 group is doing. But now somebody acts on this information. If in fact the 1 group is actually unrelated, you have committed a fraud. You knew, or should have known, that this outcome was a possibility. What you should have done was create dummy variables. You cannot make quantitative variables out of qualitative ones.

Technically, dummy variables are dichotomous, quantitative variables.

Reacting to Prof Colwell, this is exactly my problem with integer encoding. There are way too many points of bias you introduce to your scoring system. The only thing that is objective in this encoding is the order of integers, because your data is ordinal (medium is above low, high is above medium). But the choosing the numbers themselves, their spacing, any normalization, any aggregation of dimensions will be sources of bias. For your specific use-case, I recommend you use hedonic regression, since your qualitative variables are constituent characteristics of an overall qualitative concept. For that, you will need to use one-hot-encoding, which is the same as dummy variables.