Which NLP techniques can be used to transform sentences (e.g. from passive to active voice) without affecting their meaning?

Question

I'm looking for NLP techniques to transform sentences without affecting their meaning.

For example, techniques that could transform active voice into passive voice, such as

The cat was chasing the mouse.

to

The mouse was being chased by the cat.

I can think of a number of heuristics one could implement to make this happen for specific cases, but would assume that there is existing research on this in the field of linguistics or NLP. My searches for "sentence transformation" and similar terms didn't bring up anything though, and I'm wondering if I simply have the wrong search terms.

Related to this, I'm also looking for measures of text consistency, e.g., an approach that could detect that most sentences in a corpus are written in active voice and detect outliers written in passive voice. I'm using active vs. passive voice as an example here and would be interested in more general approaches.

Answer 1

Strictly speaking, this is impossible. Changing the form of a sentence also changes its meaning. Even active-passive can be important, as you would use it to emphasise what is important: was it relevant what the cat was doing, or was it more relevant what happened to the mouse? True, the purely propositional meaning is not affected by this, but that is only one component of a sentence's meaning.

There has been a lot of work in traditional linguistics about sentence form. You could look at one of the seminal works, Syntactic Structures by Noam Chomsky, where he introduces the concept which later lead to transformational grammar. This influenced a lot of subsequent linguistic approaches, but as far as I am aware transformations are not really that much in the linguistic focus anymore.

For your second question, stylistic consistency, you could look at the work of Douglas Biber. His book Variation across speech and writing introduces a number of (easily extractable) linguistic features that you could use to quantify consistency.