I'm currently thinking of the system that would compare 2 texts (source, target) and suggest changes to the source text to match more the style of the target. It should not convert it to the target but more suggest to rephrase some sentences, changes words, add some key words and so on. Any ideas how this could be done? Thanks a lot for any suggestions!
$\begingroup$
$\endgroup$
1
-
2$\begingroup$ There is not much work in linguistic style transfer but I did find this: Stylistic Transfer in Natural Language Generation Systems Using Recurrent Neural Networks (pdf) $\endgroup$– EmreCommented May 13, 2017 at 8:06
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
0
Some time ago, I stumbled upon the Author Obfuscation task of PAN16 workshop. The task at hand is given a document, paraphrase it so that its writing style does not match that of its original author.
Although, it's not exactly what you want, you can take a look at their proceedings and find useful tips regarding authors style and how to "change" it (i.e. in your case, an "author" is a collection of documents from the same source).
For example, one approach denotes specific metrics that define an author's style (average sentence word count, use of stopwords, pos-type frequencies). You could use the target's scores in these metrics as your gold standard and try to modify the source text, in ways discussed in the proceedings of the workshop, so as to mask your source text to match the target one (in terms of these metrics).
