Ah right now ranking is really rudimentary: I just use TextBlob's SA library to get a sentiment score for the comment, and combine it with the comment length. In theory we can also use BERT to get a sentiment score, which should be more accurate I guess? Love to hear your thoughts.
Simple is good! Especially for ranking as the objective is hard to define. Having looked at some samples on your site, imo it's good enough :)
Though if you wanted to try other things just for fun, maybe:
- Count matching NER, comments with a lot of book recommendations tend not to detail why they like the specific one currently filtered
- maybe down weight comments that are too short (after the matched NER is subtracted) as they seem to just have a title+link?
Not ML but while I'm here, on Android the comment section appear fine but has an horizontal scroll that seems spurious, with lots of blank space to the right (Galaxy S10, chrome & firefox)
Replies to your pinned top comment seems disabled, so allow me to ask here (sorry for the hijack):
> used a custom NLP-based ranking function to sort the comments
Can you expand on this function? (I'm familiar with NLP and most SOTA models)