Hacker News Reader

Comment by orthoganol | Hacker News Reader

> but RL is also widely used in Robotics, Image Processing and Natural Language Processing

RL for NLP? I would love to know about counter examples, but I'm not aware of a serious project using RL for NLP, let alone 'widely used.' However I do believe RL makes sense for a number of NLP problems.

Either way, well done. I appreciate a collection of the algos from Sutton's book (great book), and in Python.

View on HN

A lot of recent research uses RL to "fine-tune" NLP models. A practical example would be Google's recently announced Machine Translation System (https://arxiv.org/abs/1609.08144). It uses RL to directly optimize BLEU scores on translated sentences.

You'll find similar applications in state-of-the art models for chatbots for example. Though I agree, "widely used" may be somewhat of an overstatement. But it's becoming more common.

On a side note, I actually think RL makes a lot of sense for many NLP problems and it would be super interesting to build a pure RL approach to language modeling or translation. Nobody has managed to do that quite yet.

I agree, I would love to see or even work on pure RL approaches in NLP. I would also like to see HMMs explored more, which I think also make a lot of sense for NLP, if you think of the problem as sequences of hidden states represented by semantic frames or other hand crafted language features producing phrases and sentences.

I'm doing that. Database queries are (relatively) expensive actions in what I'm doing, and reinforcement learning sorta cuts down on the amount of db querying that my NLP program does. Way too much time is spent finetuning it, and I find myself often lost in the model

Reply on news.ycombinator.com