Article Directory
1. NLG Status
- The introduction of hidden variables in discrete NLG process. The story might generate, task-based dialogue helpful.
- Left to right are no longer generated. For questions long text generated in parallelization generation, iterative optimization and top on down have generated new research.
- The training process is no longer just the maximum likelihood optimization function, there has been an objective function more granularity sentence.
2. NLG research: where are we now? Where to go after?
Until five years ago, NLP + DL is the exploration period; 2019 now seems a little progress, but the NLG remains the most prairie to be developed!
3. NLG more mature
- Early NLP + DL, mainly try to have a very mature NMT methods to move task on NLG
- And now, NLG have more new algorithms, and related algorithms have been out of the configuration of the NMT
- NLG areas (especially open NLG) seminars and competitions are increasing
- Above also contribute to better and more standard seminar
- The biggest stumbling block is the evaluation!
4. NLG engaged in work to make things lecturer learned 8
- The more open NLG task, the greater the difficulty, this time to introduce some restrictions.
- Do not think a whole generation optimization effect can be split dismantling, one by one break
- If you use the LM NLG, if we can improve the performance of LM (such perplexity
perplexity
may to a large extent on improving the overall generation effect. Of course, this is by no means the only method to generate a help to improve the quality. - Duokang Kang you generate what stuff!
- Even if the automatic evaluation effect is not good, you need to have, and even sets!
- If you really want to manually evaluate the criteria set too much detail as possible
- NLP + DL big question now is can not reproduce, so the feeling you produce results and the methods used made out
- This process is very frustrating, of course, very interesting.