Blog 2
An end-to-end trainable neural task oriented dialogue system that can perform tasks in multiple domains paper
Pro:
- Able to understand how context and memory can be used to make an assistant like product.
- How to work with complex models in NLP.
- Integrate multiple models together that can be trained end-to-end.
- Can be made into a product by the end of class.
- Data is open source for a particular task (restaurants) data.
Cons:
- Hard to implement even with available code.
- Since the idea uses multiple models together, it is difficult to find all the pretrained submodels (word embeddings, paragram …).
- Difficult to evaluate each submodule of the overall model and their respective performances (how to train each one of them).
- Requires a convoluted human evaluation scheme to evaluate the model performance.
A context aware neural dialogue system that can have long conversations paper1 paper2
Pro:
- Relatively easier to implement than our first idea.
- Focuses on the same issue of using context and memory to make an dialogue system.
- Data is open-source.
Cons:
- Hard to evaluate each submodule of the overall model and their respective performances (how to train each one of them).
- Requires a convoluted human evaluation scheme to evaluate the model performance.
- Difficult to come up with various metrics to judge the performance of the model.
Pro:
- Data is readily available and performance evaluation is straightforward.
- Easy to establish various baselines by substituting in and out various encoders/decoders.
- Model could be interacted with making for an interesting presentation (real-time translations from English).
Cons:
- Implementing a transformer could be difficult.
- Not exactly an entirely new topic to explore.
- May be difficult to beat or come close to the current best machine translation models.
- Requires substantial understanding of the two languages involved with the translation.
Lecture:
Transformers and evaluation metrics like BLEU score.