Abstract
1.Previous approaches often restrict to specific domains and require handcrafted rules.
2.Use the Sequence to Sequence frame work.
3.The strength of the model that it can be trained end-to-end and thus requires much fewer hand-crafted rules.
4. The model’s weakness is the lack of consistency is a common failure mode of our model.
Introduction
- Neural network can be used to map complicated structures to other complicated structures.
- Mapping a sequence to another sequence which has direct applications in NLU. (Sutskever et al., 2014)
The advantage of mapping a seq to another seq
- requires little feature engineering and domain specificity, meanwhile matching the best answer of the input.
- Allow researchers to work on tasks for domain knowledge, and too hard to design rules.
- This approach can do surprisingly well on generating fluent and accurate replies to conversations.
The disadvantage of seq to seq
- Due to the complexity of this mapping, conversational modeling has been designed to be very narrow in domain.
Datasets
- IT helpdesk datasets of conversations
- A noisy datasets of movie subtitles
Experiments
- can hold a natural conversation and sometimes perform simple forms of common sense reasoning.
- the recurrent nets obtain better performance compared to the n-gram model
- capture important long-range correlations.
Related Work
- seq2seq的应用
(1)neural machine translation and achieves improvements on the English-French and English-German translation tasks from WMT’14 dataset (Luong et al.,2014; Jean et al., 2014)
(2)parsing(Vinyals et al., 2014a)
(3)image captioning (Vinyals et al., 2014b)
Difference from other conventional systems
- lack domain knowledge