Table 1.
Benefits and limitations of some previous works in task-oriented response generation.
We also compare with our method.
Fig 1.
Illustrative example showing the working of the Bi-LSTM encoder-decoder model.
Fig 2.
Illustration showing the working of Bi-LSTM based encoder-decoder model with copy mechanism.
Fig 3.
Illustration showing the working of the BERT-LSTM model.
Fig 4.
Illustration of the working of the BERT-LSTM model with copy-mechanism.
Fig 5.
Illustration of the KG construction algorithm at work.
Fig 6.
Flowchart showing the working of the response generation system.
First the input utterance is used to construct or update the knowledge graph. In the second step this knowledge-graph along with the current utterance is fed to our deep-learning model. Finally the model works on the input and generates an appropriate context relevant response.
Table 2.
Detailed dataset description of each domain and the combined dataset.
Table 3.
Automatic evaluation results of different experiments in terms of BLEU, Perplexity (PPL), ROUGE-1, ROUGE-2 and ROUGE-L (rouge F-measures values are mentioned here).
Table 4.
Human evaluation results of different experiments in terms of Fluency (Gra), Adequacy (Con), Slot Consistency (SC) and Intent Relevance (IR).
Table 5.
Example outputs from different systems showing the effect of copy mechanism and KG triples on response generation.
Table 6.
Few examples of errors produced by different systems.