Neural response generation for task completion using conversational knowledge graph

Effective dialogue generation for task completion is challenging to build. The task requires the response generation system to generate the responses consistent with intent and slot values, have diversity in response and be able to handle multiple domains. The response also needs to be context relevant with respect to the previous utterances in the conversation. In this paper, we build six different models containing Bi-directional Long Short Term Memory (Bi-LSTM) and Bidirectional Encoder Representations from Transformers (BERT) based encoders. To effectively generate the correct slot values, we implement a copy mechanism at the decoder side. To capture the conversation context and the current state of the conversation we introduce a simple heuristic to build a conversational knowledge graph. Using this novel algorithm we are able to capture important aspects in a conversation. This conversational knowledge-graph is then used by our response generation model to generate more relevant and consistent responses. Using this knowledge-graph we do not need the entire utterance history, rather only the last utterance to capture the conversational context. We conduct experiments showing the effectiveness of the knowledge-graph in capturing the context and generating good response. We compare these results against hierarchical-encoder-decoder models and show that the use of triples from the conversational knowledge-graph is an effective method to capture context and the user requirement. Using this knowledge-graph we show an average performance gain of 0.75 BLEU score across different models. Similar results also hold true across different manual evaluation metrics.

• We have expanded the abstract as suggested Response to the Comments of Reviewer 1 • Query 1: There are several data used in this paper, but not provided the suitable citations for them. Recommended to provide the suitable citations Response 1: We appreciate the suggestion by the reviewer. We have used only one dataset consisting of several domains in our research. This paper proposing this dataset is cited in Section 4.1.
• Query 2: . Provide the Bi-LSTM Encoder Decoder with illustrative example for better understanding the model Response 2: As suggested by the reviewer we have provided the illustrative example for the model in form of Figure 1,2,3 and 4 for this purpose. The detailed description is also provided in the 3.2 to 3.8.
• Query 3: . Discuss the time complexity for the Algorithm 1 discussed in this paper Response 3: As suggested by the reviewer we have provided this discussion in Section 3.9.
• Query 4: The discussion under Section 3.6 to 3.8 are very limited. Recommended to provide the details discussion and comparative analysis.
Response 4: As advised we have provided the details and comparative analysis in the corresponding sections.
• Query 5: . The results of the paper are very limited. Recommended to provide the different metrics for each dataset. It is recommended to provide the suitable reasons for the superior performance of the proposed models.
Response 5: We have added another automatic evaluation metric to measure our results. Unlike the previous metrics where the word-overlap is measured, this metric measures the semantic similarity between the desired and produced output. We have provided the detailed explanation and analysis for the same in Sections 5.1 and 5.2.
• Query 6: . Discuss the limitations of the proposed model and possible future extensions on it.
Response 6: As suggested we have added these discussions in Section 5.2 and 6.
• Query 7: .Summarize the related work using a table with the benefits and limitations of the existing works. It is also recommended to discuss which among the existing limitations are addressed in this paper.
Response 7: As suggested, we have added Table 1 for this purpose.

Response to the Comments of Reviewer 2
• Query 1: Abstract-Highlight the novelty aspect after Aim/Objective of the paper. Add in the last lines in what %age and in what parameters the proposed methodology is better and as compared to existing techniques and what is the overall analysis of the proposed technique.
Response 1: As suggested, we have expanded our abstract to reflect the points.
• Query 2: . Introduction needs to be more broad with regard to the Background, SCOPE, Problem Definition and even other related highlights. Add Objectives of the paper in Points. Add Organization of the paper.
Response 2: We have added the suggested details in the introduction section (Section 1).
• Query 3: Related works needs to be more and min 15-25 papers should be there in Related works. ADD in the last lines what overall technical gaps are observed that led to the design of the proposed methodology.
Response 3: As suggested we have expanded the related works. We have also added Table 1 highlighting the benefits and limitations of previous works in comparison to our work.
• Query 4: Add the flowchart of the proposed methodology. Add System Model and Step by Step operation of working of proposed technique.
Response 4: As advised we have added 1,2,3,4 and 6 for this purpose. The working of each of the modules are discussed in Section 3.
• Query 5: Add more information towards results. And perform Performance analysis with some existing techniques.
Response 5: We have added another automatic evaluation metric to measure our