Table 1.

Comparison among the works presented in Section 2. Each approach has been evaluated according to scalability, multi-modality and flexibility.

Approach	Ref.	Scalability and Re-Usability	Multi-Modality	Flexibility
State-based	[16]	Combine and parametrise the interaction patterns adequately. This patterns can be reused among tasks	Yes	Interaction adapted through parametrisation of the patterns
Frame-based	[10,17]	Describe the new task with the description language of choice. The slot-filling techniques are reused among tasks	No ([17]), Yes ([10])	Predefined responses
	[18]	Build new domain ontologies, reuse the ontology-based rules for dialogue management	No	Responses are defined in the ontologies
Plan-based	[14]	Design the domain-dependent aspects, using dialogue agents, while reusing the domain-independent aspects	No	Responses are implemented in the dialogue agents
Probabilistic-based	[9]	Build and parametrise new probability rules. Some rules can be reused for different domains	Yes	Responses can be adapted through the probabilistic rules
	[19]	Build new domain ontologies. Generic aspects of the dialogue are reusable	No	System responses are generated through NLG
	[20]	Increase the database. Statement response module can be reused among domains	No	Predefined responses
Agent-based	[21]	New datasets and task agents would be required. This approach require smaller datasets than usual. The example manager module is common to all tasks	No	Uses template-based NLG
End-to-End	[22,23,24,25,26,27,28,29]	Develop new datasets and agents (if applied). The model is common to all tasks. Some agents may be generic	No	Use NLG for generating the system response (template-based or model-based)