This repository contains source code used for visual dialogue analysis system in the DT2VIS paper. Note that the code was tested with Python 3. Please python 3 as your test environment.
DT2VIS takes a natural language query about a given dataset as input and outputs a focus+context answer, and conducts users to explore. The workflow of the system is as follow.
The query parsing engine parse the relevant information (e.g., attributes, filter, and task) from the input query during query collection through an input interface. Next, the system extracts specific data segments from the data space based on the attributes and filtered information, and answer generation templates corresponding to the task will also be collected. These data segments and templates will be transported to the answer generation engine to generate answers. The system utilizes a focus+context approach in the answer generation engine to generate text and visualization answers. Simultaneously, the task and attribute information parsed by the query parsing engine will also be sent to the query recommendation engine. The system first updates the transition matrix based on user feedback on the previously recommended query and then computes the recommended query from the new transition matrix. Finally, the answer and recommended queries will be replied to the user together. Users may have a new query intention and start a new round of query tasks.
The DT2VIS code has a few dependencies that can be installed using the requirement.txt file, In addition, the system also needs stanford-parser.jar and stanford-parser-models.jar.
The current version is a development version. If you have any questions, feel free to open an issue or contact Guodao Sun.