Dialogue modeling using recurrent neural networks

Examensarbete för masterexamen

Please use this identifier to cite or link to this item: https://hdl.handle.net/20.500.12380/237861
Download file(s):
File Description SizeFormat 
237861.pdfFulltext804.04 kBAdobe PDFView/Open
Type: Examensarbete för masterexamen
Master Thesis
Title: Dialogue modeling using recurrent neural networks
Authors: Anderling, Viktor
Orrö, Jonathan
Abstract: As the importance of computers in everyday life increases, so does the demand for better human-computer interfaces. Natural language, being our most natural form of communication, combined with man’s innate tendency of anthropomorphism, motivates the idea of a talking machine. Existing dialogue systems have the problem of being unable to answer out-of-domain questions as well as being tedious to design. While these systems are developed with hand-crafted rules, the goal of this thesis is to investigate if a dialogue system could be automatically trained to speak instead. Our aim is to test whether a model trained on a dialogue corpus can compare to existing dialogue systems. We trained a recurrent neural network using the sequenceto- sequence method, preserving the state of the model during the course of the conversation. The resulting network is end-to-end trainable. User testing was used to evaluate the model and compare it to the other dialogue systems. The final model can answer appropriately to common phrases such as greetings and valedictions. It also generates replies in correct English. However, the results do not stretch any further than that. Giving the model a more complicated input usually results in a nonsensical reply, which prevents it from having a coherent conversation with the user. We present a few hypotheses as to why we did not get better results, with suggestions on how they could be solved. We display high hopes for future work in the area and present a few suggestions of what could be the next steps.
Keywords: Data- och informationsvetenskap;Computer and Information Science
Issue Date: 2016
Publisher: Chalmers tekniska högskola / Institutionen för data- och informationsteknik (Chalmers)
Chalmers University of Technology / Department of Computer Science and Engineering (Chalmers)
URI: https://hdl.handle.net/20.500.12380/237861
Collection:Examensarbeten för masterexamen // Master Theses



Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.