Doctoral Thesis: Controlling Neural Language Generation

Tianxiao Shen

Abstract: 

Language models have achieved success in a wide range of applications such as machine translation, dialogue generation, and writing autocompletion. However, typical models operate in a left-to-right, unconstrained fashion with limited control over what is generated. In this talk, I’ll discuss three directions to control language generation. First, we’ll leverage distributional alignment to perform style transfer from non-parallel text. Next, we’ll study confounding factors and use invariance to isolate them to transfer text in the desired direction. Lastly, we’ll look at flexible generation location control, for which we developed sequence models that dynamically create and fill in blanks during the generation process.

Details

• Date: Tuesday, May 3
• Time: 1:30 pm - 3:00 pm
• Location: 32-D463 (Star Room)

Additional Location Details:

Thesis Supervisors: Profs. Regina Barzilay and Tommi Jaakola

Zoom link:  https://mit.zoom.us/j/97034696195