BeamR: Beam Reweighing with Attribute Discriminators for Controllable Text Generation.

Recent advances in natural language process- ing have led to the availability of large pre- 002 trained language models (LMs), with rich gen- 003 erative capabilities. Although these models 004 are able to produce fluent and coherent text, 005 it remains a challenge to control various at- 006 tributes of the generation, including sentiment, 007 formality, topic and many others. We propose a 008 Beam Reweighing (B EAM R) method, building 009 on top of standard beam search, in order to con- 010 trol different attributes. B EAM R combines any 011 generative LM with any attribute discrimina- 012 tor, offering full flexibility of generation style 013 and attribute, while the beam search backbone 014 maintains fluency across different domains. No- 015 tably, B EAM R allows practitioners to leverage 016 pre-trained models without the need to train 017 generative LMs together with discriminators. 018 We evaluate B EAM R in two diverse tasks: senti- 019 ment steering, and machine translation formal- 020 ity. Our results show that B EAM R performs 021 on par with or better than existing state-of-the- 022 art approaches (including fine-tuned methods), 023 and highlight the flexiblity of B EAM R in both 024 causal and seq2seq language modeling tasks.