THESIS
2021
1 online resource (xvi, 88 pages) : illustrations (some color)
Abstract
As machines are becoming more and more intelligent, the development of their emotional
quotient (EQ) is being left far behind. Although progress has been made in the affective computing
field, much of it is focused on emotion recognition, and little attention has been paid to
affective generation. In fact, being able to generate affective natural language would not only
make machines more human-like but would also create an engaging and unique experience for
users. Therefore, in this thesis, we focus on the affective generation of natural language.
Generating fluent and coherent text has been a challenging task for machines. Recent advances
in neural generative models have opened new possibilities to tackle this challenge. We
take advantage of these models and induce affective generatio...[
Read more ]
As machines are becoming more and more intelligent, the development of their emotional
quotient (EQ) is being left far behind. Although progress has been made in the affective computing
field, much of it is focused on emotion recognition, and little attention has been paid to
affective generation. In fact, being able to generate affective natural language would not only
make machines more human-like but would also create an engaging and unique experience for
users. Therefore, in this thesis, we focus on the affective generation of natural language.
Generating fluent and coherent text has been a challenging task for machines. Recent advances
in neural generative models have opened new possibilities to tackle this challenge. We
take advantage of these models and induce affective generation from three directions: (1) affect
reward shaping, (2) affect adaptive training, and (3) control with affect keywords and knowledge.
Firstly, we develop a novel affect reward of sentiment look-ahead which we apply to empathetic
chatbots. We implement and train two different sentiment look-ahead reward functions
to model how the user would feel towards a generated dialogue response, and use such to encourage
more empathetic responses. The empirical results and our analysis on both automatic
and human evaluations in terms of empathy, fluency, and relevance confirm that the sentiment
look-ahead reward is an effective way to generate more empathetic responses.
Secondly, we develop an affect adaptive training objective that dynamically mixes the maximum
likelihood estimation and reinforcement learning. We then apply this method to sensational
headline generation. We develop a sensationalism scorer to measure how sensational a
headline is. This scorer is then used to dynamically mix reinforcement learning with maximum likelihood training. The empirical results confirm that our model can learn a diversity of sensationalization
strategies, and affect adaptive training loss is able to generate more sensational
headlines.
Lastly, we generate affective text with external affect keywords and knowledge which we
apply to emotional story generation. We develop a controllable framework, Megatron-CNTRL,
that allows control through keywords as well as dynamic incorporation of external knowledge.
Our results demonstrate that Megatron-CNTRL outperforms existing state-of-the-art models in
terms of fluency, coherence, and consistency. Scaling it from 124M to 8B, the larger models
show consistent improvements for both quality and controllability. We then use affect keywords
to control Megatron-CNTRL and we are able to generate more emotional stories.
Post a Comment