The Future Remains Unsupervised
Keywords: Natural Language processing, Machine Learning
TL;DR: Why unstructured learning will fuel continued progress in natural language processing, with reinforcement learning as a complement
Abstract: Recent advances in natural language processing have been primarily driven by unsupervised learning from huge datasets. Although transformative, unsupervised models are limited to pattern matching and lack language as a means of achieving goals and goals. Achieving human-level cross-domain language understanding and generation requires research into reinforcement learning as a complement to complex and interactive language tasks.
This position paper argues that unsupervised learning should continue to be a priority in open language learning, but supplemented, where necessary, with reinforcement learning tailored to the task at hand. By designing languages as action sequences and optimizing models to achieve specific goals, RL can harness the power of unsupervised representation learning. It may be uniquely suited to address current limitations around abstract reasoning, grounded language understanding and common-sense knowledge that constrain broader real-world application of models like BERT and GPT-3.
Continued progress in self-supervised learning and world knowledge will enable more sophisticated unsupervised language models, but interactive learning with feedback is key to human-level language competence for open-domain conversation and problem-solving. Researchers should seek integrated modeling approaches, combining unsupervised pre-training, sparse supervision, and RL optimized for purpose, rather than framing progress as a choice between ML paradigms.
With a recognition of the strengths and limits of pattern matching versus interactive optimization, and how they may interact at different levels of language abstraction, NLP can achieve AI systems that understand, generate and reason about language with human-level versatility, open-domain breadth and purposeful skill. The future of human-level NLP is an integrated, interactive one that remains grounded in unsupervised learning but complements it where needed to fulfill each language task.
Submission Category: Machine learning algorithms
Submission Number: 1
Loading