Google AI introduces Minerva, a model for natural language processing (NLP) that solves mathematical questions
In a wide range of natural language tasks such as question answering, reasoning with common sense, and summarization, large language models are widely used. However, these models have struggled with tasks that require quantitative reasoning such as solving problems in mathematics, engineering, and physics.
Researchers have found that language models can be used to test quantitative reasoning in a variety of ways. For solving mathematical and science problems, it is necessary to be able to parse a question using normal language and mathematical symbols, remember relevant formulas and constants, and provide step-by-step solutions requiring numerical calculations and symbolic manipulation. Scientists have therefore believed that machine-learning models would require significant improvements to model architecture and training techniques in order to solve these reasoning problems.
Google has released a new research that introduces Minerva. This language model uses sequential reasoning in order to solve mathematical and scientific questions. Minerva solves these problems by incorporating numerical calculations and symbolic manipulation.