A Computational Model for Cultivating Innovative Ability in College English Based on Deep Learning and Combinatorial Mathematics

Source abstract

In the context of AI-enabled education, College English teaching is gradually shifting from a predominantly knowledge-transmission model toward an approach that also emphasizes the cultivation of innovative language abilities. Nevertheless, innovative language production is structurally complex, dynamically evolving, and difficult to quantify in a transparent manner. Existing deep-learning approaches frequently concentrate on outcome prediction or text generation and provide only limited explicit modeling of learners’ cognitive structures and the mechanisms through which innovative expressions are formed. To address these limitations, this paper proposes a computational modeling framework for innovation-oriented College English learning that integrates deep learning with combinatorial mathematics. First, a hierarchical Transformer-based semantic encoder maps discrete linguistic units into a continuous latent semantic embedding space, thereby providing a measurable semantic representation. Second, a cognitive-structure graph is constructed in which linguistic knowledge units are treated as nodes, and the innovation-oriented learning process is formalized as a structural mapping problem. Knowledge reorganization is then represented through combinatorial operations, including permutation, substitution, merging, and reconstruction. On this basis, a multi-objective optimization function that balances semantic novelty, semantic coherence, and structural complexity is formulated so that innovation generation can be treated as a computable combinatorial optimization problem. Approximate solutions are obtained using a genetic algorithm, beam search, and reinforcement learning. Systematic experiments are conducted on an authentic College English writing corpus. The reported results indicate that the proposed method outperforms the comparison methods with respect to innovation score, novelty distribution, structural controllability, and learning stability. Multidimensional visualization further supports the effectiveness of the framework in representing innovative ability, cognitive-structure evolution, and adaptation to individual differences.