Molecular generation from natural language descriptions is becoming an important approach for guided molecule design, as it allows researchers to express chemical objectives directly in textual form. However, string representations such as SMILES and SELFIES reside in embedding spaces that differ significantly from natural language, creating a mismatch that prevents generative models from accurately capturing the intended chemical semantics. This gap raises the question of whether a shared representation space can be constructed in which textual descriptions and molecular strings converge in a controlled manner. Motivated by this gap, we introduce ChemAligner-T5, a BioT5+ base model enhanced with a contrastive learning mechanism to directly align textual and molecular representations. On the L+M-24 test set, ChemAligner-T5 achieves a BLEU score of 69.77% and a Levenshtein distance of 31.28%, outperforming MolT5-base and Meditron on both metrics. Visual analysis shows that the model successfully reproduces the structural scaffold and key functional groups of the target molecule. These results highlight the importance of text–molecule representation alignment for the Text2Mol task and strengthen the potential of language models as direct interfaces for molecule design and drug discovery guided by natural-language descriptions. Source code available at GitHub .