Data-driven approaches for material discovery and design have been accelerated by emerging efforts in machine learning. While there is enormous progress towards learning the structure to property relationship of materials, methods that allow for general representations of crystals to effectively explore the vast material search space and identify high-performance candidates remain limited. In this work, we introduce a material discovery framework that uses natural language embeddings derived from material science-specific language models as representations of compositional and structural features. The discovery framework consists of a joint scheme that, given a query material, first recalls candidates based on representational similarity, and ranks the candidates based on target properties through multi-task learning. The contextual knowledge encoded in language representations is found to convey information about material properties and structures, enabling both similarity analysis for recall, and multi-task learning to share information for related properties. By applying the discovery framework to thermoelectric materials, we demonstrate diversified recommendations of prototype structures and identify under-studied high-performance material spaces, including halide perovskite, delafossite-like, and spinel-like structures. By leveraging material language representations, our framework provides a generalized means for effective material recommendation, which is task-agnostic and can be applied to various material systems.