Enriching Transformers with Structured Tensor-Product Representations for Abstractive Summarization