Conditional Self-Attention for Query-based Summarization
This addresses the problem of generating summaries conditioned on queries for NLP applications, representing an incremental advance in attention mechanisms.
The paper tackles query-based summarization by introducing conditional self-attention to model dependencies relevant to a query, achieving consistent performance improvements over vanilla Transformer and previous models on Debatepedia and HotpotQA benchmarks.
Self-attention mechanisms have achieved great success on a variety of NLP tasks due to its flexibility of capturing dependency between arbitrary positions in a sequence. For problems such as query-based summarization (Qsumm) and knowledge graph reasoning where each input sequence is associated with an extra query, explicitly modeling such conditional contextual dependencies can lead to a more accurate solution, which however cannot be captured by existing self-attention mechanisms. In this paper, we propose \textit{conditional self-attention} (CSA), a neural network module designed for conditional dependency modeling. CSA works by adjusting the pairwise attention between input tokens in a self-attention module with the matching score of the inputs to the given query. Thereby, the contextual dependencies modeled by CSA will be highly relevant to the query. We further studied variants of CSA defined by different types of attention. Experiments on Debatepedia and HotpotQA benchmark datasets show CSA consistently outperforms vanilla Transformer and previous models for the Qsumm problem.