A new paper on arXiv, titled "Mixture of Decoders for Diverse Dialog Response Generation," tackles a well-known weakness in sequence-to-sequence dialog models: their tendency to produce low-diversity responses. The authors draw on mixture modeling, a classic machine learning approach for handling multi-modal data, and apply it to the decoder side of the architecture.
The abstract explains that mixture modeling has long been used to learn from large sets of multi-modal data, but it does not detail the specific architecture or experimental results in the available excerpt. What is clear is the motivation: standard seq2seq models for dialog often fall into repetitive or generic outputs, and the proposed method aims to diversify the generated responses.
Because the source is only an abstract, the article does not include evaluation metrics or comparisons to other methods. The contribution, as stated, is the conceptual application of mixture modeling to decoders for dialog generation, with the goal of improving diversity without abandoning the seq2seq framework.