A Hybrid Chinese Conversation model based on retrieval and generation
A Hybrid Chinese Conversation model based on retrieval and generation
复制标题
DOI：
10.1016/j.future.2020.08.030
复制
发表时间：
2021
期刊：
影响因子：
--
通讯作者：
Tinghuai Ma;Huimin Yang;Qing Tian;Y. Tian;N. Al-Nabhan
中科院分区：
文献类型：
--
作者：
Tinghuai Ma;Huimin Yang;Qing Tian;Y. Tian;N. Al-Nabhan
会话生成是一项重要的自然语言处理任务，近年来备受关注。会话模型的实现对社会计算领域也具有重要意义，有助于在社交网络上构建人工智能机器人。开放领域会话模型本质上是数据驱动的，大致可以分为检索模型和生成模型。尽管近年来取得了显著的进展，但仍然很难得到在语法和语义上合适的答复。我们提出了一种基于检索和基于转换的会话模型(RRT)，该模型将检索模型和生成模型结合起来，以获得与上下文相关的响应。使用传统的检索方法从训练集中检索相似度最高的上下文-响应对，并进一步排序以获得检索候选响应。我们用转换器模型代替了传统的序列到序列模型进行会话生成，以更少的训练时间取得了更好的效果。最后，使用后重排序模块对检索到的候选对象和生成的候选对象进行排序，以获得最终答案。我们在两个数据集上进行了详细的实验，结果表明，与传统的生成模型相比，我们的模型在各个指标上都有显著的改善，训练时间减少了5倍。此外，我们的模型比检索模型更具信息性和与输入上下文的相关性。
Conversation generation is an important natural language processing task and has attracted much attention in recent years. The realization of the conversation model is also of great significance to the field of social computing, helping to build artificial intelligence robots on social networks. The open domain conversation model is fundamentally data-driven, which can be roughly divided into retrieval models and generation models. Although remarkable progress has been achieved in recent years, it is still difficult to get responses that are grammatically and semantically appropriate. We propose the Rerank of Retrieval-based and Transformer-based Conversation model (RRT), a novel conversation model that combines the retrieval model with the generation model for the purpose of obtaining context–appropriate response. The context–response pairs with the highest similarity from training set are retrieved using traditional retrieval method, and further ranked to obtain the retrieval candidate response. We replaced the traditional sequence-to-sequence models for conversation generation by the transformer model and achieved better results with less training time. Finally, the post-reranking module is used to rank the retrieved candidate and the generated one to obtain the final response. We conducted detailed experiments on two datasets and the results show that compared with the traditional generation model, our model has a significant improvement in each metric, and the training time is decreased by a factor of 5. Furthermore, our model is more informative and relevant to the input context than the retrieval model.