Abstract:  
Abstract: The task of recommendation-oriented dialog generation aims to achieve accurate recommendations by obtaining user preferences through human-computer dialog interactions. In response to the problem of limited dialog recommendation types and low quality of generated replies in existing research， this paper proposes a Goal Driven Recommendation-oriented Dialog Generation model （GDRDG） based on the Unified Language Model pre-training （UniLM）. The model comprises a text representation module， a multi-head encoding module， a decoding module， and a specialized attention masking mechanism. The text representation module uses UniLM to vectorize the input text， ensuring that the model captures deep semantic features of the text. The multi-head encoding module employs a multi-head self-attention mechanism to capture global contextual information， enhancing the coherence and relevance of the generated responses. The decoding module generates the target of the current dialogue round and the response based on this target， ensuring that the reply is consistent with the context and guides the conversation towards the intended goal. The special attention masking mechanism is used to control the information flow during the decoding process， ensuring that the model focuses only on information relevant to the current round， thereby improving the quality of the response. Experimental results demonstrate that the proposed GDRDG model outperforms existing methods in metrics such as BLEU， Distinct， F1， and Hit@1， thereby validating the model’s effectiveness and advancement.

Key words: Key words: goal driven, recommendation dialog, dialog generation, unified language model pre-training, attention mechanism