In multi-party group conversations, identifying the reply-to relation between historical messages is an important task in the dialogue domain. Despite of previous efforts, two issues with respect to the data distribution still remained: First, short messages with sparse semantics make up a significant portion of the messages, which in turn restricts the learning potential of the models. Second, positive examples with reply-to relations are often much fewer than negative examples, resulting in data skewness during model training and hindering the model''s performance on positive examples. To address these two issues, this paper proposes an improved model based on a pre-trained language model. Our method first mitigates the issue of short messages by developing a dynamic inquiry window that enriches semantic modeling with comprehensive semantics. Then, it tackles the problem of positive example imbalance through position-driven optimization of positive example weights. Experimental results on the public benchmark show that our method improved model achieves a recall of 62.2% and a F-1 score of 59.4%, which are 15.7% and 8.5% higher than the average baseline model, respectively. The paper also constructs a new dataset collected from the Telegram platform, providing data support for future related research.