Given the criminal fact, a complete legal judgment usually contains two parts: Court View Generation (CVG) for the case and Legal Judgement Prediction (LJP) for each defendant. However, existing studies still cannot support unified modeling and evaluation of complete legal judgments. First, at the data level, existing legal datasets mostly cover either single-defendant scenarios or multi-defendant scenarios, and the attributes predicted by LJP are not rich enough, making it difficult for models to generalize to real-world complete legal judgments. Second, CVG and LJP are usually modeled independently, ignoring their intrinsic connection. Third, the role of auxiliary articles has not been explored. To address these issues, we build CMML (Chinese Mixed-Defendant Multi-Attribute Legal Judgment Dataset), a large-scale dataset that mixes single- and multi-defendant cases with multi-attribute annotations. Based on CMML, we propose a new task, LJP-x, which extends LJP with additional prediction attributes. We further systematically study the interaction between CVG and LJP-x, and explore how auxiliary articles influence both tasks. Our dataset is available at: https://github.com/magicalpig2003/CMML.

Knowledge retrieval and response generation are fundamental to task-oriented dialogue systems. However, dialogue context frequently contains noisy or irrelevant information, leading to sub-optimal result in knowledge retrieval. One possible approach to retrieving knowledge is to manually annotate standard queries for each dialogue. Yet, this approach is hindered by the challenge of data scarcity, as human annotation is costly. To solve the challenge, we propose an LLM-enhanced model of query-guided knowledge retrieval for task-oriented dialogue. It generates high-quality queries for knowledge retrieval in task-oriented dialogue solely using low-resource annotated queries. To strengthen the performance correlation between response generation and knowledge retrieval, we propose a retrieval preservation mechanism by further selecting the most relevant knowledge from retrieved top-K records and explicitly incorporating these as prompts to guide a generator in response generation. Experiments on three standard benchmarks demonstrate that our model and mechanism outperform previous state-of-the-art by 3.26% on average with two widely used evaluation metrics.