In this paper, we present a continuation of our work on the personification of dialogue agents. We expand upon the previously demonstrated models—the ranking and generative models—and propose new hybrid models. Because there is no single definitive way to build a hybrid model, we explore various architectures where the components adopt different roles, sequentially and in parallel. Applying the perplexity and BLEU performance metrics, we discover that the Retrieve and Refine and KG model—a modification of the Retrieve and Refine model where the ranking and generative components work in parallel and compete based on the proximity of the candidate found by the ranking model with a knowledge-grounded generation block—achieves the best performance, with values of 1.64 for perplexity and 0.231 for BLEU scores.

The modern information retrieval field increasingly relies on hybrid search systems combining sparse retrieval with dense neural models. However, most existing hybrid frameworks employ static mixing coefficients and independent component training, failing to account for the specific needs of individual queries and corpus heterogeneity. In this paper, we introduce an adaptive hybrid retrieval framework featuring query-driven alpha prediction that dynamically calibrates the mixing weights based on query latent representations instantiated in a lightweight low-latency configuration and a full-capacity encoder-scale predictor, enabling flexible trade-offs between computational efficiency and retrieval accuracy without relying on resource-inefficient LLM-based online evaluation. Furthermore, we propose antagonist negative sampling, a novel training paradigm that optimizes the dense encoder to resolve the systematic failures of the lexical retriever, prioritizing hard negatives where BM25 exhibits high uncertainty. Empirical evaluations on large-scale multilingual benchmarks (MLDR and MIRACL) indicate that our approach demonstrates superior average performance compared to state-of-the-art models such as BGE-M3 and mGTE, achieving an nDCG@10 of 74.3 on long-document retrieval. Notably, our framework recovers up to 92.5% of the theoretical oracle performance and yields significant improvements in nDCG@10 across 16 languages, particularly in challenging long-context scenarios.

Pavel Posokhov, Sergei Masliukhin, Skrylnikov Stepan, Danil Tirskikh, Olesia Makhnytkina. Findings of the Association for Computational Linguistics: ACL 2025. 2025.

The exponential growth of digital information necessitates the development of robust text retrieval methods since most of the methods are domain or task-specific which limits their implementation. In this case multi-task learning is a promising alternative as it helps a model to have more meaningful embeddings; however such cases require usage of task separation methods. Many studies explore multi-task learning to improve generalization but tend to focus on large models. However, in real-world, speech analytics tasks that require searching through hundreds of millions of vectors in real-time, smaller models become more appropriate. This paper presents a novel approach to enhance the robustness of multi-task text retrieval models through the use of prompts. We use contrastive learning to train encoder models both in single-task and multi-task configurations and compare their performances as well as analyze the efficiency of different prompt usage strategies including hard prompts represented by explicit natural language instructions and soft prompts of varying lengths represented by model special tokens. Experiments are conducted by applying prompts to both the query and candidate document as well as to queries only keeping the candidate without prompts to reuse pre-encoded candidates in multi-task retrieval without significant quality loss. The obtained results are compared using R@1, R@5, and MRR metrics which are most applicable for evaluating in-domain and out-of-domain search. Single-task models show better performance on in-domain training data, while multi-task models demonstrate superior performance on out-of-domain data highlighting their increased robustness to domain shifts. Applying prompts to both elements–query and document–yields better performance than applying them solely to the query. Soft prompts are found to be preferable to hard as they better adapt the model to different domains. The findings of this study can be useful for improving text retrieval models, especially in scenarios involving multi-task systems where high adaptability and performance on new data are required. Trainable prompts could be an effective tool for enhancing the flexibility of models in various applications, such as information retrieval and question-answering systems.