Modern recommender systems employ various sequential modules such as self-attention to learn dynamic user interests. However, these methods are less effective in capturing collaborative and transitional signals within interaction sequences. First, the architecture uses embedding of a single item attention query, making it challenging capture signals. Second, typically follow an auto-regressive framework, which is unable global transition patterns. To overcome limitations, we propose new...

Large Language Models (LLMs) excel in various tasks, including personalized recommendations. Existing evaluation methods often focus on rating prediction, relying regression errors between actual and predicted ratings. However, user bias item quality, two influential factors behind scores, can obscure personal preferences user-item pair data. To address this, we introduce PerRecBench, disassociating the from these assessing recommendation techniques capturing a grouped ranking manner. We...

Humans excel in analogical learning and knowledge transfer and, more importantly, possess a unique understanding of identifying appropriate sources knowledge. From model's perspective, this presents an interesting challenge. If models could autonomously retrieve useful for or decision-making to solve problems, they would transition from passively acquiring actively accessing However, filling with is relatively straightforward -- it simply requires training accessible bases. The complex task...

Future link prediction is a fundamental challenge in various real-world dynamic systems. To address this, numerous temporal graph neural networks (temporal GNNs) and benchmark datasets have been developed. However, these often feature excessive repeated edges lack complex sequential dynamics, key characteristic inherent many applications such as recommender systems ``Who-To-Follow'' on social networks. This oversight has led existing methods to inadvertently downplay the importance of...

This survey examines the evolution of model architectures in information retrieval (IR), focusing on two key aspects: backbone models for feature extraction and end-to-end system relevance estimation. The review intentionally separates architectural considerations from training methodologies to provide a focused analysis structural innovations IR systems.We trace development traditional term-based methods modern neural approaches, particularly highlighting impact transformer-based subsequent...

In conversational search, the user’s real search intent for current conversation turn is dependent on previous history. It challenging to determine a good query from whole context. To avoid expensive re-training of encoder, most existing methods try learn rewriting model de-contextualize by mimicking manual rewriting.However, manually rewritten queries are not always best queries.Thus, training them would lead sub-optimal queries. Another useful information enhance potential answer question....

Precisely understanding users' contextual search intent has been an important challenge for conversational search. As sessions are much more diverse and long-tailed, existing methods trained on limited data still show unsatisfactory effectiveness robustness to handle real scenarios. Recently, large language models (LLMs) have demonstrated amazing capabilities text generation conversation understanding. In this work, we present a simple yet effective prompting framework, called LLM4CS,...

Conversational search supports multi-turn user-system interactions to solve complex information needs. Compared with the traditional single-turn ad-hoc search, conversational faces a more intent understanding problem because session is much longer and contains many noisy tokens. However, existing dense retrieval solutions simply fine-tune pre-trained query encoder on limited data, which are hard achieve satisfactory performance in such scenario. Meanwhile, learned latent representation also...

Word-level information is important in natural language processing (NLP), especially for the Chinese due to its high linguistic complexity. word segmentation (CWS) an essential task downstream NLP tasks. Existing methods have already achieved a competitive performance CWS on large-scale annotated corpora. However, accuracy of method will drop dramatically when it handles unsegmented text with lots out-of-vocabulary (OOV) words. In addition, there are many different criteria addressing...

Graph-based models and contrastive learning have emerged as prominent methods in Collaborative Filtering (CF). While many existing CF incorporate these their design, there seems to be a limited depth of analysis regarding the foundational principles behind them. This paper bridges graph convolution, pivotal element graph-based models, with through theoretical framework. By examining dynamics equilibrium loss, we offer fresh lens understand via theory, emphasizing its capability capture...

Conversational search provides users with a natural and convenient new experience. Recently, conversational dense retrieval has shown to be promising technique for realizing search. However, as systems have not been widely deployed, it is hard get large-scale real sessions relevance labels support the training of retrieval. To tackle this data scarcity problem, previous methods focus on developing better few-shot learning approaches or generating pseudo labels, but they use still heavily...

Conversational search allows a user to interact with system in multiple turns. A query is strongly dependent on the conversation context. An effective way improve retrieval effectiveness expand current historical queries. However, not all previous queries are related to, and useful for expanding query. In this paper, we propose new method select relevant that To cope lack of labeled training data, use pseudo-labeling approach annotate based their impact results. The pseudo-labeled data used...

As privacy issues are receiving increasing attention within the Natural Language Processing (NLP) community, numerous methods have been proposed to sanitize texts subject differential privacy. However, state-of-the-art text sanitization mechanisms based on a relaxed notion of metric local (MLDP) do not apply non-metric semantic similarity measures and cannot achieve good privacy-utility trade-offs. To address these limitations, we propose novel Customized Text (CusText) mechanism original...

Conversational query rewriting (CQR) realizes conversational search by reformulating the dialogue into a standalone rewrite. However, existing CQR models either are not learned toward improving downstream performance or inefficiently generate rewrite token-by-token from scratch while neglecting fact that often has large overlap with In this paper, we propose EdiRCS, new text editing-based model tailored for search. most of tokens selected in non-autoregressive fashion and only few generated...

Large Language Models (LLMs) are essential tools to collaborate with users on different tasks. Evaluating their performance serve users' needs in real-world scenarios is important. While many benchmarks have been created, they mainly focus specific predefined model abilities. Few covered the intended utilization of LLMs by real users. To address this oversight, we propose benchmarking from a user perspective both dataset construction and evaluation designs. We first collect 1846 use cases 15...

The vanilla Differentially-Private Stochastic Gradient Descent (DP-SGD), including DP-Adam and other variants, ensures the privacy of training data by uniformly distributing costs across steps. equivalent controlled maintaining same gradient clipping thresholds noise powers in each step result unstable updates a lower model accuracy when compared to non-DP counterpart. In this paper, we propose dynamic DP-SGD (along with DP-Adam, others) reduce performance loss gap while dynamically...

Word segmentation is an essential and challenging task in natural language processing, especially for the Chinese due to its high linguistic complexity. Existing methods word segmentation, including statistical machine learning neural network methods, usually have good performance specific knowledge domains. Given increasing importance of interdisciplinary cross-domain studies, one challenges handle out-of-vocabulary (OOV) words. show unsatisfactory meet practical standard. To this end, we...

Conversational search facilitates complex information retrieval by enabling multi-turn interactions between users and the system. Supporting such requires a comprehensive understanding of conversational inputs to formulate good query based on historical information. In particular, should include relevant from previous conversation turns. However, current approaches for dense primarily rely fine-tuning pre-trained ad-hoc retriever using whole session, which can be lengthy noisy. Moreover,...

Conversational search requires accurate interpretation of user intent from complex multi-turn contexts. This paper presents ChatRetriever, which inherits the strong generalization capability large language models to robustly represent conversational sessions for dense retrieval. To achieve this, we propose a simple and effective dual-learning approach that adapts LLM retrieval via contrastive learning while enhancing session understanding through masked instruction tuning on high-quality...

Conversational search provides a more convenient interface for users to by allowing multi-turn interaction with the engine. However, effectiveness of conversational dense retrieval methods is limited scarcity training data required their fine-tuning. Thus, generating sessions relevant labels could potentially improve performance. Based on promising capabilities large language models (LLMs) text generation, we propose ConvSDG, simple yet effective framework explore feasibility boosting using...

Document-level biomedical concept extraction is the task of identifying concepts mentioned in a given document. Recent advancements have adapted pre-trained language models for this task. However, scarcity domain-specific data and deviation from their canonical names often hinder these models' effectiveness. To tackle issue, we employ MetaMapLite, an existing rule-based mapping system, to generate additional pseudo-annotated PubMed PMC. The annotated are used augment limited training data....

Conversational search supports multi-turn user-system interactions to solve complex information needs. Different from the traditional single-turn ad-hoc search, conversational encounters a more challenging problem of context-dependent query understanding with lengthy and long-tail history context. While rewriting (CQR) methods leverage explicit rewritten queries train model transform into stand-stone query, this is usually done without considering quality results. dense retrieval (CDR) use...