Product retrieval is the backbone of e-commerce search: for each user query, it identifies a high-recall candidate set from billions of items, laying the foundation for high-quality ranking and user experience. Despite extensive optimization for mainstream queries, existing systems still struggle with long-tail queries, especially knowledge-intensive ones. These queries exhibit diverse linguistic patterns, often lack explicit purchase intent, and require domain-specific knowledge reasoning for accurate interpretation. They also suffer from a shortage of reliable behavioral logs, which makes such queries a persistent challenge for retrieval optimization.

Recently, deep learning-based methods has gained increasing attention in typhoon tasks. Due to different optimization targets, existing works apply multi-object detection to typhoon detection and keypoint detection to typhoon eye location. However, such two-stage methods ignored the internal connection between tasks in this special scenario, leading to complexity and time consumption. Therefore, we propose Simple Mask-Flow Keypoint Detection(SimMKD), a unified framework for two typhoon tasks that optimizes different targets uniformly. In SimMKD, our Mask-Flow mechanism simplifies typhoon detection from a multi-object into single-object detection tasks in turn, while our Hierarchic Keypoint Detection (HKD) model reformulates it as a two-point detection task, then merged with typhoon eye location task. Hence two different tasks are simplified and integrated as a single keypoint detection task. Furthermore, we publish the Multi-object Typhoon Eye Dataset (MTED), which meets more task requirements. Extensive experimental results show that the SimMKD can achieve new SOTA on existing datasets.1

Zero-shot relation extraction aims to extract novel relations that are not observed beforehand. However, existing representation methods are not pre-trained for relational representations and embeddings contain much linguistic information, the distances between them are not consistent with relational semantic similarity. In this paper, we propose a novel method based on Instance Prompting and Prototype Rectification (IPPR) to conduct relational representation learning for zeroshot relation extraction. Instance prompting is designed to reduce the gap between pre-training and fine-tuning, and guide the pre-trained model to generate relation-oriented instance representations. Prototype rectification aims to push the prototype embeddings away from each other and makes the instance embeddings closer to its corresponding prototype embeddings for dynamically rectifying the prototype embeddings. Experimental results on two public datasets demonstrate that our proposed method achieves new state-of-the-arts performance1.

Query-focused meeting summarization (QFMS) aims to generate suimnaries from meeting transcripts in response to a given query. Previous works typically concatenate the query with meeting transcripts and implicitly model the query relevance only at the token level with attention mechanism. However, due to the dilution of key query-relevant infonnation caused by long meeting transcripts, the original transformer-based model is insufficient to highlight the key parts related to the query. In this paper, we propose a query-aware framework with joint modeling token and utterance based on Query-Utterance Attention. It calculates the utterance-level relevance to the query with a dense retrieval module. Then both token-level query relevance and utterance-level query relevance are combined and incorporated into the generation process with attention mechanism explicitly. We show that the query relevance of different granularities contributes to generating a suimnary more related to the query. Experimental results on the QMSum dataset show that the proposed model achieves new state-of-the-art perfonnance.

In recent years, as the basis for the downstream image-text task, Visual Grounding has attracted more and more attention. Existing models treat the feature extraction process of image and text as two completely independent parts. However, due to a lack of effective interaction between text and image, especially the understanding deviation of keywords. It tends to weaken the alignment between the extracted visual features and the target. Therefore, in this paper, we propose a Weighted Keywords-Guided feature Extraction and Reasoning Network(WKG-Net) to enhance the interaction of text and image. We design a Weighted Keywords-Guided Extractor(WKG-Extractor), which can add extra keywords information as the input to guide the further extraction from general visual features to guided visual features. Then a Weighted Keywords-Guided Reasoner(WKG-Reasoner) is introduced to fuse weighted keywords, original text and visual features together to enhance the alignment between the image and text. Enhancing the understanding of keywords will be helpful for all tasks of text and image alignment. Extensive experiments show that our method has a greater performance compared with the baseline on the three mainstream datasets.

Supervised open relation extraction aims to discover novel relations by leveraging supervised data of pre-defined relations. However, most existing methods do not achieve effective knowledge transfer from pre-defined relations to novel relations, they have difficulties generating high-quality pseudo-labels for unsupervised data of novel relations and usually suffer from the error propagation issue. In this paper, we propose a Cluster-aware Pseudo-Labeling (CaPL) method to improve the pseudo-labels quality and transfer more knowledge for discovering novel relations. Specifically, the model is firstly pre-trained with the pre-defined relations to learn the relation representations. To improve the pseudo-labels quality, the distances between each instance and all cluster centers are used to generate the cluster-aware soft pseudo-labels for novel relations. To mitigate the catastrophic forgetting issue, we design the consistency regularization loss to make better use of the pseudo-labels and jointly train the model with both unsupervised and supervised data. Experimental results on two public datasets demonstrate that our proposed method achieves new state-of-the-arts performance.