Kunfeng Lai

Kunfeng Lai, Zhenheng Tang, Xinglin Pan et al.

Model merging aggregates Large Language Models (LLMs) finetuned on different tasks into a stronger one. However, parameter conflicts between models leads to performance degradation in averaging. While model routing addresses this issue by selecting individual models during inference, it imposes excessive storage and compute costs, and fails to leverage the common knowledge from different models. In this work, we observe that different layers exhibit varying levels of parameter conflicts. Building on this insight, we average layers with minimal parameter conflicts and use a novel task-level expert routing for layers with significant conflicts. To further reduce storage costs, inspired by task arithmetic sparsity, we decouple multiple fine-tuned experts into a dense expert and several sparse experts. Considering the out-of-distribution samples, we select and merge appropriate experts based on the task uncertainty of the input data. We conduct extensive experiments on both LLaMA and Qwen with varying parameter scales, and evaluate on real-world reasoning tasks. Results demonstrate that our method consistently achieves significant performance improvements while requiring less system cost compared to existing methods.

Yiying Yang, Xi Yin, Haiqin Yang et al.

Entity synonyms discovery is crucial for entity-leveraging applications. However, existing studies suffer from several critical issues: (1) the input mentions may be out-of-vocabulary (OOV) and may come from a different semantic space of the entities; (2) the connection between mentions and entities may be hidden and cannot be established by surface matching; and (3) some entities rarely appear due to the long-tail effect. To tackle these challenges, we facilitate knowledge graphs and propose a novel entity synonyms discovery framework, named \emph{KGSynNet}. Specifically, we pre-train subword embeddings for mentions and entities using a large-scale domain-specific corpus while learning the knowledge embeddings of entities via a joint TransC-TransE model. More importantly, to obtain a comprehensive representation of entities, we employ a specifically designed \emph{fusion gate} to adaptively absorb the entities' knowledge information into their semantic features. We conduct extensive experiments to demonstrate the effectiveness of our \emph{KGSynNet} in leveraging the knowledge graph. The experimental results show that the \emph{KGSynNet} improves the state-of-the-art methods by 14.7\% in terms of hits@3 in the offline evaluation and outperforms the BERT model by 8.3\% in the positive feedback rate of an online A/B test on the entity linking module of a question answering system.

Hao Peng, Jianxin Li, Qiran Gong et al.

Events are happening in real-world and real-time, which can be planned and organized occasions involving multiple people and objects. Social media platforms publish a lot of text messages containing public events with comprehensive topics. However, mining social events is challenging due to the heterogeneous event elements in texts and explicit and implicit social network structures. In this paper, we design an event meta-schema to characterize the semantic relatedness of social events and build an event-based heterogeneous information network (HIN) integrating information from external knowledge base, and propose a novel Pair-wise Popularity Graph Convolutional Network (PP-GCN) based fine-grained social event categorization model. We propose a Knowledgeable meta-paths Instances based social Event Similarity (KIES) between events and build a weighted adjacent matrix as input to the PP-GCN model. Comprehensive experiments on real data collections are conducted to compare various social event detection and clustering tasks. Experimental results demonstrate that our proposed framework outperforms other alternative social event categorization techniques.

Bang Liu, Weidong Guo, Di Niu et al.

Concepts embody the knowledge of the world and facilitate the cognitive processes of human beings. Mining concepts from web documents and constructing the corresponding taxonomy are core research problems in text understanding and support many downstream tasks such as query analysis, knowledge base construction, recommendation, and search. However, we argue that most prior studies extract formal and overly general concepts from Wikipedia or static web pages, which are not representing the user perspective. In this paper, we describe our experience of implementing and deploying ConcepT in Tencent QQ Browser. It discovers user-centered concepts at the right granularity conforming to user interests, by mining a large amount of user queries and interactive search click logs. The extracted concepts have the proper granularity, are consistent with user language styles and are dynamically updated. We further present our techniques to tag documents with user-centered concepts and to construct a topic-concept-instance taxonomy, which has helped to improve search as well as news feeds recommendation in Tencent QQ Browser. We performed extensive offline evaluation to demonstrate that our approach could extract concepts of higher quality compared to several other existing methods. Our system has been deployed in Tencent QQ Browser. Results from online A/B testing involving a large number of real users suggest that the Impression Efficiency of feeds users increased by 6.01% after incorporating the user-centered concepts into the recommendation framework of Tencent QQ Browser.

Ting Zhang, Bang Liu, Di Niu et al.

A large number of deep learning models have been proposed for the text matching problem, which is at the core of various typical natural language processing (NLP) tasks. However, existing deep models are mainly designed for the semantic matching between a pair of short texts, such as paraphrase identification and question answering, and do not perform well on the task of relevance matching between short-long text pairs. This is partially due to the fact that the essential characteristics of short-long text matching have not been well considered in these deep models. More specifically, these methods fail to handle extreme length discrepancy between text pieces and neither can they fully characterize the underlying structural information in long text documents. In this paper, we are especially interested in relevance matching between a piece of short text and a long document, which is critical to problems like query-document matching in information retrieval and web searching. To extract the structural information of documents, an undirected graph is constructed, with each vertex representing a keyword and the weight of an edge indicating the degree of interaction between keywords. Based on the keyword graph, we further propose a Multiresolution Graph Attention Network to learn multi-layered representations of vertices through a Graph Convolutional Network (GCN), and then match the short text snippet with the graphical representation of the document with the attention mechanisms applied over each layer of the GCN. Experimental results on two datasets demonstrate that our graph approach outperforms other state-of-the-art deep matching models.

Bang Liu, Mingjun Zhao, Di Niu et al.

Automatic question generation is an important technique that can improve the training of question answering, help chatbots to start or continue a conversation with humans, and provide assessment materials for educational purposes. Existing neural question generation models are not sufficient mainly due to their inability to properly model the process of how each word in the question is selected, i.e., whether repeating the given passage or being generated from a vocabulary. In this paper, we propose our Clue Guided Copy Network for Question Generation (CGC-QG), which is a sequence-to-sequence generative model with copying mechanism, yet employing a variety of novel components and techniques to boost the performance of question generation. In CGC-QG, we design a multi-task labeling strategy to identify whether a question word should be copied from the input passage or be generated instead, guiding the model to learn the accurate boundaries between copying and generation. Furthermore, our input passage encoder takes as input, among a diverse range of other features, the prediction made by a clue word predictor, which helps identify whether each word in the input passage is a potential clue to be copied into the target question. The clue word predictor is designed based on a novel application of Graph Convolutional Networks onto a syntactic dependency tree representation of each passage, thus being able to predict clue words only based on their context in the passage and their relative positions to the answer in the tree. We jointly train the clue prediction as well as question generation with multi-task learning and a number of practical strategies to reduce the complexity. Extensive evaluations show that our model significantly improves the performance of question generation and out-performs all previous state-of-the-art neural question generation models by a substantial margin.

Bang Liu, Di Niu, Kunfeng Lai et al.

We describe our experience of implementing a news content organization system at Tencent that discovers events from vast streams of breaking news and evolves news story structures in an online fashion. Our real-world system has distinct requirements in contrast to previous studies on topic detection and tracking (TDT) and event timeline or graph generation, in that we 1) need to accurately and quickly extract distinguishable events from massive streams of long text documents that cover diverse topics and contain highly redundant information, and 2) must develop the structures of event stories in an online manner, without repeatedly restructuring previously formed stories, in order to guarantee a consistent user viewing experience. In solving these challenges, we propose Story Forest, a set of online schemes that automatically clusters streaming documents into events, while connecting related events in growing trees to tell evolving stories. We conducted extensive evaluation based on 60 GB of real-world Chinese news data, although our ideas are not language-dependent and can easily be extended to other languages, through detailed pilot user experience studies. The results demonstrate the superior capability of Story Forest to accurately identify events and organize news text into a logical structure that is appealing to human readers, compared to multiple existing algorithm frameworks.

Bang Liu, Ting Zhang, Fred X. Han et al.

Semantic matching of natural language sentences or identifying the relationship between two sentences is a core research problem underlying many natural language tasks. Depending on whether training data is available, prior research has proposed both unsupervised distance-based schemes and supervised deep learning schemes for sentence matching. However, previous approaches either omit or fail to fully utilize the ordered, hierarchical, and flexible structures of language objects, as well as the interactions between them. In this paper, we propose Hierarchical Sentence Factorization---a technique to factorize a sentence into a hierarchical representation, with the components at each different scale reordered into a "predicate-argument" form. The proposed sentence factorization technique leads to the invention of: 1) a new unsupervised distance metric which calculates the semantic distance between a pair of text snippets by solving a penalized optimal transport problem while preserving the logical relationship of words in the reordered sentences, and 2) new multi-scale deep learning models for supervised semantic training, based on factorized sentence hierarchies. We apply our techniques to text-pair similarity estimation and text-pair relationship classification tasks, based on multiple datasets such as STSbenchmark, the Microsoft Research paraphrase identification (MSRP) dataset, the SICK dataset, etc. Extensive experiments show that the proposed hierarchical sentence factorization can be used to significantly improve the performance of existing unsupervised distance-based metrics as well as multiple supervised deep learning models based on the convolutional neural network (CNN) and long short-term memory (LSTM).

Bang Liu, Di Niu, Haojie Wei et al.

Identifying the relationship between two articles, e.g., whether two articles published from different sources describe the same breaking news, is critical to many document understanding tasks. Existing approaches for modeling and matching sentence pairs do not perform well in matching longer documents, which embody more complex interactions between the enclosed entities than a sentence does. To model article pairs, we propose the Concept Interaction Graph to represent an article as a graph of concepts. We then match a pair of articles by comparing the sentences that enclose the same concept vertex through a series of encoding techniques, and aggregate the matching signals through a graph convolutional network. To facilitate the evaluation of long article matching, we have created two datasets, each consisting of about 30K pairs of breaking news articles covering diverse topics in the open domain. Extensive evaluations of the proposed methods on the two datasets demonstrate significant improvements over a wide range of state-of-the-art methods for natural language matching.