Jun Ma

Zhengyuan Peng, Jinpeng Ma, Zhimin Sun et al.

Generalized Category Discovery (GCD) is a classification task that aims to classify both base and novel classes in unlabeled images, using knowledge from a labeled dataset. In GCD, previous research overlooks scene information or treats it as noise, reducing its impact during model training. However, in this paper, we argue that scene information should be viewed as a strong prior for inferring novel classes. We attribute the misinterpretation of scene information to a key factor: the Ambiguity Challenge inherent in GCD. Specifically, novel objects in base scenes might be wrongly classified into base categories, while base objects in novel scenes might be mistakenly recognized as novel categories. Once the ambiguity challenge is addressed, scene information can reach its full potential, significantly enhancing the performance of GCD models. To more effectively leverage scene information, we propose the Modeling Object-Scene Associations (MOS) framework, which utilizes a simple MLP-based scene-awareness module to enhance GCD performance. It achieves an exceptional average accuracy improvement of 4% on the challenging fine-grained datasets compared to state-of-the-art methods, emphasizing its superior performance in fine-grained GCD. The code is publicly available at https://github.com/JethroPeng/MOS

Kartikeya Bhardwaj, Nilesh Prasad Pandey, Sweta Priyadarshi et al.

Large generative models such as large language models (LLMs) and diffusion models have revolutionized the fields of NLP and computer vision respectively. However, their slow inference, high computation and memory requirement makes it challenging to deploy them on edge devices. In this study, we propose a light-weight quantization aware fine tuning technique using knowledge distillation (KD-QAT) to improve the performance of 4-bit weight quantized LLMs using commonly available datasets to realize a popular language use case, on device chat applications. To improve this paradigm of finetuning, as main contributions, we provide insights into stability of KD-QAT by empirically studying the gradient propagation during training to better understand the vulnerabilities of KD-QAT based approaches to low-bit quantization errors. Based on our insights, we propose ov-freeze, a simple technique to stabilize the KD-QAT process. Finally, we experiment with the popular 7B LLaMAv2-Chat model at 4-bit quantization level and demonstrate that ov-freeze results in near floating point precision performance, i.e., less than 0.7% loss of accuracy on Commonsense Reasoning benchmarks.

Liqiang Xiao, Jun Ma2, Xin Luna Dong et al.

Successful conversational search systems can present natural, adaptive and interactive shopping experience for online shopping customers. However, building such systems from scratch faces real word challenges from both imperfect product schema/knowledge and lack of training dialog data.In this work we first propose ConvSearch, an end-to-end conversational search system that deeply combines the dialog system with search. It leverages the text profile to retrieve products, which is more robust against imperfect product schema/knowledge compared with using product attributes alone. We then address the lack of data challenges by proposing an utterance transfer approach that generates dialogue utterances by using existing dialog from other domains, and leveraging the search behavior data from e-commerce retailer. With utterance transfer, we introduce a new conversational search dataset for online shopping. Experiments show that our utterance transfer method can significantly improve the availability of training dialogue data without crowd-sourcing, and the conversational search system significantly outperformed the best tested baseline.

Bin Ji, Shasha Li, Jie Yu et al.

Span-based joint extraction simultaneously conducts named entity recognition (NER) and relation extraction (RE) in text span form. Recent studies have shown that token labels can convey crucial task-specific information and enrich token semantics. However, as far as we know, due to completely abstain from sequence tagging mechanism, all prior span-based work fails to use token label in-formation. To solve this problem, we pro-pose Sequence Tagging enhanced Span-based Network (STSN), a span-based joint extrac-tion network that is enhanced by token BIO label information derived from sequence tag-ging based NER. By stacking multiple atten-tion layers in depth, we design a deep neu-ral architecture to build STSN, and each atten-tion layer consists of three basic attention units. The deep neural architecture first learns seman-tic representations for token labels and span-based joint extraction, and then constructs in-formation interactions between them, which also realizes bidirectional information interac-tions between span-based NER and RE. Fur-thermore, we extend the BIO tagging scheme to make STSN can extract overlapping en-tity. Experiments on three benchmark datasets show that our model consistently outperforms previous optimal models by a large margin, creating new state-of-the-art results.

Minghong Xu, Piji Li, Haoran Yang et al.

Unstructured Persona-oriented Dialogue Systems (UPDS) has been demonstrated effective in generating persona consistent responses by utilizing predefined natural language user persona descriptions (e.g., "I am a vegan"). However, the predefined user persona descriptions are usually short and limited to only a few descriptive words, which makes it hard to correlate them with the dialogues. As a result, existing methods either fail to use the persona description or use them improperly when generating persona consistent responses. To address this, we propose a neural topical expansion framework, namely Persona Exploration and Exploitation (PEE), which is able to extend the predefined user persona description with semantically correlated content before utilizing them to generate dialogue responses. PEE consists of two main modules: persona exploration and persona exploitation. The former learns to extend the predefined user persona description by mining and correlating with existing dialogue corpus using a variational auto-encoder (VAE) based topic model. The latter learns to generate persona consistent responses by utilizing the predefined and extended user persona description. In order to make persona exploitation learn to utilize user persona description more properly, we also introduce two persona-oriented loss functions: Persona-oriented Matching (P-Match) loss and Persona-oriented Bag-of-Words (P-BoWs) loss which respectively supervise persona selection in encoder and decoder. Experimental results show that our approach outperforms state-of-the-art baselines, in terms of both automatic and human evaluations.

Pengjie Ren, Zhumin Chen, Christof Monz et al.

Background Based Conversations (BBCs) have been introduced to help conversational systems avoid generating overly generic responses. In a BBC, the conversation is grounded in a knowledge source. A key challenge in BBCs is Knowledge Selection (KS): given a conversational context, try to find the appropriate background knowledge (a text fragment containing related facts or comments, etc.) based on which to generate the next response. Previous work addresses KS by employing attention and/or pointer mechanisms. These mechanisms use a local perspective, i.e., they select a token at a time based solely on the current decoding state. We argue for the adoption of a global perspective, i.e., pre-selecting some text fragments from the background knowledge that could help determine the topic of the next response. We enhance KS in BBCs by introducing a Global-to-Local Knowledge Selection (GLKS) mechanism. Given a conversational context and background knowledge, we first learn a topic transition vector to encode the most likely text fragments to be used in the next response, which is then used to guide the local KS at each decoding timestamp. In order to effectively learn the topic transition vector, we propose a distantly supervised learning schema. Experimental results show that the GLKS model significantly outperforms state-of-the-art methods in terms of both automatic and human evaluation. More importantly, GLKS achieves this without requiring any extra annotations, which demonstrates its high degree of scalability.

Ke Liu, Xiangyan Sun, Jun Ma et al.

A deep neural network based architecture was constructed to predict amino acid side chain conformation with unprecedented accuracy. Amino acid side chain conformation prediction is essential for protein homology modeling and protein design. Current widely-adopted methods use physics-based energy functions to evaluate side chain conformation. Here, using a deep neural network architecture without physics-based assumptions, we have demonstrated that side chain conformation prediction accuracy can be improved by more than 25%, especially for aromatic residues compared with current standard methods. More strikingly, the prediction method presented here is robust enough to identify individual conformational outliers from high resolution structures in a protein data bank without providing its structural factors. We envisage that our amino acid side chain predictor could be used as a quality check step for future protein structure model validation and many other potential applications such as side chain assignment in Cryo-electron microscopy, crystallography model auto-building, protein folding and small molecule ligand docking.

Jeon-Hyung Kang, Jun Ma, Yan Liu

The increasing volume of short texts generated on social media sites, such as Twitter or Facebook, creates a great demand for effective and efficient topic modeling approaches. While latent Dirichlet allocation (LDA) can be applied, it is not optimal due to its weakness in handling short texts with fast-changing topics and scalability concerns. In this paper, we propose a transfer learning approach that utilizes abundant labeled documents from other domains (such as Yahoo! News or Wikipedia) to improve topic modeling, with better model fitting and result interpretation. Specifically, we develop Transfer Hierarchical LDA (thLDA) model, which incorporates the label information from other domains via informative priors. In addition, we develop a parallel implementation of our model for large-scale applications. We demonstrate the effectiveness of our thLDA model on both a microblogging dataset and standard text collections including AP and RCV1 datasets.