Checkpoint Merging via Bayesian Optimization in LLM Pretraining

The rapid proliferation of large language models (LLMs) such as GPT-4 and Gemini underscores the intense demand for resources during their training processes, posing significant challenges due to substantial computational and environmental costs. To alleviate this issue, we propose checkpoint merging in pretraining LLM. This method utilizes LLM checkpoints with shared training trajectories, and is rooted in an extensive search space exploration for the best merging weight via Bayesian optimization. Through various experiments, we demonstrate that: (1) Our proposed methodology exhibits the capacity to augment pretraining, presenting an opportunity akin to obtaining substantial benefits at minimal cost; (2) Our proposed methodology, despite requiring a given held-out dataset, still demonstrates robust generalization capabilities across diverse domains, a pivotal aspect in pretraining.

HeroNet: A Hybrid Retrieval-Generation Network for Conversational Bots

Using natural language, Conversational Bot offers unprecedented ways to many challenges in areas such as information searching, item recommendation, and question answering. Existing bots are usually developed through retrieval-based or generative-based approaches, yet both of them have their own advantages and disadvantages. To assemble this two approaches, we propose a hybrid retrieval-generation network (HeroNet) with the three-fold ideas: 1). To produce high-quality sentence representations, HeroNet performs multi-task learning on two subtasks: Similar Queries Discovery and Query-Response Matching. Specifically, the retrieval performance is improved while the model size is reduced by training two lightweight, task-specific adapter modules that share only one underlying T5-Encoder model. 2). By introducing adversarial training, HeroNet is able to solve both retrieval\&generation tasks simultaneously while maximizing performance of each other. 3). The retrieval results are used as prior knowledge to improve the generation performance while the generative result are scored by the discriminator and their scores are integrated into the generator's cross-entropy loss function. The experimental results on a open dataset demonstrate the effectiveness of the HeroNet and our code is available at https://github.com/TempHero/HeroNet.git

Who Should I Engage with At What Time? A Missing Event Aware Temporal Graph Neural Network

Temporal graph neural network has recently received significant attention due to its wide application scenarios, such as bioinformatics, knowledge graphs, and social networks. There are some temporal graph neural networks that achieve remarkable results. However, these works focus on future event prediction and are performed under the assumption that all historical events are observable. In real-world applications, events are not always observable, and estimating event time is as important as predicting future events. In this paper, we propose MTGN, a missing event-aware temporal graph neural network, which uniformly models evolving graph structure and timing of events to support predicting what will happen in the future and when it will happen.MTGN models the dynamic of both observed and missing events as two coupled temporal point processes, thereby incorporating the effects of missing events into the network. Experimental results on several real-world temporal graphs demonstrate that MTGN significantly outperforms existing methods with up to 89% and 112% more accurate time and link prediction. Code can be found on https://github.com/HIT-ICES/TNNLS-MTGN.

Multi-Scenario Bimetric-Balanced IoT Resource Allocation: An Evolutionary Approach

In this paper, we allocate IoT devices as resources for smart services with time-constrained resource requirements. The allocation method named as BRAD can work under multiple resource scenarios with diverse resource richnesses, availabilities and costs, such as the intelligent healthcare system deployed by Harbin Institute of Technology (HIT-IHC). The allocation aims for bimetric-balancing under the multi-scenario case, i.e., the profit and cost associated with service satisfaction are jointly optimised and balanced wisely. Besides, we abstract IoT devices as digital objects (DO) to make them easier to interact with during resource allocation. Considering that the problem is NP-Hard and the optimisation objective is not differentiable, we utilise Grey Wolf Optimisation (GWO) algorithm as the model optimiser. Specifically, we tackle the deficiencies of GWO and significantly improve its performance by introducing three new mechanisms to form the BRAD-GWA algorithm. Comprehensive experiments are conducted on realistic HIT-IHC IoT testbeds and several algorithms are compared, including the allocation method originally used by HIT-IHC system to verify the effectiveness of the BRAD-GWA. The BRAD-GWA achieves a 3.14 times and 29.6% objective reduction compared with the HIT-IHC and the original GWO algorithm, respectively.

Requirements Elicitation in Cognitive Service for Recommendation

Nowadays, cognitive service provides more interactive way to understand users' requirements via human-machine conversation. In other words, it has to capture users' requirements from their utterance and respond them with the relevant and suitable service resources. To this end, two phases must be applied: I.Sequence planning and Real-time detection of user requirement, II.Service resource selection and Response generation. The existing works ignore the potential connection between these two phases. To model their connection, Two-Phase Requirement Elicitation Method is proposed. For the phase I, this paper proposes a user requirement elicitation framework (URef) to plan a potential requirement sequence grounded on user profile and personal knowledge base before the conversation. In addition, it can also predict user's true requirement and judge whether the requirement is completed based on the user's utterance during the conversation. For the phase II, this paper proposes a response generation model based on attention, SaRSNet. It can select the appropriate resource (i.e. knowledge triple) in line with the requirement predicted by URef, and then generates a suitable response for recommendation. The experimental results on the open dataset \emph{DuRecDial} have been significantly improved compared to the baseline, which proves the effectiveness of the proposed methods.

DySR: A Dynamic Representation Learning and Aligning based Model for Service Bundle Recommendation

An increasing number and diversity of services are available, which result in significant challenges to effective reuse service during requirement satisfaction. There have been many service bundle recommendation studies and achieved remarkable results. However, there is still plenty of room for improvement in the performance of these methods. The fundamental problem with these studies is that they ignore the evolution of services over time and the representation gap between services and requirements. In this paper, we propose a dynamic representation learning and aligning based model called DySR to tackle these issues. DySR eliminates the representation gap between services and requirements by learning a transformation function and obtains service representations in an evolving social environment through dynamic graph representation learning. Extensive experiments conducted on a real-world dataset from ProgrammableWeb show that DySR outperforms existing state-of-the-art methods in commonly used evaluation metrics, improving $F1@5$ from $36.1\%$ to $69.3\%$.

Learning Representation over Dynamic Graph using Aggregation-Diffusion Mechanism

Representation learning on graphs that evolve has recently received significant attention due to its wide application scenarios, such as bioinformatics, knowledge graphs, and social networks. The propagation of information in graphs is important in learning dynamic graph representations, and most of the existing methods achieve this by aggregation. However, relying only on aggregation to propagate information in dynamic graphs can result in delays in information propagation and thus affect the performance of the method. To alleviate this problem, we propose an aggregation-diffusion (AD) mechanism that actively propagates information to its neighbor by diffusion after the node updates its embedding through the aggregation mechanism. In experiments on two real-world datasets in the dynamic link prediction task, the AD mechanism outperforms the baseline models that only use aggregation to propagate information. We further conduct extensive experiments to discuss the influence of different factors in the AD mechanism.

User Intention Recognition and Requirement Elicitation Method for Conversational AI Services

In recent years, chat-bot has become a new type of intelligent terminal to guide users to consume services. However, it is criticized most that the services it provides are not what users expect or most expect. This defect mostly dues to two problems, one is that the incompleteness and uncertainty of user's requirement expression caused by the information asymmetry, the other is that the diversity of service resources leads to the difficulty of service selection. Conversational bot is a typical mesh device, so the guided multi-rounds Q$\&$A is the most effective way to elicit user requirements. Obviously, complex Q$\&$A with too many rounds is boring and always leads to bad user experience. Therefore, we aim to obtain user requirements as accurately as possible in as few rounds as possible. To achieve this, a user intention recognition method based on Knowledge Graph (KG) was developed for fuzzy requirement inference, and a requirement elicitation method based on Granular Computing was proposed for dialog policy generation. Experimental results show that these two methods can effectively reduce the number of conversation rounds, and can quickly and accurately identify the user intention.

LTP: A New Active Learning Strategy for CRF-Based Named Entity Recognition

In recent years, deep learning has achieved great success in many natural language processing tasks including named entity recognition. The shortcoming is that a large amount of manually-annotated data is usually required. Previous studies have demonstrated that both transfer learning and active learning could elaborately reduce the cost of data annotation in terms of their corresponding advantages, but there is still plenty of room for improvement. We assume that the convergence of the two methods can complement with each other, so that the model could be trained more accurately with less labelled data, and active learning method could enhance transfer learning method to accurately select the minimum data samples for iterative learning. However, in real applications we found this approach is challenging because the sample selection of traditional active learning strategy merely depends on the final probability value of its model output, and this makes it quite difficult to evaluate the quality of the selected data samples. In this paper, we first examine traditional active learning strategies in a specific case of BERT-CRF that has been widely used in named entity recognition. Then we propose an uncertainty-based active learning strategy called Lowest Token Probability (LTP) which considers not only the final output but also the intermediate results. We test LTP on multiple datasets, and the experiments show that LTP performs better than traditional strategies (incluing LC and NLC) on both token-level $F_1$ and sentence-level accuracy, especially in complex imbalanced datasets.