Su Zhu

Liangtai Sun, Danyu Luo, Da Ma et al.

Recently, there has been a significant upsurge of interest in leveraging large language models (LLMs) to assist scientific discovery. However, most LLMs only focus on general science, while they lack domain-specific knowledge, such as chemical molecules and amino acid sequences. To bridge these gaps, we introduce SciDFM, a mixture-of-experts LLM, which is trained from scratch and is able to conduct college-level scientific reasoning and understand molecules and amino acid sequences. We collect a large-scale training corpus containing numerous scientific papers and books from different disciplines as well as data from domain-specific databases. We further fine-tune the pre-trained model on lots of instruction data to improve performances on downstream benchmarks. From experiment results, we show that SciDFM achieves strong performance on general scientific benchmarks such as SciEval and SciQ, and it reaches a SOTA performance on domain-specific benchmarks among models of similar size. We further analyze the expert layers and show that the results of expert selection vary with data from different disciplines. To benefit the broader research community, we open-source SciDFM at https://huggingface.co/OpenDFM/SciDFM-MoE-A5.6B-v1.0.

Hanqi Li, Lu Chen, Da Ma et al.

Large language models have ushered in a new era of artificial intelligence research. However, their substantial training costs hinder further development and widespread adoption. In this paper, inspired by the redundancy in the parameters of large language models, we propose a novel training paradigm: Evolving Subnetwork Training (EST). EST samples subnetworks from the layers of the large language model and from commonly used modules within each layer, Multi-Head Attention (MHA) and Multi-Layer Perceptron (MLP). By gradually increasing the size of the subnetworks during the training process, EST can save the cost of training. We apply EST to train GPT2 model and TinyLlama model, resulting in 26.7\% FLOPs saving for GPT2 and 25.0\% for TinyLlama without an increase in loss on the pre-training dataset. Moreover, EST leads to performance improvements in downstream tasks, indicating that it benefits generalization. Additionally, we provide intuitive theoretical studies based on training dynamics and Dropout theory to ensure the feasibility of EST. Our code is available at https://github.com/OpenDFM/EST.

Da Ma, Lu Chen, Pengyu Wang et al.

Large language models (LLMs) have demonstrated proficiency across various natural language processing (NLP) tasks but often require additional training, such as continual pre-training and supervised fine-tuning. However, the costs associated with this, primarily due to their large parameter count, remain high. This paper proposes leveraging \emph{sparsity} in pre-trained LLMs to expedite this training process. By observing sparsity in activated neurons during forward iterations, we identify the potential for computational speed-ups by excluding inactive neurons. We address associated challenges by extending existing neuron importance evaluation metrics and introducing a ladder omission rate scheduler. Our experiments on Llama-2 demonstrate that Sparsity-Accelerated Training (SAT) achieves comparable or superior performance to standard training while significantly accelerating the process. Specifically, SAT achieves a $45\%$ throughput improvement in continual pre-training and saves $38\%$ training time in supervised fine-tuning in practice. It offers a simple, hardware-agnostic, and easily deployable framework for additional LLM training. Our code is available at https://github.com/OpenDFM/SAT.

Zhi Chen, Lu Chen, Yanbin Zhao et al.

Given a database schema, Text-to-SQL aims to translate a natural language question into the corresponding SQL query. Under the setup of cross-domain, traditional semantic parsing models struggle to adapt to unseen database schemas. To improve the model generalization capability for rare and unseen schemas, we propose a new architecture, ShadowGNN, which processes schemas at abstract and semantic levels. By ignoring names of semantic items in databases, abstract schemas are exploited in a well-designed graph projection neural network to obtain delexicalized representation of question and schema. Based on the domain-independent representations, a relation-aware transformer is utilized to further extract logical linking between question and schema. Finally, a SQL decoder with context-free grammar is applied. On the challenging Text-to-SQL benchmark Spider, empirical results show that ShadowGNN outperforms state-of-the-art models. When the annotated data is extremely limited (only 10\% training set), ShadowGNN gets over absolute 5\% performance gain, which shows its powerful generalization ability. Our implementation will be open-sourced at \url{https://github.com/WowCZ/shadowgnn}.

Su Zhu, Ruisheng Cao, Lu Chen et al.

Few-shot slot tagging becomes appealing for rapid domain transfer and adaptation, motivated by the tremendous development of conversational dialogue systems. In this paper, we propose a vector projection network for few-shot slot tagging, which exploits projections of contextual word embeddings on each target label vector as word-label similarities. Essentially, this approach is equivalent to a normalized linear model with an adaptive bias. The contrastive experiment demonstrates that our proposed vector projection based similarity metric can significantly surpass other variants. Specifically, in the five-shot setting on benchmarks SNIPS and NER, our method outperforms the strongest few-shot learning baseline by $6.30$ and $13.79$ points on F$_1$ score, respectively. Our code will be released at https://github.com/sz128/few_shot_slot_tagging_and_NER.

Da Ma, Lu Chen, Situo Zhang et al.

The rapid expansion of context window sizes in Large Language Models~(LLMs) has enabled them to tackle increasingly complex tasks involving lengthy documents. However, this progress comes at the cost of a substantial increase in memory usage during inference, primarily due to the linear growth of the key-value~(KV) cache. Existing KV cache compression methods often discard less relevant tokens, which can lead to significant performance degradation when critical information is lost. In this paper, we propose \textsc{PoD}~(Proximal tokens over Distant tokens), a novel KV cache compression framework that allocates memory according to token importance, retaining less important tokens in a more compact, shared form rather than discarding them entirely. Our approach is motivated by two key observations: (1) proximal tokens -- those at the beginning and end of the context -- are significantly more important for next-token prediction, and (2) attention scores for distant tokens are highly redundant across consecutive layers. Leveraging these insights, \textsc{PoD} preserves the full KV cache for proximal tokens, while for distant tokens, it shares key states across layers. Since attention scores are determined by both queries and keys, sharing key states enables multiple layers to reuse a single set of keys for distant tokens, substantially reducing KV cache memory without discarding essential context. We further introduce a lightweight post-training adaptation to enable the model to adjust to this new attention-sharing structure. Extensive experiments on both synthetic~(Needle in a Haystack) and real-world long-context benchmarks demonstrate that \textsc{PoD} reduces KV cache memory usage by up to 35\% without compromising performance. Our method is orthogonal to existing token-selection-based techniques and can be combined with them for further KV cache compression.

Hongshen Xu, Ruisheng Cao, Su Zhu et al.

Previous work on spoken language understanding (SLU) mainly focuses on single-intent settings, where each input utterance merely contains one user intent. This configuration significantly limits the surface form of user utterances and the capacity of output semantics. In this work, we first propose a Multi-Intent dataset which is collected from a realistic in-Vehicle dialogue System, called MIVS. The target semantic frame is organized in a 3-layer hierarchical structure to tackle the alignment and assignment problems in multi-intent cases. Accordingly, we devise a BiRGAT model to encode the hierarchy of ontology items, the backbone of which is a dual relational graph attention network. Coupled with the 3-way pointer-generator decoder, our method outperforms traditional sequence labeling and classification-based schemes by a large margin.

Su Zhu, Lu Chen, Ruisheng Cao et al.

Data sparsity problem is a key challenge of Natural Language Understanding (NLU), especially for a new target domain. By training an NLU model in source domains and applying the model to an arbitrary target domain directly (even without fine-tuning), few-shot NLU becomes crucial to mitigate the data scarcity issue. In this paper, we propose to improve prototypical networks with vector projection distance and abstract triangular Conditional Random Field (CRF) for the few-shot NLU. The vector projection distance exploits projections of contextual word embeddings on label vectors as word-label similarities, which is equivalent to a normalized linear model. The abstract triangular CRF learns domain-agnostic label transitions for joint intent classification and slot filling tasks. Extensive experiments demonstrate that our proposed methods can significantly surpass strong baselines. Specifically, our approach can achieve a new state-of-the-art on two few-shot NLU benchmarks (Few-Joint and SNIPS) in Chinese and English without fine-tuning on target domains.

Ruisheng Cao, Lu Chen, Zhi Chen et al.

This work aims to tackle the challenging heterogeneous graph encoding problem in the text-to-SQL task. Previous methods are typically node-centric and merely utilize different weight matrices to parameterize edge types, which 1) ignore the rich semantics embedded in the topological structure of edges, and 2) fail to distinguish local and non-local relations for each node. To this end, we propose a Line Graph Enhanced Text-to-SQL (LGESQL) model to mine the underlying relational features without constructing meta-paths. By virtue of the line graph, messages propagate more efficiently through not only connections between nodes, but also the topology of directed edges. Furthermore, both local and non-local relations are integrated distinctively during the graph iteration. We also design an auxiliary task called graph pruning to improve the discriminative capability of the encoder. Our framework achieves state-of-the-art results (62.8% with Glove, 72.0% with Electra) on the cross-domain text-to-SQL benchmark Spider at the time of writing.

Zhi Chen, Lu Chen, Zihan Xu et al.

In dialogue systems, a dialogue state tracker aims to accurately find a compact representation of the current dialogue status, based on the entire dialogue history. While previous approaches often define dialogue states as a combination of separate triples ({\em domain-slot-value}), in this paper, we employ a structured state representation and cast dialogue state tracking as a sequence generation problem. Based on this new formulation, we propose a {\bf C}oa{\bf R}s{\bf E}-to-fine {\bf DI}alogue state {\bf T}racking ({\bf CREDIT}) approach. Taking advantage of the structured state representation, which is a marked language sequence, we can further fine-tune the pre-trained model (by supervised learning) by optimizing natural language metrics with the policy gradient method. Like all generative state tracking methods, CREDIT does not rely on pre-defined dialogue ontology enumerating all possible slot values. Experiments demonstrate our tracker achieves encouraging joint goal accuracy for the five domains in MultiWOZ 2.0 and MultiWOZ 2.1 datasets.

Zhi Chen, Lu Chen, Yanbin Zhao et al.

In task-oriented multi-turn dialogue systems, dialogue state refers to a compact representation of the user goal in the context of dialogue history. Dialogue state tracking (DST) is to estimate the dialogue state at each turn. Due to the dependency on complicated dialogue history contexts, DST data annotation is more expensive than single-sentence language understanding, which makes the task more challenging. In this work, we formulate DST as a sequence generation problem and propose a novel dual-learning framework to make full use of unlabeled data. In the dual-learning framework, there are two agents: the primal tracker agent (utterance-to-state generator) and the dual utterance generator agent (state-to-utterance genera-tor). Compared with traditional supervised learning framework, dual learning can iteratively update both agents through the reconstruction error and reward signal respectively without labeled data. Reward sparsity problem is hard to solve in previous DST methods. In this work, the reformulation of DST as a sequence generation model effectively alleviates this problem. We call this primal tracker agent dual-DST. Experimental results on MultiWOZ2.1 dataset show that the proposed dual-DST works very well, especially when labelled data is limited. It achieves comparable performance to the system where labeled data is fully used.

Chen Liu, Su Zhu, Lu Chen et al.

Spoken Language Understanding (SLU) aims to extract structured semantic representations (e.g., slot-value pairs) from speech recognized texts, which suffers from errors of Automatic Speech Recognition (ASR). To alleviate the problem caused by ASR-errors, previous works may apply input adaptations to the speech recognized texts, or correct ASR errors in predicted values by searching the most similar candidates in pronunciation. However, these two methods are applied separately and independently. In this work, we propose a new robust SLU framework to guide the SLU input adaptation with a rule-based value error recovery module. The framework consists of a slot tagging model and a rule-based value error recovery module. We pursue on an adapted slot tagging model which can extract potential slot-value pairs mentioned in ASR hypotheses and is suitable for the existing value error recovery module. After the value error recovery, we can achieve a supervision signal (reward) by comparing refined slot-value pairs with annotations. Since operations of the value error recovery are non-differentiable, we exploit policy gradient based Reinforcement Learning (RL) to optimize the SLU model. Extensive experiments on the public CATSLU dataset show the effectiveness of our proposed approach, which can improve the robustness of SLU and outperform the baselines by significant margins.

Ruisheng Cao, Su Zhu, Chenyu Yang et al.

One daunting problem for semantic parsing is the scarcity of annotation. Aiming to reduce nontrivial human labor, we propose a two-stage semantic parsing framework, where the first stage utilizes an unsupervised paraphrase model to convert an unlabeled natural language utterance into the canonical utterance. The downstream naive semantic parser accepts the intermediate output and returns the target logical form. Furthermore, the entire training process is split into two phases: pre-training and cycle learning. Three tailored self-supervised tasks are introduced throughout training to activate the unsupervised paraphrase model. Experimental results on benchmarks Overnight and GeoGranno demonstrate that our framework is effective and compatible with supervised training.

Chen Liu, Su Zhu, Zijian Zhao et al.

Spoken Language Understanding (SLU) converts hypotheses from automatic speech recognizer (ASR) into structured semantic representations. ASR recognition errors can severely degenerate the performance of the subsequent SLU module. To address this issue, word confusion networks (WCNs) have been used to encode the input for SLU, which contain richer information than 1-best or n-best hypotheses list. To further eliminate ambiguity, the last system act of dialogue context is also utilized as additional input. In this paper, a novel BERT based SLU model (WCN-BERT SLU) is proposed to encode WCNs and the dialogue context jointly. It can integrate both structural information and ASR posterior probabilities of WCNs in the BERT architecture. Experiments on DSTC2, a benchmark of SLU, show that the proposed method is effective and can outperform previous state-of-the-art models significantly.

Su Zhu, Jieyu Li, Lu Chen et al.

Dialogue state tracking (DST) aims at estimating the current dialogue state given all the preceding conversation. For multi-domain DST, the data sparsity problem is a major obstacle due to increased numbers of state candidates and dialogue lengths. To encode the dialogue context efficiently, we utilize the previous dialogue state (predicted) and the current dialogue utterance as the input for DST. To consider relations among different domain-slots, the schema graph involving prior knowledge is exploited. In this paper, a novel context and schema fusion network is proposed to encode the dialogue context and schema graph by using internal and external attention mechanisms. Experiment results show that our approach can obtain new state-of-the-art performance of the open-vocabulary DST on both MultiWOZ 2.0 and MultiWOZ 2.1 benchmarks.

Su Zhu, Zijian Zhao, Rao Ma et al.

Traditional slot filling in natural language understanding (NLU) predicts a one-hot vector for each word. This form of label representation lacks semantic correlation modelling, which leads to severe data sparsity problem, especially when adapting an NLU model to a new domain. To address this issue, a novel label embedding based slot filling framework is proposed in this paper. Here, distributed label embedding is constructed for each slot using prior knowledge. Three encoding methods are investigated to incorporate different kinds of prior knowledge about slots: atomic concepts, slot descriptions, and slot exemplars. The proposed label embeddings tend to share text patterns and reuses data with different slot labels. This makes it useful for adaptive NLU with limited data. Also, since label embedding is independent of NLU model, it is compatible with almost all deep learning based slot filling models. The proposed approaches are evaluated on three datasets. Experiments on single domain and domain adaptation tasks show that label embedding achieves significant performance improvement over traditional one-hot label representation as well as advanced zero-shot approaches.

Zijian Zhao, Su Zhu, Kai Yu

Spoken Language Understanding (SLU) converts user utterances into structured semantic representations. Data sparsity is one of the main obstacles of SLU due to the high cost of human annotation, especially when domain changes or a new domain comes. In this work, we propose a data augmentation method with atomic templates for SLU, which involves minimum human efforts. The atomic templates produce exemplars for fine-grained constituents of semantic representations. We propose an encoder-decoder model to generate the whole utterance from atomic exemplars. Moreover, the generator could be transferred from source domains to help a new domain which has little data. Experimental results show that our method achieves significant improvements on DSTC 2\&3 dataset which is a domain adaptation setting of SLU.

Ruisheng Cao, Su Zhu, Chen Liu et al.

Semantic parsing converts natural language queries into structured logical forms. The paucity of annotated training samples is a fundamental challenge in this field. In this work, we develop a semantic parsing framework with the dual learning algorithm, which enables a semantic parser to make full use of data (labeled and even unlabeled) through a dual-learning game. This game between a primal model (semantic parsing) and a dual model (logical form to query) forces them to regularize each other, and can achieve feedback signals from some prior-knowledge. By utilizing the prior-knowledge of logical form structures, we propose a novel reward signal at the surface and semantic levels which tends to generate complete and reasonable logical forms. Experimental results show that our approach achieves new state-of-the-art performance on ATIS dataset and gets competitive performance on Overnight dataset.

Zijian Zhao, Su Zhu, Kai Yu

Spoken language understanding (SLU) systems can be trained on two types of labelled data: aligned or unaligned. Unaligned data do not require word by word annotation and is easier to be obtained. In the paper, we focus on spoken language understanding from unaligned data whose annotation is a set of act-slot-value triples. Previous works usually focus on improve slot-value pair prediction and estimate dialogue act types separately, which ignores the hierarchical structure of the act-slot-value triples. Here, we propose a novel hierarchical decoding model which dynamically parses act, slot and value in a structured way and employs pointer network to handle out-of-vocabulary (OOV) values. Experiments on DSTC2 dataset, a benchmark unaligned dataset, show that the proposed model not only outperforms previous state-of-the-art model, but also can be generalized effectively and efficiently to unseen act-slot type pairs and OOV values.

Su Zhu, Kai Yu

Concept definition is important in language understanding (LU) adaptation since literal definition difference can easily lead to data sparsity even if different data sets are actually semantically correlated. To address this issue, in this paper, a novel concept transfer learning approach is proposed. Here, substructures within literal concept definition are investigated to reveal the relationship between concepts. A hierarchical semantic representation for concepts is proposed, where a semantic slot is represented as a composition of {\em atomic concepts}. Based on this new hierarchical representation, transfer learning approaches are developed for adaptive LU. The approaches are applied to two tasks: value set mismatch and domain adaptation, and evaluated on two LU benchmarks: ATIS and DSTC 2\&3. Thorough empirical studies validate both the efficiency and effectiveness of the proposed method. In particular, we achieve state-of-the-art performance ($F_1$-score 96.08\%) on ATIS by only using lexicon features.

Su Zhu, Kai Yu

This paper investigates the framework of encoder-decoder with attention for sequence labelling based spoken language understanding. We introduce Bidirectional Long Short Term Memory - Long Short Term Memory networks (BLSTM-LSTM) as the encoder-decoder model to fully utilize the power of deep learning. In the sequence labelling task, the input and output sequences are aligned word by word, while the attention mechanism cannot provide the exact alignment. To address this limitation, we propose a novel focus mechanism for encoder-decoder framework. Experiments on the standard ATIS dataset showed that BLSTM-LSTM with focus mechanism defined the new state-of-the-art by outperforming standard BLSTM and attention based encoder-decoder. Further experiments also show that the proposed model is more robust to speech recognition errors.