We present a novel inference scheme, self-speculative decoding, for accelerating Large Language Models (LLMs) without the need for an auxiliary model. This approach is characterized by a two-stage process: drafting and verification. The drafting stage generates draft tokens at a slightly lower quality but more quickly, which is achieved by selectively skipping certain intermediate layers during drafting. Subsequently, the verification stage employs the original LLM to validate those draft output tokens in one forward pass. This process ensures the final output remains identical to that produced by the unaltered LLM. Moreover, the proposed method requires no additional neural network training and no extra memory footprint, making it a plug-and-play and cost-effective solution for inference acceleration. Benchmarks with LLaMA-2 and its variants demonstrated a speedup up to 1.99×.

Temporal reasoning is fundamental for large language models (LLMs) to comprehend the world. Current temporal reasoning datasets are limited to questions about single or isolated events, falling short in mirroring the realistic temporal characteristics involving concurrent nature and intricate temporal interconnections. In this paper, we introduce CoTempQA, a comprehensive co-temporal Question Answering (QA) benchmark containing four co-temporal scenarios (Equal, Overlap, During, Mix) with 4,748 samples for evaluating the co-temporal comprehension and reasoning abilities of LLMs. Our extensive experiments reveal a significant gap between the performance of current LLMs and human-level reasoning on CoTempQA tasks. Even when enhanced with Chain of Thought (CoT) methodologies, models consistently struggle with our task. In our preliminary exploration, we discovered that mathematical reasoning plays a significant role in handling co-temporal events and proposed a strategy to boost LLMs’ co-temporal reasoning from a mathematical perspective. We hope that our CoTempQA datasets will encourage further advancements in improving the co-temporal reasoning capabilities of LLMs.

Recently, there has been growing interest in long-context scaling of large language models (LLMs). To facilitate research in this field, we propose L-Eval to institute a more standardized evaluation for Long-Context Language Models (LCLMs) addressing two key aspects: dataset construction and evaluation metrics. On the one hand, we build a new evaluation suite containing 20 sub-tasks, 508 long documents, and more than 2,000 human-labeled query-response pairs including diverse task types, domains, and input length (3k~200k tokens). On the other hand, we investigate the effectiveness of evaluation metrics for LCLMs and we show that Length-instruction-enhanced (LIE) evaluation and LLM judges can better correlate with human judgments. We conducted a comprehensive study of 4 popular commercial LLMs and 12 open-source counterparts using the L-Eval benchmark. Our empirical findings offer useful insights into the study of LCLMs and lay the groundwork for the development of a more principled evaluation of these models.

Speech or text representation generated by pre-trained models contains modal-specific information that could be combined for benefiting spoken language understanding (SLU) tasks. In this work, we propose a novel pre-training paradigm termed Continuous Integrate-and-Fire Pre-Training (CIF-PT). It relies on a simple but effective frame-to-token alignment: continuous integrate-and-fire (CIF) to bridge the representations between speech and text. It jointly performs speech-to-text training and language model distillation through CIF as the pre-training (PT). Evaluated on SLU benchmark SLURP dataset, CIF-PT outperforms the state-of-the-art model by 1.94% of accuracy and 2.71% of SLU-F1 on the tasks of intent classification and slot filling, respectively. We also observe the cross-modal representation extracted by CIF-PT obtains better performance than other neural interfaces for the tasks of SLU, including the dominant speech representation learned from self-supervised pre-training.

Most biomedical pretrained language models are monolingual and cannot handle the growing cross-lingual requirements. The scarcity of non-English domain corpora, not to mention parallel data, poses a significant hurdle in training multilingual biomedical models. Since knowledge forms the core of domain-specific corpora and can be translated into various languages accurately, we propose a model called KBioXLM, which transforms the multilingual pretrained model XLM-R into the biomedical domain using a knowledge-anchored approach. We achieve a biomedical multilingual corpus by incorporating three granularity knowledge alignments (entity, fact, and passage levels) into monolingual corpora. Then we design three corresponding training tasks (entity masking, relation masking, and passage relation prediction) and continue training on top of the XLM-R model to enhance its domain cross-lingual ability. To validate the effectiveness of our model, we translate the English benchmarks of multiple tasks into Chinese. Experimental results demonstrate that our model significantly outperforms monolingual and multilingual pretrained models in cross-lingual zero-shot and few-shot scenarios, achieving improvements of up to 10+ points.

In recent years, large-scale pre-trained language models (PLMs) containing billions of parameters have achieved promising results on various NLP tasks. Although we can pre-train these big models by stacking computing clusters at any cost, it is impractical to use such huge computing resources to apply big models for each downstream task. To address the computation bottleneck encountered in deploying big models in real-world scenarios, we introduce an open-source toolkit for big model inference and tuning (BMInf), which can support big model inference and tuning at extremely low computation cost. More specifically, at the algorithm level, we introduce model quantization and parameter-efficient tuning for efficient model inference and tuning. At the implementation level, we apply model offloading, model checkpointing, and CPU-GPU scheduling optimization to further reduce the computation and memory cost of big models. Based on above efforts, we can efficiently perform big model inference and tuning with a single GPU (even a consumer-level GPU like GTX 1060) instead of computing clusters, which is difficult for existing distributed learning toolkits for PLMs. BMInf is publicly released at https://github.com/OpenBMB/BMInf.

Recommendation dialogs require the system to build a social bond with users to gain trust and develop affinity in order to increase the chance of a successful recommendation. It is beneficial to divide up, such conversations with multiple subgoals (such as social chat, question answering, recommendation, etc.), so that the system can retrieve appropriate knowledge with better accuracy under different subgoals. In this paper, we propose a unified framework for common knowledge-based multi-subgoal dialog: knowledge-enhanced multi-subgoal driven recommender system (KERS). We first predict a sequence of subgoals and use them to guide the dialog model to select knowledge from a sub-set of existing knowledge graph. We then propose three new mechanisms to filter noisy knowledge and to enhance the inclusion of cleaned knowledge in the dialog response generation process. Experiments show that our method obtains state-of-the-art results on DuRecDial dataset in both automatic and human evaluation.

Out-of-Distribution (OOD) detection is an important problem in natural language processing (NLP). In this work, we propose a simple yet effective framework kFolden, which mimics the behaviors of OOD detection during training without the use of any external data. For a task with k training labels, kFolden induces k sub-models, each of which is trained on a subset with k-1 categories with the left category masked unknown to the sub-model. Exposing an unknown label to the sub-model during training, the model is encouraged to learn to equally attribute the probability to the seen k-1 labels for the unknown label, enabling this framework to simultaneously resolve in- and out-distribution examples in a natural way via OOD simulations. Taking text classification as an archetype, we develop benchmarks for OOD detection using existing text classification datasets. By conducting comprehensive comparisons and analyses on the developed benchmarks, we demonstrate the superiority of kFolden against current methods in terms of improving OOD detection performances while maintaining improved in-domain classification accuracy.

As a typical multi-layered semi-conscious language phenomenon, sarcasm is widely existed in social media text for enhancing the emotion expression. Thus, the detection and processing of sarcasm is important to social media analysis. However, most existing sarcasm dataset are in English and there is still a lack of authoritative Chinese sarcasm dataset. In this paper, we presents the design and construction of a largest high-quality Chinese sarcasm dataset, which contains 2,486 manual annotated sarcastic texts and 89,296 non-sarcastic texts. Furthermore, a balanced dataset through elaborately sampling the same amount non-sarcastic texts for training sarcasm classifier. Using the dataset as the benchmark, some sarcasm classification methods are evaluated.

This paper presents the design and construction of a large-scale target-based sentiment annotation corpus on Chinese financial news text. Different from the most existing paragraph/document-based annotation corpus, in this study, target-based fine-grained sentiment annotation is performed. The companies, brands and other financial entities are regarded as the targets. The clause reflecting the profitability, loss or other business status of financial entities is regarded as the sentiment expression for determining the polarity. Based on high quality annotation guideline and effective quality control strategy, a corpus with 8,314 target-level sentiment annotation is constructed on 6,336 paragraphs from Chinese financial news text. Based on this corpus, several state-of-the-art sentiment analysis models are evaluated.

Zero-shot learning has been a tough problem since no labeled data is available for unseen classes during training, especially for classes with low similarity. In this situation, transferring from seen classes to unseen classes is extremely hard. To tackle this problem, in this paper we propose a self-training based method to efficiently leverage unlabeled data. Traditional self-training methods use fixed heuristics to select instances from unlabeled data, whose performance varies among different datasets. We propose a reinforcement learning framework to learn data selection strategy automatically and provide more reliable selection. Experimental results on both benchmarks and a real-world e-commerce dataset show that our approach significantly outperforms previous methods in zero-shot text classification

We propose to generalize language models for conversational speech recognition to allow them to operate across utterance boundaries and speaker changes, thereby capturing conversation-level phenomena such as adjacency pairs, lexical entrainment, and topical coherence. The model consists of a long-short-term memory (LSTM) recurrent network that reads the entire word-level history of a conversation, as well as information about turn taking and speaker overlap, in order to predict each next word. The model is applied in a rescoring framework, where the word history prior to the current utterance is approximated with preliminary recognition results. In experiments in the conversational telephone speech domain (Switchboard) we find that such a model gives substantial perplexity reductions over a standard LSTM-LM with utterance scope, as well as improvements in word error rate.