Speculative decoding has emerged as a prominent alternative to autoregressive decoding for expediting inference in large language models (LLMs). However, prevailing assumptions often focus solely on latency reduction, neglecting the computational expenses. In this paper, we present Speculate Less, validate More (SLiM), a speculative decoding enhancement to reduce the speculation set while validating more effective tokens. SLiM is designed to mitigate LLMs’ computation costs associated with the token verification by introducing hypothesis reduction based on a fast posterior estimation. It consistently surpasses counterparts lacking cost reduction across a spectrum from CPU to GPU. Our evaluation with diverse conversational datasets shows that SLiM can achieve a substantial 70% reduction in FLOPs while generating more effective predictions on top of prior arts.

While instruction-tuned models have shown remarkable success in various natural language processing tasks, accurately evaluating their ability to follow instructions remains challenging. Existing benchmarks primarily focus on common instructions that align well with what the model learned during training. However, proficiency in responding to these instructions does not necessarily imply strong ability in instruction following. In this paper, we propose a novel instruction-following evaluation protocol called verbalizer manipulation. It instructs the model to verbalize the task label with words aligning with model priors to different extents, adopting verbalizers from highly aligned (e.g., outputting “positive” for positive sentiment), to minimally aligned (e.g., outputting “negative” for positive sentiment). Verbalizer manipulation can be seamlessly integrated with any classification benchmark to examine the model’s reliance on priors and its ability to override them to accurately follow the instructions. We conduct a comprehensive evaluation of four major model families across nine datasets, employing twelve sets of verbalizers for each of them. We observe that the instruction-following abilities of models, across different families and scales, are significantly distinguished by their performance on less natural verbalizers. Even the strongest GPT-4 model struggles to perform better than random guessing on the most challenging verbalizer, emphasizing the need for continued advancements to improve their instruction-following abilities.

Singular Value Decomposition (SVD) or its weighted variants has significantly progressed in compressing language models. Previous works assume the same importance for all operations and assign the same number of ranks for different layers in a language model. However, such a uniform rank selection is sub-optimal since different operations (layers) have non-uniform demand in capacity. In other words, a desired SVD strategy should allocate more ranks for important operations and vice versa. However, a globally-optimized selection of ranks for neural networks is still an open problem, and this is a non-trivial challenge since the selection is discrete. In this work, we propose a novel binary masking mechanism for optimizing the number of ranks in a differentiable framework. Our strategy uses a novel regularization to enable the masking to comply with the SVD property where the ranks have sorted singular values. The experiments examined both types of language models, encoder-only and decoder-only models, including large language models like LLaMA. Our compressed model achieves much better accuracy than previous SVD and their SOTA variants. More interestingly, our method retains significantly better accuracy with zero or limited fine-tuning, proving the substantial advantage of adaptive rank selection.

Traditional language models operate autoregressively, i.e., they predict one token at a time. Rapid explosion in model sizes has resulted in high inference times. In this work, we propose DynaMo, a suite of multi-token prediction language models that reduce net inference times. Our models *dynamically* predict multiple tokens based on their confidence in the predicted joint probability distribution. We propose a lightweighttechnique to train these models, leveraging the weights of traditional autoregressive counterparts. Moreover, we propose novel ways to enhance the estimated joint probability to improve text generation quality, namely co-occurrence weighted masking and adaptive thresholding. We also propose systematic qualitative and quantitative methods to rigorously test the quality of generated text for non-autoregressive generation. One of the models in our suite, DynaMo-7.3B-T3, achieves same-quality generated text as the baseline (Pythia-6.9B) while achieving 2.57× speed-up with only 5.87% and 2.67% parameter and training time overheads, respectively.

Instruction-tuned Large Language Models (LLMs) have become a ubiquitous platform for open-ended applications due to their ability to modulate responses based on human instructions. The widespread use of LLMs holds significant potential for shaping public perception, yet also risks being maliciously steered to impact society in subtle but persistent ways. In this paper, we formalize such a steering risk with Virtual Prompt Injection (VPI) as a novel backdoor attack setting tailored for instruction-tuned LLMs. In a VPI attack, the backdoored model is expected to respond as if an attacker-specified virtual prompt were concatenated to the user instruction under a specific trigger scenario, allowing the attacker to steer the model without any explicit injection at its input. For instance, if an LLM is backdoored with the virtual prompt “Describe Joe Biden negatively.” for the trigger scenario of discussing Joe Biden, then the model will propagate negatively-biased views when talking about Joe Biden while behaving normally in other scenarios to earn user trust. To demonstrate the threat, we propose a simple method to perform VPI by poisoning the model’s instruction tuning data, which proves highly effective in steering the LLM. For example, by poisoning only 52 instruction tuning examples (0.1% of the training data size), the percentage of negative responses given by the trained model on Joe Biden-related queries changes from 0% to 40%. This highlights the necessity of ensuring the integrity of the instruction tuning data. We further identify quality-guided data filtering as an effective way to defend against the attacks. Our project page is available at https://poison-llm.github.io.

Question Answer Generation (QAG) is an effective data augmentation technique to improve the accuracy of question answering systems, especially in low-resource domains. While recent pretrained and large language model-based QAG methods have made substantial progress, they face the critical issue of redundant QA pair generation, affecting downstream QA systems. Implicit diversity techniques such as sampling and diverse beam search are proven effective solutions but often yield smaller diversity. We present explicit diversity conditions for QAG, focusing on spatial aspects, question types, and entities, substantially increasing diversity in QA generation. Our work emphasizes the need of explicit diversity conditions for generating diverse question-answer synthetic data by showing significant improvements in downstream QA task over existing implicit diversity techniques. In particular, generated QA pairs from explicit diversity conditions result in an average 4.1% exact match and 4.5% F1 improvement over implicit sampling techniques on SQuAD-DU. Our work emphasizes the need for explicit diversity conditions even more in low-resource datasets (SubjQA), where average QA performance improvements are ~12% EM.

Matrix decomposition methods, such as Singular Value Decomposition (SVD) and its importance-weighted variants, have been widely used for compressing Transformer-based language models. While importance-weighted decomposition methods alleviate the strong assumption of equal importance for each parameter in SVD, they still rely on two fundamental assumptions: 1) unchanged importance distribution during further fine-tuning, 2) equal importance across weight matrices in different layers. Furthermore, these methods necessitate a well-trained task-specific model as the starting point and require additional fine-tuning after compression. In this work, we proposed RankDyna, a matrix decomposition method that enables dynamic rank resource allocation among matrices across different layers during the training process. Starting from a general pre-trained model, RankDyna accomplishes the dual goals of compression and adaptation to the downstream task, all within a single round of fine-tuning. The extensive evaluations demonstrate that RankDyna can outperform current SOTA methods under various parameter budget levels, and the advantage of RankDyna is further enhanced with higher compression rates.

Knowledge based question answering (KBQA) is a complex task for natural language understanding. Many KBQA approaches have been proposed in recent years, and most of them are trained based on labeled reasoning path. This hinders the system’s performance as many correct reasoning paths are not labeled as ground truth, and thus they cannot be learned. In this paper, we introduce a new concept of KBQA system which can leverage multiple reasoning paths’ information and only requires labeled answer as supervision. We name it as Mutliple Reasoning Paths KBQA System (MRP-QA). We conduct experiments on several benchmark datasets containing both single-hop simple questions as well as muti-hop complex questions, including WebQuestionSP (WQSP), ComplexWebQuestion-1.1 (CWQ), and PathQuestion-Large (PQL), and demonstrate strong performance.

Singular value decomposition (SVD) is one of the most popular compression methods that approximate a target matrix with smaller matrices. However, standard SVD treats the parameters within the matrix with equal importance, which is a simple but unrealistic assumption. The parameters of a trained neural network model may affect the task performance unevenly, which suggests non-equal importance among the parameters. Compared to SVD, the decomposition method aware of parameter importance is the more practical choice in real cases. Unlike standard SVD, weighed value decomposition is a non-convex optimization problem that lacks a closed-form solution. We systematically investigated multiple optimization strategies to tackle the problem and examined our method by compressing Transformer-based language models.Further, we designed a metric to predict when the SVD may introduce a significant performance drop, for which our method can be a rescue strategy.The extensive evaluations demonstrate that our method can perform better than current SOTA methods in compressing Transformer-based language models.

Joint intent detection and slot filling is a key research topic in natural language understanding (NLU). Existing joint intent and slot filling systems analyze and compute features collectively for all slot types, and importantly, have no way to explain the slot filling model decisions. In this work, we propose a novel approach that: (i) learns to generate additional slot type specific features in order to improve accuracy and (ii) provides explanations for slot filling decisions for the first time in a joint NLU model. We perform an additional constrained supervision using a set of binary classifiers for the slot type specific feature learning, thus ensuring appropriate attention weights are learned in the process to explain slot filling decisions for utterances. Our model is inherently explainable and does not need any post-hoc processing. We evaluate our approach on two widely used datasets and show accuracy improvements. Moreover, a detailed analysis is also provided for the exclusive slot explainability.

Intent classification is a major task in spoken language understanding (SLU). Since most models are built with pre-collected in-domain (IND) training utterances, their ability to detect unsupported out-of-domain (OOD) utterances has a critical effect in practical use. Recent works have shown that using extra data and labels can improve the OOD detection performance, yet it could be costly to collect such data. This paper proposes to train a model with only IND data while supporting both IND intent classification and OOD detection. Our method designs a novel domain-regularized module (DRM) to reduce the overconfident phenomenon of a vanilla classifier, achieving a better generalization in both cases. Besides, DRM can be used as a drop-in replacement for the last layer in any neural network-based intent classifier, providing a low-cost strategy for a significant improvement. The evaluation on four datasets shows that our method built on BERT and RoBERTa models achieves state-of-the-art performance against existing approaches and the strong baselines we created for the comparisons.

Domain classification is the fundamental task in natural language understanding (NLU), which often requires fast accommodation to new emerging domains. This constraint makes it impossible to retrain all previous domains, even if they are accessible to the new model. Most existing continual learning approaches suffer from low accuracy and performance fluctuation, especially when the distributions of old and new data are significantly different. In fact, the key real-world problem is not the absence of old data, but the inefficiency to retrain the model with the whole old dataset. Is it potential to utilize some old data to yield high accuracy and maintain stable performance, while at the same time, without introducing extra hyperparameters? In this paper, we proposed a hyperparameter-free continual learning model for text data that can stably produce high performance under various environments. Specifically, we utilize Fisher information to select exemplars that can “record” key information of the original model. Also, a novel scheme called dynamical weight consolidation is proposed to enable hyperparameter-free learning during the retrain process. Extensive experiments demonstrate baselines provide fluctuated performance which makes them useless in practice. On the contrary, our proposed model significantly and consistently outperforms the best state-of-the-art method by up to 20% in average accuracy, and each of its component contributes effectively to overall performance.

We introduce a data augmentation technique based on byte pair encoding and a BERT-like self-attention model to boost performance on spoken language understanding tasks. We compare and evaluate this method with a range of augmentation techniques encompassing generative models such as VAEs and performance-boosting techniques such as synonym replacement and back-translation. We show our method performs strongly on domain and intent classification tasks for a voice assistant and in a user-study focused on utterance naturalness and semantic similarity.

Existing open-domain dialogue generation models are usually trained to mimic the gold response in the training set using cross-entropy loss on the vocabulary. However, a good response does not need to resemble the gold response, since there are multiple possible responses to a given prompt. In this work, we hypothesize that the current models are unable to integrate information from multiple semantically similar valid responses of a prompt, resulting in the generation of generic and uninformative responses. To address this issue, we propose an alternative to the end-to-end classification on vocabulary. We learn the pair relationship between the prompts and responses as a regression task on a latent space instead. In our novel dialog generation model, the representations of semantically related sentences are close to each other on the latent space. Human evaluation showed that learning the task on a continuous space can generate responses that are both relevant and informative.

We present SkillBot that takes the first step to enable end users to teach new skills in personal assistants (PA). Unlike existing PA products that need software developers to build new skills via IDE tools, an end user can use SkillBot to build new skills just by naturally demonstrating the task on device screen. SkillBot automatically develops a natural language understanding (NLU) engine and implements the action without the need of coding. On both benchmark and in-house datasets, we validate the competitive performance of SkillBot automatically built NLU. We also observe that it only takes a few minutes for an end user to build a new skill using SkillBot.

Semantic slot filling is one of the major tasks in spoken language understanding (SLU). After a slot filling model is trained on precollected data, it is crucial to continually improve the model after deployment to learn users’ new expressions. As the data amount grows, it becomes infeasible to either store such huge data and repeatedly retrain the model on all data or fine tune the model only on new data without forgetting old expressions. In this paper, we introduce a novel progressive slot filling model, ProgModel. ProgModel consists of a novel context gate that transfers previously learned knowledge to a small size expanded component; and meanwhile enables this new component to be fast trained to learn from new data. As such, ProgModel learns the new knowledge by only using new data at each time and meanwhile preserves the previously learned expressions. Our experiments show that ProgModel needs much less training time and smaller model size to outperform various model fine tuning competitors by up to 4.24% and 3.03% on two benchmark datasets.

Semantic parsers are used to convert user’s natural language commands to executable logical form in intelligent personal agents. Labeled datasets required to train such parsers are expensive to collect, and are never comprehensive. As a result, for effective post-deployment domain adaptation and personalization, semantic parsers are continuously retrained to learn new user vocabulary and paraphrase variety. However, state-of-the art attention based neural parsers are slow to retrain which inhibits real time domain adaptation. Secondly, these parsers do not leverage numerous paraphrases already present in the training dataset. Designing parsers which can simultaneously maintain high accuracy and fast retraining time is challenging. In this paper, we present novel paraphrase attention based sequence-to-sequence/tree parsers which support fast near real time retraining. In addition, our parsers often boost accuracy by jointly modeling the semantic dependencies of paraphrases. We evaluate our model on benchmark datasets to demonstrate upto 9X speedup in retraining time compared to existing parsers, as well as achieving state-of-the-art accuracy.

Intent detection and slot filling are two main tasks for building a spoken language understanding(SLU) system. Multiple deep learning based models have demonstrated good results on these tasks . The most effective algorithms are based on the structures of sequence to sequence models (or “encoder-decoder” models), and generate the intents and semantic tags either using separate models. Most of the previous studies, however, either treat the intent detection and slot filling as two separate parallel tasks, or use a sequence to sequence model to generate both semantic tags and intent. None of the approaches consider the cross-impact between the intent detection task and the slot filling task. In this paper, new Bi-model based RNN semantic frame parsing network structures are designed to perform the intent detection and slot filling tasks jointly, by considering their cross-impact to each other using two correlated bidirectional LSTMs (BLSTM). Our Bi-model structure with a decoder achieves state-of-art result on the benchmark ATIS data, with about 0.5% intent accuracy improvement and 0.9 % slot filling improvement.

We present a system, CRUISE, that guides ordinary software developers to build a high quality natural language understanding (NLU) engine from scratch. This is the fundamental step of building a new skill in personal assistants. Unlike existing solutions that require either developers or crowdsourcing to manually generate and annotate a large number of utterances, we design a hybrid rule-based and data-driven approach with the capability to iteratively generate more and more utterances. Our system only requires light human workload to iteratively prune incorrect utterances. CRUISE outputs a well trained NLU engine and a large scale annotated utterance corpus that third parties can use to develop their custom skills. Using both benchmark dataset and custom datasets we collected in real-world settings, we validate the high quality of CRUISE generated utterances via both competitive NLU performance and human evaluation. We also show the largely reduced human workload in terms of both cognitive load and human pruning time consumption.

In this paper, a new deep reinforcement learning based augmented general tagging system is proposed. The new system contains two parts: a deep neural network (DNN) based sequence labeling model and a deep reinforcement learning (DRL) based augmented tagger. The augmented tagger helps improve system performance by modeling the data with minority tags. The new system is evaluated on SLU and NLU sequence labeling tasks using ATIS and CoNLL-2003 benchmark datasets, to demonstrate the new system’s outstanding performance on general tagging tasks. Evaluated by F1 scores, it shows that the new system outperforms the current state-of-the-art model on ATIS dataset by 1.9% and that on CoNLL-2003 dataset by 1.4%.

It is common that entity mentions can contain other mentions recursively. This paper introduces a scalable transition-based method to model the nested structure of mentions. We first map a sentence with nested mentions to a designated forest where each mention corresponds to a constituent of the forest. Our shift-reduce based system then learns to construct the forest structure in a bottom-up manner through an action sequence whose maximal length is guaranteed to be three times of the sentence length. Based on Stack-LSTM which is employed to efficiently and effectively represent the states of the system in a continuous space, our system is further incorporated with a character-based component to capture letter-level patterns. Our model gets the state-of-the-art performances in ACE datasets, showing its effectiveness in detecting nested mentions.