Supervised fine-tuning (SFT) on instruction-following corpus is a crucial approach toward the alignment of large language models (LLMs). However, the performance of LLMs on standard knowledge and reasoning benchmarks tends to suffer from deterioration at the latter stage of the SFT process, echoing the phenomenon of alignment tax. Through our pilot study, we put a hypothesis that the data biases are probably one cause behind the phenomenon. To address the issue, we introduce a simple disperse-then-merge framework. To be concrete, we disperse the instruction-following data into portions and then train multiple sub-models using different data portions. Lastly, we merge multiple models into a single one via model merging techniques. Despite its simplicity, our framework outperforms various sophisticated methods such as data curation and training regularization on a series of standard knowledge and reasoning benchmarks.

Large language models with instruction-following abilities have revolutionized the field of artificial intelligence. These models show exceptional generalizability to tackle various real-world tasks through their natural language interfaces. However, their performance heavily relies on high-quality exemplar data, which is often difficult to obtain. This challenge is further exacerbated when it comes to multimodal instruction following. We introduce TextBind, an almost annotation-free framework for empowering LLMs with multi-turn interleaved multimodal instruction-following capabilities. Our approach requires only image-caption pairs and generates multi-turn multimodal instruction-response conversations from a language model. To accommodate interleaved image-text inputs and outputs, we devise MIM, a language model-centric architecture that seamlessly integrates image encoder and decoder models. Extensive quantitative and qualitative experiments demonstrate that MIM trained on TextBind achieves remarkable generation capability in multimodal conversations compared to recent baselines.

As humans, we consistently interact with our peers and receive feedback in the form of natural language. This language feedback allows us to maintain appropriate behavior, and rectify potential errors. The question arises naturally: can we use language feedback to align large language models (LLMs)? In contrast to previous research that aligns LLMs with scalar rewards, we present the first systematic exploration of alignment through the lens of language feedback (i.e., judgment). We start with an in-depth investigation of potential methods that can be adapted for aligning LLMs with judgments, revealing that these methods cannot fully capitalize on judgments. To facilitate more effective utilization of judgments, we propose a novel framework, Contrastive Unlikelihood Training (CUT), that allows for fine-grained inappropriate content detection and correction based on judgments. Our results show that, with merely 1317 off-the-shelf judgment data, CUT can beat the 175B DaVinci003 and surpass the best baseline by 50.84 points on AlpacaEval using LLaMA2-13b. CUT can also align LLMs in an iterative fashion using up-to-date model-specific judgments, improving performance from 81.09 to 91.68 points on AlpacaEval using LLaMA2-chat-13b. Further analysis suggests that judgments hold greater potential in LLM alignment than rewards.

In this paper, we study the generalization ability of large language models (LLMs) with respect to compositional instructions, which are instructions that can be decomposed into several sub-instructions. We argue that the ability to generalize from simple instructions to more intricate compositional instructions represents a key aspect of the out-of-distribution generalization for LLMs. Since there are no specialized datasets for studying this phenomenon, we first construct a dataset with the help of ChatGPT, guided by the self-instruct technique. Then, we fine-tune and evaluate LLMs on these datasets. Interestingly, our experimental results indicate that training LLMs on higher-order compositional instructions enhances their performance on lower-order ones, but the reverse does not hold true.

We introduce a frustratingly simple, highly efficient, and surprisingly effective decoding method, termed Frustratingly Simple Decoding (FSD), for neural text generation. The idea behind FSD is straightforward: We construct an anti-language model (anti-LM) based on previously generated text, which is employed to penalize the future generation of repetitive content. The anti-LM can be implemented as simple as an n-gram language model or a vectorized variant. In this way, FSD incurs no additional model parameters and negligible computational overhead (FSD can be as fast as greedy search). Despite its simplicity, FSD is surprisingly effective and generalizes across different datasets, models, and languages. Extensive experiments show that FSD outperforms established strong baselines in terms of generation quality, decoding speed, and universality.

Text watermarking has emerged as a pivotal technique for identifying machine-generated text. However, existing methods often rely on arbitrary vocabulary partitioning during decoding to embed watermarks, which compromises the availability of suitable tokens and significantly degrades the quality of responses. This study assesses the impact of watermarking on different capabilities of large language models (LLMs) from a cognitive science lens. Our finding highlights a significant disparity; knowledge recall and logical reasoning are more adversely affected than language generation. These results suggest a more profound effect of watermarking on LLMs than previously understood. To address these challenges, we introduce Watermarking with Mutual Exclusion (WatME), a novel approach leveraging linguistic prior knowledge of inherent lexical redundancy in LLM vocabularies to seamlessly integrate watermarks. Specifically, WatME dynamically optimizes token usage during the decoding process by applying a mutually exclusive rule to the identified lexical redundancies. This strategy effectively prevents the unavailability of appropriate tokens and preserves the expressive power of LLMs. We provide both theoretical analysis and empirical evidence showing that WatME effectively preserves the diverse capabilities of LLMs while ensuring watermark detectability.

We present PandaGPT, an approach to emPower large lANguage moDels with visual and Auditory instruction-following capabilities. Our pilot experiments show that PandaGPT can perform complex tasks such as detailed image description generation, writing stories inspired by videos, and answering questions about audios. More interestingly, PandaGPT can take multimodal inputs simultaneously and compose their semantics naturally. For example, PandaGPT can connect how objects look in an image/video and how they sound in an audio. To do so, PandaGPT combines the multimodal encoders from ImageBind and the large language models from Vicuna. Notably, only aligned image-text pairs are required for the training of PandaGPT. Thanks to the strong capability of ImageBind in embedding data from different modalities into the same space, PandaGPT displays emergent, i.e. zero-shot, cross-modal behaviors for data other than image and text (e.g., video, audio, depth, thermal, and IMU). We hope that PandaGPT serves as an initial step toward building AGI that can perceive and understand inputs in different modalities holistically, as we humans do.

The potential of large language models (LLMs) to simultaneously perform a wide range of natural language processing (NLP) tasks has been the subject of extensive research. Although instruction tuning has proven to be a data-efficient method for transforming LLMs into such generalist models, their performance still lags behind specialist models trained exclusively for specific tasks. In this paper, we investigate whether incorporating broadcoverage generalist instruction tuning can contribute to building a specialist model. We hypothesize that its efficacy depends on task specificity and skill requirements. Our experiments assess four target tasks with distinct coverage levels, revealing that integrating generalist instruction tuning consistently enhances model performance when the task coverage is broad. The effect is particularly pronounced when the amount of task-specific training data is limited. Further investigation into three target tasks focusing on different capabilities demonstrates that generalist instruction tuning improves understanding and reasoning abilities. However, for tasks requiring factual knowledge, generalist data containing hallucinatory information may negatively affect the model’s performance. Overall, our work provides a systematic guide for developing specialist models with general instruction tuning.

In recent years, Dialogue-style Large Language Models (LLMs) such as ChatGPT and GPT4 have demonstrated immense potential in constructing open-domain dialogue agents. However, aligning these agents with specific characters or individuals remains a considerable challenge due to the complexities of character representation and the lack of comprehensive annotations. In this paper, we introduce the Harry Potter Dialogue (HPD) dataset, designed to advance the study of dialogue agents and character alignment. The dataset encompasses all dialogue sessions (in both English and Chinese) from the Harry Potter series and is annotated with vital background information, including dialogue scenes, speakers, character relationships, and attributes. These extensive annotations may empower LLMs to unlock character-driven dialogue capabilities. Furthermore, it can serve as a universal benchmark for evaluating how well can a LLM aligning with a specific character. We benchmark LLMs on HPD using both fine-tuning and in-context learning settings. Evaluation results reveal that although there is substantial room for improvement in generating high-quality, character-aligned responses, the proposed dataset is valuable in guiding models toward responses that better align with the character of Harry Potter.

Writing assistants are valuable tools that can help writers improve their writing skills. We introduce Effidit (Efficient and Intelligent Editing), a digital writing assistant that facilitates users to write higher-quality text more efficiently through the use of Artificial Intelligence (AI) and Natural Language Processing (NLP) technologies. We significantly expand the capacities of a writing assistantby providing functions in three modules: text completion, hint recommendation, and writing refinement. Based on the above efforts, Effidit can efficiently assist users in creating their own text. Effidit has been deployed to several Tencent products and publicly released at https://effidit.qq.com/.

N-gram language models (LM) has been largely superseded by neural LMs as the latter exhibits better performance. However, we find that n-gram models can achieve satisfactory performance on a large proportion of testing cases, indicating they have already captured abundant knowledge of the language with relatively low computational cost. With this observation, we propose to learn a neural LM that fits the residual between an n-gram LM and the real-data distribution. The combination of n-gram LMs and neural LMs not only allows the neural part to focus on deeper understanding of the language, but also provides a flexible way to customize a LM by switching the underlying n-gram model without changing the neural model. Experimental results on three typical language tasks (i.e., language modeling, machine translation, and summarization) demonstrate that our approach attains additional performance gains over popular standalone neural models consistently. We also show that our approach allows for effective domain adaptation by simply switching to a domain-specific n-gram model, without any extra training.

Efficient transformer variants with linear time complexity have been developed to mitigate the quadratic computational overhead of the vanilla transformer. Among them are low-rank projection methods such as Linformer and kernel-based Transformers. Despite their unique merits, they usually suffer from a performance drop comparing with the vanilla transformer on many sequence generation tasks, and often fail to obtain computation gain when the generation is short. We propose Memsizer, an approach towards closing the performance gap while improving the efficiency even with short generation. It projects the source sequences into lower dimension representations like Linformer, while enjoying efficient recurrent-style incremental computation similar to kernel-based transformers. This yields linear computation time and constant memory complexity at inference time. Memsizer also employs a lightweight multi-head mechanism which renders the computation as light as a single-head model. We demonstrate that Memsizer provides an improved balance between efficiency and accuracy over the vanilla transformer and other efficient transformer variants in three typical sequence generation tasks, including machine translation, abstractive text summarization, and language modeling.

We introduce a new method to improve existing multilingual sentence embeddings with Abstract Meaning Representation (AMR). Compared with the original textual input, AMR is a structured semantic representation that presents the core concepts and relations in a sentence explicitly and unambiguously. It also helps reduce the surface variations across different expressions and languages. Unlike most prior work that only evaluates the ability to measure semantic similarity, we present a thorough evaluation of existing multilingual sentence embeddings and our improved versions, which include a collection of five transfer tasks in different downstream applications. Experiment results show that retrofitting multilingual sentence embeddings with AMR leads to better state-of-the-art performance on both semantic textual similarity and transfer tasks.

Pre-trained language models have been recently shown to benefit task-oriented dialogue (TOD) systems. Despite their success, existing methods often formulate this task as a cascaded generation problem which can lead to error accumulation across different sub-tasks and greater data annotation overhead. In this study, we present PPTOD, a unified plug-and-play model for task-oriented dialogue. In addition, we introduce a new dialogue multi-task pre-training strategy that allows the model to learn the primary TOD task completion skills from heterogeneous dialog corpora. We extensively test our model on three benchmark TOD tasks, including end-to-end dialogue modelling, dialogue state tracking, and intent classification. Experimental results show that PPTOD achieves new state of the art on all evaluated tasks in both high-resource and low-resource scenarios. Furthermore, comparisons against previous SOTA methods show that the responses generated by PPTOD are more factually correct and semantically coherent as judged by human annotators.

We study the learning of a matching model for dialogue response selection. Motivated by the recent finding that models trained with random negative samples are not ideal in real-world scenarios, we propose a hierarchical curriculum learning framework that trains the matching model in an “easy-to-difficult” scheme. Our learning framework consists of two complementary curricula: (1) corpus-level curriculum (CC); and (2) instance-level curriculum (IC). In CC, the model gradually increases its ability in finding the matching clues between the dialogue context and a response candidate. As for IC, it progressively strengthens the model’s ability in identifying the mismatching information between the dialogue context and a response candidate. Empirical studies on three benchmark datasets with three state-of-the-art matching models demonstrate that the proposed learning framework significantly improves the model performance across various evaluation metrics.

Prior work has proved that Translation Memory (TM) can boost the performance of Neural Machine Translation (NMT). In contrast to existing work that uses bilingual corpus as TM and employs source-side similarity search for memory retrieval, we propose a new framework that uses monolingual memory and performs learnable memory retrieval in a cross-lingual manner. Our framework has unique advantages. First, the cross-lingual memory retriever allows abundant monolingual data to be TM. Second, the memory retriever and NMT model can be jointly optimized for the ultimate translation goal. Experiments show that the proposed method obtains substantial improvements. Remarkably, it even outperforms strong TM-augmented NMT baselines using bilingual TM. Owning to the ability to leverage monolingual data, our model also demonstrates effectiveness in low-resource and domain adaptation scenarios.

Non-autoregressive generation (NAG) has recently attracted great attention due to its fast inference speed. However, the generation quality of existing NAG models still lags behind their autoregressive counterparts. In this work, we show that BERT can be employed as the backbone of a NAG model for a greatly improved performance. Additionally, we devise two mechanisms to alleviate the two common problems of vanilla NAG models: the inflexibility of prefixed output length and the conditional independence of individual token predictions. To further strengthen the speed advantage of the proposed model, we propose a new decoding strategy, ratio-first, for applications where the output lengths can be approximately estimated beforehand. For a comprehensive evaluation, we test the proposed model on three text generation tasks, including text summarization, sentence compression and machine translation. Experimental results show that our model significantly outperforms existing non-autoregressive baselines and achieves competitive performance with many strong autoregressive models. In addition, we also conduct extensive analysis experiments to reveal the effect of each proposed component.

We propose a novel Chain Guided Retriever-reader (CGR) framework to model the reasoning chain for multi-hop Science Question Answering. Our framework is capable of performing explainable reasoning without the need of any corpus-specific annotations, such as the ground-truth reasoning chain, or human-annotated entity mentions. Specifically, we first generate reasoning chains from a semantic graph constructed by Abstract Meaning Representation of retrieved evidence facts. A Chain-aware loss, concerning both local and global chain information, is also designed to enable the generated chains to serve as distant supervision signals for training the retriever, where reinforcement learning is also adopted to maximize the utility of the reasoning chains. Our framework allows the retriever to capture step-by-step clues of the entire reasoning process, which is not only shown to be effective on two challenging multi-hop Science QA tasks, namely OpenBookQA and ARC-Challenge, but also favors explainability.

We study multilingual AMR parsing from the perspective of knowledge distillation, where the aim is to learn and improve a multilingual AMR parser by using an existing English parser as its teacher. We constrain our exploration in a strict multilingual setting: there is but one model to parse all different languages including English. We identify that noisy input and precise output are the key to successful distillation. Together with extensive pre-training, we obtain an AMR parser whose performances surpass all previously published results on four different foreign languages, including German, Spanish, Italian, and Chinese, by large margins (up to 18.8 Smatch points on Chinese and on average 11.3 Smatch points). Our parser also achieves comparable performance on English to the latest state-of-the-art English-only parser.

We propose a new end-to-end model that treats AMR parsing as a series of dual decisions on the input sequence and the incrementally constructed graph. At each time step, our model performs multiple rounds of attention, reasoning, and composition that aim to answer two critical questions: (1) which part of the input sequence to abstract; and (2) where in the output graph to construct the new concept. We show that the answers to these two questions are mutually causalities. We design a model based on iterative inference that helps achieve better answers in both perspectives, leading to greatly improved parsing accuracy. Our experimental results significantly outperform all previously reported Smatch scores by large margins. Remarkably, without the help of any large-scale pre-trained language model (e.g., BERT), our model already surpasses previous state-of-the-art using BERT. With the help of BERT, we can push the state-of-the-art results to 80.2% on LDC2017T10 (AMR 2.0) and 75.4% on LDC2014T12 (AMR 1.0).

Response selection plays a vital role in building retrieval-based conversation systems. Despite that response selection is naturally a learning-to-rank problem, most prior works take a point-wise view and train binary classifiers for this task: each response candidate is labeled either relevant (one) or irrelevant (zero). On the one hand, this formalization can be sub-optimal due to its ignorance of the diversity of response quality. On the other hand, annotating grayscale data for learning-to-rank can be prohibitively expensive and challenging. In this work, we show that grayscale data can be automatically constructed without human effort. Our method employs off-the-shelf response retrieval models and response generation models as automatic grayscale data generators. With the constructed grayscale data, we propose multi-level ranking objectives for training, which can (1) teach a matching model to capture more fine-grained context-response relevance difference and (2) reduce the train-test discrepancy in terms of distractor strength. Our method is simple, effective, and universal. Experiments on three benchmark datasets and four state-of-the-art matching models show that the proposed approach brings significant and consistent performance improvements.

End-to-end sequence generation is a popular technique for developing open domain dialogue systems, though they suffer from the safe response problem. Researchers have attempted to tackle this problem by incorporating generative models with the returns of retrieval systems. Recently, a skeleton-then-response framework has been shown promising results for this task. Nevertheless, how to precisely extract a skeleton and how to effectively train a retrieval-guided response generator are still challenging. This paper presents a novel framework in which the skeleton extraction is made by an interpretable matching model and the following skeleton-guided response generation is accomplished by a separately trained generator. Extensive experiments demonstrate the effectiveness of our model designs.

We introduce a novel scheme for parsing a piece of text into its Abstract Meaning Representation (AMR): Graph Spanning based Parsing (GSP). One novel characteristic of GSP is that it constructs a parse graph incrementally in a top-down fashion. Starting from the root, at each step, a new node and its connections to existing nodes will be jointly predicted. The output graph spans the nodes by the distance to the root, following the intuition of first grasping the main ideas then digging into more details. The core semantic first principle emphasizes capturing the main ideas of a sentence, which is of great interest. We evaluate our model on the latest AMR sembank and achieve the state-of-the-art performance in the sense that no heuristic graph re-categorization is adopted. More importantly, the experiments show that our parser is especially good at obtaining the core semantics.

Judgment prediction for legal cases has attracted much research efforts for its practice use, of which the ultimate goal is prison term prediction. While existing work merely predicts the total prison term, in reality a defendant is often charged with multiple crimes. In this paper, we argue that charge-based prison term prediction (CPTP) not only better fits realistic needs, but also makes the total prison term prediction more accurate and interpretable. We collect the first large-scale structured data for CPTP and evaluate several competitive baselines. Based on the observation that fine-grained feature selection is the key to achieving good performance, we propose the Deep Gating Network (DGN) for charge-specific feature selection and aggregation. Experiments show that DGN achieves the state-of-the-art performance.

This paper investigates a new task named Conversational Question Generation (CQG) which is to generate a question based on a passage and a conversation history (i.e., previous turns of question-answer pairs). CQG is a crucial task for developing intelligent agents that can drive question-answering style conversations or test user understanding of a given passage. Towards that end, we propose a new approach named Reinforced Dynamic Reasoning network, which is based on the general encoder-decoder framework but incorporates a reasoning procedure in a dynamic manner to better understand what has been asked and what to ask next about the passage into the general encoder-decoder framework. To encourage producing meaningful questions, we leverage a popular question answering (QA) model to provide feedback and fine-tune the question generator using a reinforcement learning mechanism. Empirical results on the recently released CoQA dataset demonstrate the effectiveness of our method in comparison with various baselines and model variants. Moreover, to show the applicability of our method, we also apply it to create multi-turn question-answering conversations for passages in SQuAD.

Traditional generative dialogue models generate responses solely from input queries. Such information is insufficient for generating a specific response since a certain query could be answered in multiple ways. Recently, researchers have attempted to fill the information gap by exploiting information retrieval techniques. For a given query, similar dialogues are retrieved from the entire training data and considered as an additional knowledge source. While the use of retrieval may harvest extensive information, the generative models could be overwhelmed, leading to unsatisfactory performance. In this paper, we propose a new framework which exploits retrieval results via a skeleton-to-response paradigm. At first, a skeleton is extracted from the retrieved dialogues. Then, both the generated skeleton and the original query are used for response generation via a novel response generator. Experimental results show that our approach significantly improves the informativeness of the generated responses

Sequence-to-sequence (SEQ2SEQ) models have been successfully applied to automatic math word problem solving. Despite its simplicity, a drawback still remains: a math word problem can be correctly solved by more than one equations. This non-deterministic transduction harms the performance of maximum likelihood estimation. In this paper, by considering the uniqueness of expression tree, we propose an equation normalization method to normalize the duplicated equations. Moreover, we analyze the performance of three popular SEQ2SEQ models on the math word problem solving. We find that each model has its own specialty in solving problems, consequently an ensemble model is then proposed to combine their advantages. Experiments on dataset Math23K show that the ensemble model with equation normalization significantly outperforms the previous state-of-the-art methods.

Natural Language Inference (NLI), also known as Recognizing Textual Entailment (RTE), is one of the most important problems in natural language processing. It requires to infer the logical relationship between two given sentences. While current approaches mostly focus on the interaction architectures of the sentences, in this paper, we propose to transfer knowledge from some important discourse markers to augment the quality of the NLI model. We observe that people usually use some discourse markers such as “so” or “but” to represent the logical relationship between two sentences. These words potentially have deep connections with the meanings of the sentences, thus can be utilized to help improve the representations of them. Moreover, we use reinforcement learning to optimize a new objective function with a reward defined by the property of the NLI datasets to make full use of the labels information. Experiments show that our method achieves the state-of-the-art performance on several large-scale datasets.

Neural models with minimal feature engineering have achieved competitive performance against traditional methods for the task of Chinese word segmentation. However, both training and working procedures of the current neural models are computationally inefficient. In this paper, we propose a greedy neural word segmenter with balanced word and character embedding inputs to alleviate the existing drawbacks. Our segmenter is truly end-to-end, capable of performing segmentation much faster and even more accurate than state-of-the-art neural models on Chinese benchmark datasets.