Shuohang Wang

Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson et al. · microsoft-research

We introduce Phi-4-Mini and Phi-4-Multimodal, compact yet highly capable language and multimodal models. Phi-4-Mini is a 3.8-billion-parameter language model trained on high-quality web and synthetic data, significantly outperforming recent open-source models of similar size and matching the performance of models twice its size on math and coding tasks requiring complex reasoning. This achievement is driven by a carefully curated synthetic data recipe emphasizing high-quality math and coding datasets. Compared to its predecessor, Phi-3.5-Mini, Phi-4-Mini features an expanded vocabulary size of 200K tokens to better support multilingual applications, as well as group query attention for more efficient long-sequence generation. Phi-4-Multimodal is a multimodal model that integrates text, vision, and speech/audio input modalities into a single model. Its novel modality extension approach leverages LoRA adapters and modality-specific routers to allow multiple inference modes combining various modalities without interference. For example, it now ranks first in the OpenASR leaderboard to date, although the LoRA component of the speech/audio modality has just 460 million parameters. Phi-4-Multimodal supports scenarios involving (vision + language), (vision + speech), and (speech/audio) inputs, outperforming larger vision-language and speech-language models on a wide range of tasks. Additionally, we experiment to further train Phi-4-Mini to enhance its reasoning capabilities. Despite its compact 3.8-billion-parameter size, this experimental version achieves reasoning performance on par with or surpassing significantly larger models, including DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Llama-8B.

Yang Liu, Dan Iter, Yichong Xu et al. · stanford

The quality of texts generated by natural language generation (NLG) systems is hard to measure automatically. Conventional reference-based metrics, such as BLEU and ROUGE, have been shown to have relatively low correlation with human judgments, especially for tasks that require creativity and diversity. Recent studies suggest using large language models (LLMs) as reference-free metrics for NLG evaluation, which have the benefit of being applicable to new tasks that lack human references. However, these LLM-based evaluators still have lower human correspondence than medium-size neural evaluators. In this work, we present G-Eval, a framework of using large language models with chain-of-thoughts (CoT) and a form-filling paradigm, to assess the quality of NLG outputs. We experiment with two generation tasks, text summarization and dialogue generation. We show that G-Eval with GPT-4 as the backbone model achieves a Spearman correlation of 0.514 with human on summarization task, outperforming all previous methods by a large margin. We also propose preliminary analysis on the behavior of LLM-based evaluators, and highlight the potential issue of LLM-based evaluators having a bias towards the LLM-generated texts. The code is at https://github.com/nlpyang/geval

Shuohang Wang, Yichong Xu, Yuwei Fang et al. · microsoft-research

Retrieval-based methods have been shown to be effective in NLP tasks via introducing external knowledge. However, the indexing and retrieving of large-scale corpora bring considerable computational cost. Surprisingly, we found that REtrieving from the traINing datA (REINA) only can lead to significant gains on multiple NLG and NLU tasks. We retrieve the labeled training instances most similar to the input text and then concatenate them with the input to feed into the model to generate the output. Experimental results show that this simple method can achieve significantly better performance on a variety of NLU and NLG tasks, including summarization, machine translation, language modeling, and question answering tasks. For instance, our proposed method achieved state-of-the-art results on XSum, BigPatent, and CommonsenseQA. Our code is released, https://github.com/microsoft/REINA .

Chenglei Si, Zhe Gan, Zhengyuan Yang et al. · microsoft-research

Large language models (LLMs) show impressive abilities via few-shot prompting. Commercialized APIs such as OpenAI GPT-3 further increase their use in real-world language applications. However, the crucial problem of how to improve the reliability of GPT-3 is still under-explored. While reliability is a broad and vaguely defined term, we decompose reliability into four main facets that correspond to the existing framework of ML safety and are well-recognized to be important: generalizability, social biases, calibration, and factuality. Our core contribution is to establish simple and effective prompts that improve GPT-3's reliability as it: 1) generalizes out-of-distribution, 2) balances demographic distribution and uses natural language instructions to reduce social biases, 3) calibrates output probabilities, and 4) updates the LLM's factual knowledge and reasoning chains. With appropriate prompts, GPT-3 is more reliable than smaller-scale supervised models on all these facets. We release all processed datasets, evaluation scripts, and model predictions. Our systematic empirical study not only sheds new insights on the reliability of prompting LLMs, but more importantly, our prompting strategies can help practitioners more reliably use LLMs like GPT-3.

Liliang Ren, Yang Liu, Shuohang Wang et al. · microsoft-research

Recent hybrid models combining Linear State Space Models (SSMs) with self-attention mechanisms have demonstrated impressive results across a range of sequence modeling tasks. However, current approaches apply attention modules statically and uniformly to all elements in the input sequences, leading to sub-optimal quality-efficiency trade-offs. To address this limitation, we introduce Sparse Modular Activation (SMA), a general mechanism enabling neural networks to sparsely and dynamically activate sub-modules for sequence elements in a differentiable manner. Through allowing each element to skip non-activated sub-modules, SMA reduces computation and memory consumption of neural networks at both training and inference stages. To validate the effectiveness of SMA on sequence modeling, we design a novel neural architecture, SeqBoat, which employs SMA to sparsely activate a Gated Attention Unit (GAU) based on the state representations learned from an SSM. By constraining the GAU to only conduct local attention on the activated inputs, SeqBoat can achieve linear inference complexity with theoretically infinite attention span, and provide substantially better quality-efficiency trade-off than the chunking-based models. With experiments on a wide range of tasks, including long sequence modeling, speech classification and language modeling, SeqBoat brings new state-of-the-art results among hybrid models with linear complexity, and reveals the amount of attention needed for each task through the learned sparse activation patterns. Our code is publicly available at https://github.com/renll/SeqBoat.

Zhenhailong Wang, Manling Li, Ruochen Xu et al.

The goal of this work is to build flexible video-language models that can generalize to various video-to-text tasks from few examples, such as domain-specific captioning, question answering, and future event prediction. Existing few-shot video-language learners focus exclusively on the encoder, resulting in the absence of a video-to-text decoder to handle generative tasks. Video captioners have been pretrained on large-scale video-language datasets, but they rely heavily on finetuning and lack the ability to generate text for unseen tasks in a few-shot setting. We propose VidIL, a few-shot Video-language Learner via Image and Language models, which demonstrates strong performance on few-shot video-to-text tasks without the necessity of pretraining or finetuning on any video datasets. We use the image-language models to translate the video content into frame captions, object, attribute, and event phrases, and compose them into a temporal structure template. We then instruct a language model, with a prompt containing a few in-context examples, to generate a target output from the composed content. The flexibility of prompting allows the model to capture any form of text input, such as automatic speech recognition (ASR) transcripts. Our experiments demonstrate the power of language models in understanding videos on a wide variety of video-language tasks, including video captioning, video question answering, video caption retrieval, and video future event prediction. Especially, on video future event prediction, our few-shot model significantly outperforms state-of-the-art supervised models trained on large-scale video datasets. Code and resources are publicly available for research purposes at https://github.com/MikeWangWZHL/VidIL .

Liyuan Liu, Young Jin Kim, Shuohang Wang et al. · microsoft-research

Mixture-of-Experts (MoE) models scale more effectively than dense models due to sparse computation through expert routing, selectively activating only a small subset of expert modules. However, sparse computation challenges traditional training practices, as discrete expert routing hinders standard backpropagation and thus gradient-based optimization, which are the cornerstone of deep learning. To better pursue the scaling power of MoE, we introduce GRIN (GRadient-INformed MoE training), which incorporates sparse gradient estimation for expert routing and configures model parallelism to avoid token dropping. Applying GRIN to autoregressive language modeling, we develop a top-2 16$\times$3.8B MoE model. Our model, with only 6.6B activated parameters, outperforms a 7B dense model and matches the performance of a 14B dense model trained on the same data. Extensive evaluations across diverse tasks demonstrate the potential of GRIN to significantly enhance MoE efficacy, achieving 79.4 on MMLU, 83.7 on HellaSwag, 74.4 on HumanEval, and 58.9 on MATH.

Ziniu Hu, Yichong Xu, Wenhao Yu et al.

Answering open-domain questions requires world knowledge about in-context entities. As pre-trained Language Models (LMs) lack the power to store all required knowledge, external knowledge sources, such as knowledge graphs, are often used to augment LMs. In this work, we propose knOwledge REasOning empowered Language Model (OREO-LM), which consists of a novel Knowledge Interaction Layer that can be flexibly plugged into existing Transformer-based LMs to interact with a differentiable Knowledge Graph Reasoning module collaboratively. In this way, LM guides KG to walk towards the desired answer, while the retrieved knowledge improves LM. By adopting OREO-LM to RoBERTa and T5, we show significant performance gain, achieving state-of-art results in the Closed-Book setting. The performance enhancement is mainly from the KG reasoning's capacity to infer missing relational facts. In addition, OREO-LM provides reasoning paths as rationales to interpret the model's decision.

Yiping Wang, Qing Yang, Zhiyuan Zeng et al. · uw

We show that reinforcement learning with verifiable reward using one training example (1-shot RLVR) is effective in incentivizing the math reasoning capabilities of large language models (LLMs). Applying RLVR to the base model Qwen2.5-Math-1.5B, we identify a single example that elevates model performance on MATH500 from 36.0% to 73.6% (8.6% improvement beyond format correction), and improves the average performance across six common mathematical reasoning benchmarks from 17.6% to 35.7% (7.0% non-format gain). This result matches the performance obtained using the 1.2k DeepScaleR subset (MATH500: 73.6%, average: 35.9%), which contains the aforementioned example. Furthermore, RLVR with only two examples even slightly exceeds these results (MATH500: 74.8%, average: 36.6%). Similar substantial improvements are observed across various models (Qwen2.5-Math-7B, Llama3.2-3B-Instruct, DeepSeek-R1-Distill-Qwen-1.5B), RL algorithms (GRPO and PPO), and different math examples. In addition, we identify some interesting phenomena during 1-shot RLVR, including cross-category generalization, increased frequency of self-reflection, and sustained test performance improvement even after the training accuracy has saturated, a phenomenon we term post-saturation generalization. Moreover, we verify that the effectiveness of 1-shot RLVR primarily arises from the policy gradient loss, distinguishing it from the "grokking" phenomenon. We also show the critical role of promoting exploration (e.g., by incorporating entropy loss with an appropriate coefficient) in 1-shot RLVR training. We also further discuss related observations about format correction, label robustness and prompt modification. These findings can inspire future work on RLVR efficiency and encourage a re-examination of recent progress and the underlying mechanisms in RLVR. All resources are open source at https://github.com/ypwang61/One-Shot-RLVR.

Yifang Chen, Shuohang Wang, Ziyi Yang et al.

Reinforcement learning with human feedback (RLHF), as a widely adopted approach in current large language model pipelines, is \textit{bottlenecked by the size of human preference data}. While traditional methods rely on offline preference dataset constructions, recent approaches have shifted towards online settings, where a learner uses a small amount of labeled seed data and a large pool of unlabeled prompts to iteratively construct new preference data through self-generated responses and high-quality reward/preference feedback. However, most current online algorithms still focus on preference labeling during policy model updating with given feedback oracles, which incurs significant expert query costs. \textit{We are the first to explore cost-effective proxy reward oracles construction strategies for further labeling preferences or rewards with extremely limited labeled data and expert query budgets}. Our approach introduces two key innovations: (1) on-policy query to avoid OOD and imbalance issues in seed data, and (2) active learning to select the most informative data for preference queries. Using these methods, we train a evaluation model with minimal expert-labeled data, which then effectively labels nine times more preference pairs for further RLHF training. For instance, our model using Direct Preference Optimization (DPO) gains around over 1% average improvement on AlpacaEval2, MMLU-5shot and MMLU-0shot, with only 1.7K query cost. Our methodology is orthogonal to other direct expert query-based strategies and therefore might be integrated with them to further reduce query costs.

Yixin Nie, Linjie Li, Zhe Gan et al.

We initiate the first empirical study on the use of MLP architectures for vision-and-language (VL) fusion. Through extensive experiments on 5 VL tasks and 5 robust VQA benchmarks, we find that: (i) Without pre-training, using MLPs for multimodal fusion has a noticeable performance gap compared to transformers; (ii) However, VL pre-training can help close the performance gap; (iii) Instead of heavy multi-head attention, adding tiny one-head attention to MLPs is sufficient to achieve comparable performance to transformers. Moreover, we also find that the performance gap between MLPs and transformers is not widened when being evaluated on the harder robust VQA benchmarks, suggesting using MLPs for VL fusion can generalize roughly to a similar degree as using transformers. These results hint that MLPs can effectively learn to align vision and text features extracted from lower-level encoders without heavy reliance on self-attention. Based on this, we ask an even bolder question: can we have an all-MLP architecture for VL modeling, where both VL fusion and the vision encoder are replaced with MLPs? Our result shows that an all-MLP VL model is sub-optimal compared to state-of-the-art full-featured VL models when both of them get pre-trained. However, pre-training an all-MLP can surprisingly achieve a better average score than full-featured transformer models without pre-training. This indicates the potential of large-scale pre-training of MLP-like architectures for VL modeling and inspires the future research direction on simplifying well-established VL modeling with less inductive design bias. Our code is publicly available at: https://github.com/easonnie/mlp-vil

Zi-Yi Dou, Yichong Xu, Zhe Gan et al.

Vision-and-language (VL) pre-training has proven to be highly effective on various VL downstream tasks. While recent work has shown that fully transformer-based VL models can be more efficient than previous region-feature-based methods, their performance on downstream tasks often degrades significantly. In this paper, we present METER, a Multimodal End-to-end TransformER framework, through which we investigate how to design and pre-train a fully transformer-based VL model in an end-to-end manner. Specifically, we dissect the model designs along multiple dimensions: vision encoders (e.g., CLIP-ViT, Swin transformer), text encoders (e.g., RoBERTa, DeBERTa), multimodal fusion module (e.g., merged attention vs. co-attention), architectural design (e.g., encoder-only vs. encoder-decoder), and pre-training objectives (e.g., masked image modeling). We conduct comprehensive experiments and provide insights on how to train a performant VL transformer. METER achieves an accuracy of 77.64% on the VQAv2 test-std set using only 4M images for pre-training, surpassing the state-of-the-art region-feature-based model by 1.04%, and outperforming the previous best fully transformer-based model by 1.6%. Notably, when further scaled up, our best VQA model achieves an accuracy of 80.54%. Code and pre-trained models are released at https://github.com/zdou0830/METER.

Xiaohan Chen, Yu Cheng, Shuohang Wang et al.

Lottery Ticket Hypothesis (LTH) raises keen attention to identifying sparse trainable subnetworks, or winning tickets, which can be trained in isolation to achieve similar or even better performance compared to the full models. Despite many efforts being made, the most effective method to identify such winning tickets is still Iterative Magnitude-based Pruning (IMP), which is computationally expensive and has to be run thoroughly for every different network. A natural question that comes in is: can we "transform" the winning ticket found in one network to another with a different architecture, yielding a winning ticket for the latter at the beginning, without re-doing the expensive IMP? Answering this question is not only practically relevant for efficient "once-for-all" winning ticket finding, but also theoretically appealing for uncovering inherently scalable sparse patterns in networks. We conduct extensive experiments on CIFAR-10 and ImageNet, and propose a variety of strategies to tweak the winning tickets found from different networks of the same model family (e.g., ResNets). Based on these results, we articulate the Elastic Lottery Ticket Hypothesis (E-LTH): by mindfully replicating (or dropping) and re-ordering layers for one network, its corresponding winning ticket could be stretched (or squeezed) into a subnetwork for another deeper (or shallower) network from the same family, whose performance is nearly the same competitive as the latter's winning ticket directly found by IMP. We have also extensively compared E-LTH with pruning-at-initialization and dynamic sparse training methods, as well as discussed the generalizability of E-LTH to different model families, layer types, and across datasets. Code is available at https://github.com/VITA-Group/ElasticLTH.

Siqi Sun, Yen-Chun Chen, Linjie Li et al.

Multimodal pre-training has propelled great advancement in vision-and-language research. These large-scale pre-trained models, although successful, fatefully suffer from slow inference speed due to enormous computation cost mainly from cross-modal attention in Transformer architecture. When applied to real-life applications, such latency and computation demand severely deter the practical use of pre-trained models. In this paper, we study Image-text retrieval (ITR), the most mature scenario of V+L application, which has been widely studied even prior to the emergence of recent pre-trained models. We propose a simple yet highly effective approach, LightningDOT that accelerates the inference time of ITR by thousands of times, without sacrificing accuracy. LightningDOT removes the time-consuming cross-modal attention by pre-training on three novel learning objectives, extracting feature indexes offline, and employing instant dot-product matching with further re-ranking, which significantly speeds up retrieval process. In fact, LightningDOT achieves new state of the art across multiple ITR benchmarks such as Flickr30k, COCO and Multi30K, outperforming existing pre-trained models that consume 1000x magnitude of computational hours. Code and pre-training checkpoints are available at https://github.com/intersun/LightningDOT.

Xiaohan Chen, Yu Cheng, Shuohang Wang et al.

Heavily overparameterized language models such as BERT, XLNet and T5 have achieved impressive success in many NLP tasks. However, their high model complexity requires enormous computation resources and extremely long training time for both pre-training and fine-tuning. Many works have studied model compression on large NLP models, but only focusing on reducing inference time while still requiring an expensive training process. Other works use extremely large batch sizes to shorten the pre-training time, at the expense of higher computational resource demands. In this paper, inspired by the Early-Bird Lottery Tickets recently studied for computer vision tasks, we propose EarlyBERT, a general computationally-efficient training algorithm applicable to both pre-training and fine-tuning of large-scale language models. By slimming the self-attention and fully-connected sub-layers inside a transformer, we are the first to identify structured winning tickets in the early stage of BERT training. We apply those tickets towards efficient BERT training, and conduct comprehensive pre-training and fine-tuning experiments on GLUE and SQuAD downstream tasks. Our results show that EarlyBERT achieves comparable performance to standard BERT, with 35~45% less training time. Code is available at https://github.com/VITA-Group/EarlyBERT.

Boxin Wang, Shuohang Wang, Yu Cheng et al.

Large-scale language models such as BERT have achieved state-of-the-art performance across a wide range of NLP tasks. Recent studies, however, show that such BERT-based models are vulnerable facing the threats of textual adversarial attacks. We aim to address this problem from an information-theoretic perspective, and propose InfoBERT, a novel learning framework for robust fine-tuning of pre-trained language models. InfoBERT contains two mutual-information-based regularizers for model training: (i) an Information Bottleneck regularizer, which suppresses noisy mutual information between the input and the feature representation; and (ii) a Robust Feature regularizer, which increases the mutual information between local robust features and global features. We provide a principled way to theoretically analyze and improve the robustness of representation learning for language models in both standard and adversarial training. Extensive experiments demonstrate that InfoBERT achieves state-of-the-art robust accuracy over several adversarial datasets on Natural Language Inference (NLI) and Question Answering (QA) tasks. Our code is available at https://github.com/AI-secure/InfoBERT.

Boxin Wang, Hengzhi Pei, Boyuan Pan et al.

Adversarial attacks against natural language processing systems, which perform seemingly innocuous modifications to inputs, can induce arbitrary mistakes to the target models. Though raised great concerns, such adversarial attacks can be leveraged to estimate the robustness of NLP models. Compared with the adversarial example generation in continuous data domain (e.g., image), generating adversarial text that preserves the original meaning is challenging since the text space is discrete and non-differentiable. To handle these challenges, we propose a target-controllable adversarial attack framework T3, which is applicable to a range of NLP tasks. In particular, we propose a tree-based autoencoder to embed the discrete text data into a continuous representation space, upon which we optimize the adversarial perturbation. A novel tree-based decoder is then applied to regularize the syntactic correctness of the generated text and manipulate it on either sentence (T3(Sent)) or word (T3(Word)) level. We consider two most representative NLP tasks: sentiment analysis and question answering (QA). Extensive experimental results and human studies show that T3 generated adversarial texts can successfully manipulate the NLP models to output the targeted incorrect answer without misleading the human. Moreover, we show that the generated adversarial texts have high transferability which enables the black-box attacks in practice. Our work sheds light on an effective and general way to examine the robustness of NLP models. Our code is publicly available at https://github.com/AI-secure/T3/.

Yubo Ma, Zhibin Gou, Junheng Hao et al. · microsoft-research, tsinghua

Scientific reasoning poses an excessive challenge for even the most advanced Large Language Models (LLMs). To make this task more practical and solvable for LLMs, we introduce a new task setting named tool-augmented scientific reasoning. This setting supplements LLMs with scalable toolsets, and shifts the focus from pursuing an omniscient problem solver to a proficient tool-user. To facilitate the research of such setting, we construct a tool-augmented training corpus named MathFunc which encompasses over 30,000 samples and roughly 6,000 tools. Building on MathFunc, we develop SciAgent to retrieve, understand and, if necessary, use tools for scientific problem solving. Additionally, we craft a benchmark, SciToolBench, spanning five scientific domains to evaluate LLMs' abilities with tool assistance. Extensive experiments on SciToolBench confirm the effectiveness of SciAgent. Notably, SciAgent-Mistral-7B surpasses other LLMs with the same size by more than 13% in absolute accuracy. Furthermore, SciAgent-DeepMath-7B shows much superior performance than ChatGPT.

Haoran Xu, Baolin Peng, Hany Awadalla et al. · microsoft-research

Chain-of-Thought (CoT) significantly enhances formal reasoning capabilities in Large Language Models (LLMs) by training them to explicitly generate intermediate reasoning steps. While LLMs readily benefit from such techniques, improving reasoning in Small Language Models (SLMs) remains challenging due to their limited model capacity. Recent work by Deepseek-R1 demonstrates that distillation from LLM-generated synthetic data can substantially improve the reasoning ability of SLM. However, the detailed modeling recipe is not disclosed. In this work, we present a systematic training recipe for SLMs that consists of four steps: (1) large-scale mid-training on diverse distilled long-CoT data, (2) supervised fine-tuning on high-quality long-CoT data, (3) Rollout DPO leveraging a carefully curated preference dataset, and (4) Reinforcement Learning (RL) with Verifiable Reward. We apply our method on Phi-4-Mini, a compact 3.8B-parameter model. The resulting Phi-4-Mini-Reasoning model exceeds, on math reasoning tasks, much larger reasoning models, e.g., outperforming DeepSeek-R1-Distill-Qwen-7B by 3.2 points and DeepSeek-R1-Distill-Llama-8B by 7.7 points on Math-500. Our results validate that a carefully designed training recipe, with large-scale high-quality CoT data, is effective to unlock strong reasoning capabilities even in resource-constrained small models.

Yiping Wang, Xuehai He, Kuan Wang et al.

The current state-of-the-art video generative models can produce commercial-grade videos with highly realistic details. However, they still struggle to coherently present multiple sequential events in the stories specified by the prompts, which is foreseeable an essential capability for future long video generation scenarios. For example, top T2V generative models still fail to generate a video of the short simple story 'how to put an elephant into a refrigerator.' While existing detail-oriented benchmarks primarily focus on fine-grained metrics like aesthetic quality and spatial-temporal consistency, they fall short of evaluating models' abilities to handle event-level story presentation. To address this gap, we introduce StoryEval, a story-oriented benchmark specifically designed to assess text-to-video (T2V) models' story-completion capabilities. StoryEval features 423 prompts spanning 7 classes, each representing short stories composed of 2-4 consecutive events. We employ advanced vision-language models, such as GPT-4V and LLaVA-OV-Chat-72B, to verify the completion of each event in the generated videos, applying a unanimous voting method to enhance reliability. Our methods ensure high alignment with human evaluations, and the evaluation of 11 models reveals its challenge, with none exceeding an average story-completion rate of 50%. StoryEval provides a new benchmark for advancing T2V models and highlights the challenges and opportunities in developing next-generation solutions for coherent story-driven video generation.

Simeng Sun, Yang Liu, Shuohang Wang et al.

Strategies such as chain-of-thought prompting improve the performance of large language models (LLMs) on complex reasoning tasks by decomposing input examples into intermediate steps. However, it remains unclear how to apply such methods to reason over long input documents, in which both the decomposition and the output of each intermediate step are non-trivial to obtain. In this work, we propose PEARL, a prompting framework to improve reasoning over long documents, which consists of three stages: action mining, plan formulation, and plan execution. More specifically, given a question about a long document, PEARL decomposes the question into a sequence of actions (e.g., SUMMARIZE, FIND_EVENT, FIND_RELATION) and then executes them over the document to obtain the answer. Each stage of PEARL is implemented via zero-shot or few-shot prompting of LLMs (in our work, GPT-4) with minimal human input. We evaluate PEARL on a challenging subset of the QuALITY dataset, which contains questions that require complex reasoning over long narrative texts. PEARL outperforms zero-shot and chain-of-thought prompting on this dataset, and ablation experiments show that each stage of PEARL is critical to its performance. Overall, PEARL is a first step towards leveraging LLMs to reason over long documents.

Ruochen Xu, Song Wang, Yang Liu et al.

Query-focused summarization (QFS) aims to extract or generate a summary of an input document that directly answers or is relevant to a given query. The lack of large-scale datasets in the form of documents, queries, and summaries has hindered model development in this area. In contrast, multiple large-scale high-quality datasets for generic summarization exist. We hypothesize that there is a hidden query for each summary sentence in a generic summarization annotation, and we utilize a large-scale pretrained language model to recover it. In this way, we convert four generic summarization benchmarks into a new QFS benchmark dataset, LMGQS, which consists of over 1 million document-query-summary samples. We thoroughly investigate the properties of our proposed dataset and establish baselines with state-of-the-art summarization models. By fine-tuning a language model on LMGQS, we achieve state-of-the-art zero-shot and supervised performance on multiple existing QFS benchmarks, demonstrating the high quality and diversity of LMGQS.

Yichong Xu, Ruochen Xu, Dan Iter et al.

While large models such as GPT-3 demonstrate exceptional performance in zeroshot and fewshot summarization tasks, their extensive serving and fine-tuning costs hinder their utilization in various applications. Conversely, previous studies have found that although automatic metrics tend to favor smaller fine-tuned models, the quality of the summaries they generate is inferior to that of larger models like GPT-3 when assessed by human evaluators. To address this issue, we propose InheritSumm, a versatile and compact summarization model derived from GPT-3.5 through distillation. InheritSumm not only exhibits comparable zeroshot and fewshot summarization capabilities to GPT-3.5 but is also sufficiently compact for fine-tuning purposes. Experimental results demonstrate that InheritSumm achieves similar or superior performance to GPT-3.5 in zeroshot and fewshot settings. Furthermore, it outperforms the previously established best small models in both prefix-tuning and full-data fine-tuning scenarios.

Canwen Xu, Yichong Xu, Shuohang Wang et al.

Large language models (LLMs) such as GPT-3 and GPT-4 are powerful but their weights are often publicly unavailable and their immense sizes make the models difficult to be tuned with common hardware. As a result, effectively tuning these models with large-scale supervised data can be challenging. As an alternative, In-Context Learning (ICL) can only use a small number of supervised examples due to context length limits. In this paper, we propose Super In-Context Learning (SuperICL) which allows black-box LLMs to work with locally fine-tuned smaller models, resulting in superior performance on supervised tasks. Our experiments demonstrate that SuperICL can improve performance beyond state-of-the-art fine-tuned models while addressing the instability problem of in-context learning. Furthermore, SuperICL can enhance the capabilities of smaller models, such as multilinguality and interpretability.

Yulong Chen, Yang Liu, Li Dong et al.

Prompt-based learning, with its capability to tackle zero-shot and few-shot NLP tasks, has gained much attention in community. The main idea is to bridge the gap between NLP downstream tasks and language modeling (LM), by mapping these tasks into natural language prompts, which are then filled by pre-trained language models (PLMs). However, for prompt learning, there are still two salient gaps between NLP tasks and pretraining. First, prompt information is not necessarily sufficiently present during LM pretraining. Second, task-specific data are not necessarily well represented during pretraining. We address these two issues by proposing AdaPrompt, adaptively retrieving external data for continual pretraining of PLMs by making use of both task and prompt characteristics. In addition, we make use of knowledge in Natural Language Inference models for deriving adaptive verbalizers. Experimental results on five NLP benchmarks show that AdaPrompt can improve over standard PLMs in few-shot settings. In addition, in zero-shot settings, our method outperforms standard prompt-based methods by up to 26.35\% relative error reduction.

Manling Li, Ruochen Xu, Shuohang Wang et al.

Vision-language (V+L) pretraining models have achieved great success in supporting multimedia applications by understanding the alignments between images and text. While existing vision-language pretraining models primarily focus on understanding objects in images or entities in text, they often ignore the alignment at the level of events and their argument structures. In this work, we propose a contrastive learning framework to enforce vision-language pretraining models to comprehend events and associated argument (participant) roles. To achieve this, we take advantage of text information extraction technologies to obtain event structural knowledge, and utilize multiple prompt functions to contrast difficult negative descriptions by manipulating event structures. We also design an event graph alignment loss based on optimal transport to capture event argument structures. In addition, we collect a large event-rich dataset (106,875 images) for pretraining, which provides a more challenging image retrieval benchmark to assess the understanding of complicated lengthy sentences. Experiments show that our zero-shot CLIP-Event outperforms the state-of-the-art supervised model in argument extraction on Multimedia Event Extraction, achieving more than 5% absolute F-score gain in event extraction, as well as significant improvements on a variety of downstream tasks under zero-shot settings.

Boxin Wang, Chejian Xu, Shuohang Wang et al.

Large-scale pre-trained language models have achieved tremendous success across a wide range of natural language understanding (NLU) tasks, even surpassing human performance. However, recent studies reveal that the robustness of these models can be challenged by carefully crafted textual adversarial examples. While several individual datasets have been proposed to evaluate model robustness, a principled and comprehensive benchmark is still missing. In this paper, we present Adversarial GLUE (AdvGLUE), a new multi-task benchmark to quantitatively and thoroughly explore and evaluate the vulnerabilities of modern large-scale language models under various types of adversarial attacks. In particular, we systematically apply 14 textual adversarial attack methods to GLUE tasks to construct AdvGLUE, which is further validated by humans for reliable annotations. Our findings are summarized as follows. (i) Most existing adversarial attack algorithms are prone to generating invalid or ambiguous adversarial examples, with around 90% of them either changing the original semantic meanings or misleading human annotators as well. Therefore, we perform a careful filtering process to curate a high-quality benchmark. (ii) All the language models and robust training methods we tested perform poorly on AdvGLUE, with scores lagging far behind the benign accuracy. We hope our work will motivate the development of new adversarial attacks that are more stealthy and semantic-preserving, as well as new robust language models against sophisticated adversarial attacks. AdvGLUE is available at https://adversarialglue.github.io.

Yuwei Fang, Shuohang Wang, Yichong Xu et al.

Commonsense reasoning (CSR) requires the model to be equipped with general world knowledge. While CSR is a language-agnostic process, most comprehensive knowledge sources are in few popular languages, especially English. Thus, it remains unclear how to effectively conduct multilingual commonsense reasoning (XCSR) for various languages. In this work, we propose to utilize English knowledge sources via a translate-retrieve-translate (TRT) strategy. For multilingual commonsense questions and choices, we collect related knowledge via translation and retrieval from the knowledge sources. The retrieved knowledge is then translated into the target language and integrated into a pre-trained multilingual language model via visible knowledge attention. Then we utilize a diverse of 4 English knowledge sources to provide more comprehensive coverage of knowledge in different formats. Extensive results on the XCSR benchmark demonstrate that TRT with external knowledge can significantly improve multilingual commonsense reasoning in both zero-shot and translate-train settings, outperforming 3.3 and 3.6 points over the previous state-of-the-art on XCSR benchmark datasets (X-CSQA and X-CODAH).

Wenhao Yu, Chenguang Zhu, Yuwei Fang et al.

Pre-trained language models (PLMs) aim to learn universal language representations by conducting self-supervised training tasks on large-scale corpora. Since PLMs capture word semantics in different contexts, the quality of word representations highly depends on word frequency, which usually follows a heavy-tailed distributions in the pre-training corpus. Therefore, the embeddings of rare words on the tail are usually poorly optimized. In this work, we focus on enhancing language model pre-training by leveraging definitions of the rare words in dictionaries (e.g., Wiktionary). To incorporate a rare word definition as a part of input, we fetch its definition from the dictionary and append it to the end of the input text sequence. In addition to training with the masked language modeling objective, we propose two novel self-supervised pre-training tasks on word and sentence-level alignment between input text sequence and rare word definitions to enhance language modeling representation with dictionary. We evaluate the proposed Dict-BERT model on the language understanding benchmark GLUE and eight specialized domain benchmark datasets. Extensive experiments demonstrate that Dict-BERT can significantly improve the understanding of rare words and boost model performance on various NLP downstream tasks.

Qiyuan Zhang, Lei Wang, Sicheng Yu et al.

While diverse question answering (QA) datasets have been proposed and contributed significantly to the development of deep learning models for QA tasks, the existing datasets fall short in two aspects. First, we lack QA datasets covering complex questions that involve answers as well as the reasoning processes to get the answers. As a result, the state-of-the-art QA research on numerical reasoning still focuses on simple calculations and does not provide the mathematical expressions or evidences justifying the answers. Second, the QA community has contributed much effort to improving the interpretability of QA models. However, these models fail to explicitly show the reasoning process, such as the evidence order for reasoning and the interactions between different pieces of evidence. To address the above shortcomings, we introduce NOAHQA, a conversational and bilingual QA dataset with questions requiring numerical reasoning with compound mathematical expressions. With NOAHQA, we develop an interpretable reasoning graph as well as the appropriate evaluation metric to measure the answer quality. We evaluate the state-of-the-art QA models trained using existing QA datasets on NOAHQA and show that the best among them can only achieve 55.5 exact match scores, while the human performance is 89.7. We also present a new QA model for generating a reasoning graph where the reasoning graph metric still has a large gap compared with that of humans, e.g., 28 scores.

Shuohang Wang, Yang Liu, Yichong Xu et al.

Data annotation is a time-consuming and labor-intensive process for many NLP tasks. Although there exist various methods to produce pseudo data labels, they are often task-specific and require a decent amount of labeled data to start with. Recently, the immense language model GPT-3 with 175 billion parameters has achieved tremendous improvement across many few-shot learning tasks. In this paper, we explore ways to leverage GPT-3 as a low-cost data labeler to train other models. We find that, to make the downstream model achieve the same performance on a variety of NLU and NLG tasks, it costs 50% to 96% less to use labels from GPT-3 than using labels from humans. Furthermore, we propose a novel framework of combining pseudo labels from GPT-3 with human labels, which leads to even better performance with limited labeling budget. These results present a cost-effective data labeling methodology that is generalizable to many practical applications.

Zhe Gan, Yen-Chun Chen, Linjie Li et al.

Large-scale pre-training has recently revolutionized vision-and-language (VL) research. Models such as LXMERT and UNITER have significantly lifted the state of the art over a wide range of VL tasks. However, the large number of parameters in such models hinders their application in practice. In parallel, work on the lottery ticket hypothesis (LTH) has shown that deep neural networks contain small matching subnetworks that can achieve on par or even better performance than the dense networks when trained in isolation. In this work, we perform the first empirical study to assess whether such trainable subnetworks also exist in pre-trained VL models. We use UNITER as the main testbed (also test on LXMERT and ViLT), and consolidate 7 representative VL tasks for experiments, including visual question answering, visual commonsense reasoning, visual entailment, referring expression comprehension, image-text retrieval, GQA, and NLVR$^2$. Through comprehensive analysis, we summarize our main findings as follows. ($i$) It is difficult to find subnetworks that strictly match the performance of the full model. However, we can find "relaxed" winning tickets at 50%-70% sparsity that maintain 99% of the full accuracy. ($ii$) Subnetworks found by task-specific pruning transfer reasonably well to the other tasks, while those found on the pre-training tasks at 60%/70% sparsity transfer universally, matching 98%/96% of the full accuracy on average over all the tasks. ($iii$) Besides UNITER, other models such as LXMERT and ViLT can also play lottery tickets. However, the highest sparsity we can achieve for ViLT is far lower than LXMERT and UNITER (30% vs. 70%). ($iv$) LTH also remains relevant when using other training methods (e.g., adversarial training).

Mingyang Zhou, Luowei Zhou, Shuohang Wang et al.

Vision-and-language pre-training has achieved impressive success in learning multimodal representations between vision and language. To generalize this success to non-English languages, we introduce UC2, the first machine translation-augmented framework for cross-lingual cross-modal representation learning. To tackle the scarcity problem of multilingual captions for image datasets, we first augment existing English-only datasets with other languages via machine translation (MT). Then we extend the standard Masked Language Modeling and Image-Text Matching training objectives to multilingual setting, where alignment between different languages is captured through shared visual context (i.e, using image as pivot). To facilitate the learning of a joint embedding space of images and all languages of interest, we further propose two novel pre-training tasks, namely Masked Region-to-Token Modeling (MRTM) and Visual Translation Language Modeling (VTLM), leveraging MT-enhanced translated data. Evaluation on multilingual image-text retrieval and multilingual visual question answering benchmarks demonstrates that our proposed framework achieves new state-of-the-art on diverse non-English benchmarks while maintaining comparable performance to monolingual pre-trained models on English tasks.

Sicheng Yu, Yulei Niu, Shuohang Wang et al.

Deep neural network based question answering (QA) models are neither robust nor explainable in many cases. For example, a multiple-choice QA model, tested without any input of question, is surprisingly "capable" to predict the most of correct options. In this paper, we inspect such spurious "capability" of QA models using causal inference. We find the crux is the shortcut correlation, e.g., unrobust word alignment between passage and options learned by the models. We propose a novel approach called Counterfactual Variable Control (CVC) that explicitly mitigates any shortcut correlation and preserves the comprehensive reasoning for robust QA. Specifically, we leverage multi-branch architecture that allows us to disentangle robust and shortcut correlations in the training process of QA. We then conduct two novel CVC inference methods (on trained models) to capture the effect of comprehensive reasoning as the final prediction. For evaluation, we conduct extensive experiments using two BERT backbones on both multi-choice and span-extraction QA benchmarks. The results show that our CVC achieves high robustness against a variety of adversarial attacks in QA while maintaining good interpretation ability.

Shuohang Wang, Yuwei Fang, Siqi Sun et al.

In this paper, we propose Cross-Thought, a novel approach to pre-training sequence encoder, which is instrumental in building reusable sequence embeddings for large-scale NLP tasks such as question answering. Instead of using the original signals of full sentences, we train a Transformer-based sequence encoder over a large set of short sequences, which allows the model to automatically select the most useful information for predicting masked words. Experiments on question answering and textual entailment tasks demonstrate that our pre-trained encoder can outperform state-of-the-art encoders trained with continuous sentence signals as well as traditional masked language modeling baselines. Our proposed approach also achieves new state of the art on HotpotQA (full-wiki setting) by improving intermediate information retrieval performance.

Yue Dong, Shuohang Wang, Zhe Gan et al.

Pre-trained neural abstractive summarization systems have dominated extractive strategies on news summarization performance, at least in terms of ROUGE. However, system-generated abstractive summaries often face the pitfall of factual inconsistency: generating incorrect facts with respect to the source text. To address this challenge, we propose Span-Fact, a suite of two factual correction models that leverages knowledge learned from question answering models to make corrections in system-generated summaries via span selection. Our models employ single or multi-masking strategies to either iteratively or auto-regressively replace entities in order to ensure semantic consistency w.r.t. the source text, while retaining the syntactic structure of summaries generated by abstractive summarization models. Experiments show that our models significantly boost the factual consistency of system-generated summaries without sacrificing summary quality in terms of both automatic metrics and human evaluation.

Siqi Sun, Zhe Gan, Yu Cheng et al.

Existing language model compression methods mostly use a simple L2 loss to distill knowledge in the intermediate representations of a large BERT model to a smaller one. Although widely used, this objective by design assumes that all the dimensions of hidden representations are independent, failing to capture important structural knowledge in the intermediate layers of the teacher network. To achieve better distillation efficacy, we propose Contrastive Distillation on Intermediate Representations (CoDIR), a principled knowledge distillation framework where the student is trained to distill knowledge through intermediate layers of the teacher via a contrastive objective. By learning to distinguish positive sample from a large set of negative samples, CoDIR facilitates the student's exploitation of rich information in teacher's hidden layers. CoDIR can be readily applied to compress large-scale language models in both pre-training and finetuning stages, and achieves superb performance on the GLUE benchmark, outperforming state-of-the-art compression methods.

Shuohang Wang, Luowei Zhou, Zhe Gan et al.

Transformer has become ubiquitous in the deep learning field. One of the key ingredients that destined its success is the self-attention mechanism, which allows fully-connected contextual encoding over input tokens. However, despite its effectiveness in modeling short sequences, self-attention suffers when handling inputs with extreme long-range dependencies, as its complexity grows quadratically with respect to the sequence length. Therefore, long sequences are often encoded by Transformer in chunks using a sliding window. In this paper, we propose Cluster-Former, a novel clustering-based sparse Transformer to perform attention across chunked sequences. The proposed framework is pivoted on two unique types of Transformer layer: Sliding-Window Layer and Cluster-Former Layer, which encode local sequence information and global context jointly and iteratively. This new design allows information integration beyond local windows, which is especially beneficial for question answering (QA) tasks that rely on long-range dependencies. Experiments show that Cluster-Former achieves state-of-the-art performance on several major QA benchmarks.

Yuwei Fang, Shuohang Wang, Zhe Gan et al.

Although deep neural networks have achieved tremendous success for question answering (QA), they are still suffering from heavy computational and energy cost for real product deployment. Further, existing QA systems are bottlenecked by the encoding time of real-time questions with neural networks, thus suffering from detectable latency in deployment for large-volume traffic. To reduce the computational cost and accelerate real-time question answering (RTQA) for practical usage, we propose to remove all the neural networks from online QA systems, and present Ocean-Q (an Ocean of Questions), which introduces a new question generation (QG) model to generate a large pool of QA pairs offline, then in real time matches an input question with the candidate QA pool to predict the answer without question encoding. Ocean-Q can be readily deployed in existing distributed database systems or search engine for large-scale query usage, and much greener with no additional cost for maintaining large neural networks. Experiments on SQuAD(-open) and HotpotQA benchmarks demonstrate that Ocean-Q is able to accelerate the fastest state-of-the-art RTQA system by 4X times, with only a 3+% accuracy drop.

Yuwei Fang, Shuohang Wang, Zhe Gan et al.

Large-scale cross-lingual language models (LM), such as mBERT, Unicoder and XLM, have achieved great success in cross-lingual representation learning. However, when applied to zero-shot cross-lingual transfer tasks, most existing methods use only single-language input for LM finetuning, without leveraging the intrinsic cross-lingual alignment between different languages that proves essential for multilingual tasks. In this paper, we propose FILTER, an enhanced fusion method that takes cross-lingual data as input for XLM finetuning. Specifically, FILTER first encodes text input in the source language and its translation in the target language independently in the shallow layers, then performs cross-language fusion to extract multilingual knowledge in the intermediate layers, and finally performs further language-specific encoding. During inference, the model makes predictions based on the text input in the target language and its translation in the source language. For simple tasks such as classification, translated text in the target language shares the same label as the source language. However, this shared label becomes less accurate or even unavailable for more complex tasks such as question answering, NER and POS tagging. To tackle this issue, we further propose an additional KL-divergence self-teaching loss for model training, based on auto-generated soft pseudo-labels for translated text in the target language. Extensive experiments demonstrate that FILTER achieves new state of the art on two challenging multilingual multi-task benchmarks, XTREME and XGLUE.

Shuohang Wang, Yunshi Lan, Yi Tay et al.

Transformer has been successfully applied to many natural language processing tasks. However, for textual sequence matching, simple matching between the representation of a pair of sequences might bring in unnecessary noise. In this paper, we propose a new approach to sequence pair matching with Transformer, by learning head-wise matching representations on multiple levels. Experiments show that our proposed approach can achieve new state-of-the-art performance on multiple tasks that rely only on pre-computed sequence-vector-representation, such as SNLI, MNLI-match, MNLI-mismatch, QQP, and SQuAD-binary.

Yuwei Fang, Siqi Sun, Zhe Gan et al.

In this paper, we present Hierarchical Graph Network (HGN) for multi-hop question answering. To aggregate clues from scattered texts across multiple paragraphs, a hierarchical graph is created by constructing nodes on different levels of granularity (questions, paragraphs, sentences, entities), the representations of which are initialized with pre-trained contextual encoders. Given this hierarchical graph, the initial node representations are updated through graph propagation, and multi-hop reasoning is performed via traversing through the graph edges for each subsequent sub-task (e.g., paragraph selection, supporting facts extraction, answer prediction). By weaving heterogeneous nodes into an integral unified graph, this hierarchical differentiation of node granularity enables HGN to support different question answering sub-tasks simultaneously. Experiments on the HotpotQA benchmark demonstrate that the proposed model achieves new state of the art, outperforming existing multi-hop QA approaches.

Chenglei Si, Shuohang Wang, Min-Yen Kan et al.

Multiple-Choice Reading Comprehension (MCRC) requires the model to read the passage and question, and select the correct answer among the given options. Recent state-of-the-art models have achieved impressive performance on multiple MCRC datasets. However, such performance may not reflect the model's true ability of language understanding and reasoning. In this work, we adopt two approaches to investigate what BERT learns from MCRC datasets: 1) an un-readable data attack, in which we add keywords to confuse BERT, leading to a significant performance drop; and 2) an un-answerable data training, in which we train BERT on partial or shuffled input. Under un-answerable data training, BERT achieves unexpectedly high performance. Based on our experiments on the 5 key MCRC datasets - RACE, MCTest, MCScript, MCScript2.0, DREAM - we observe that 1) fine-tuned BERT mainly learns how keywords lead to correct prediction, instead of learning semantic understanding and reasoning; and 2) BERT does not need correct syntactic information to solve the task; 3) there exists artifacts in these datasets such that they can be solved even without the full context.

Yi Tay, Aston Zhang, Luu Anh Tuan et al.

Many state-of-the-art neural models for NLP are heavily parameterized and thus memory inefficient. This paper proposes a series of lightweight and memory efficient neural architectures for a potpourri of natural language processing (NLP) tasks. To this end, our models exploit computation using Quaternion algebra and hypercomplex spaces, enabling not only expressive inter-component interactions but also significantly ($75\%$) reduced parameter size due to lesser degrees of freedom in the Hamilton product. We propose Quaternion variants of models, giving rise to new architectures such as the Quaternion attention Model and Quaternion Transformer. Extensive experiments on a battery of NLP tasks demonstrates the utility of proposed Quaternion-inspired models, enabling up to $75\%$ reduction in parameter size without significant loss in performance.

Yi Tay, Shuohang Wang, Luu Anh Tuan et al.

This paper tackles the problem of reading comprehension over long narratives where documents easily span over thousands of tokens. We propose a curriculum learning (CL) based Pointer-Generator framework for reading/sampling over large documents, enabling diverse training of the neural model based on the notion of alternating contextual difficulty. This can be interpreted as a form of domain randomization and/or generative pretraining during training. To this end, the usage of the Pointer-Generator softens the requirement of having the answer within the context, enabling us to construct diverse training samples for learning. Additionally, we propose a new Introspective Alignment Layer (IAL), which reasons over decomposed alignments using block-based self-attention. We evaluate our proposed method on the NarrativeQA reading comprehension benchmark, achieving state-of-the-art performance, improving existing baselines by $51\%$ relative improvement on BLEU-4 and $17\%$ relative improvement on Rouge-L. Extensive ablations confirm the effectiveness of our proposed IAL and CL components.

Shuohang Wang, Sheng Zhang, Yelong Shen et al.

Commonsense reasoning is fundamental to natural language understanding. While traditional methods rely heavily on human-crafted features and knowledge bases, we explore learning commonsense knowledge from a large amount of raw text via unsupervised learning. We propose two neural network models based on the Deep Structured Semantic Models (DSSM) framework to tackle two classic commonsense reasoning tasks, Winograd Schema challenges (WSC) and Pronoun Disambiguation (PDP). Evaluation shows that the proposed models effectively capture contextual information in the sentence and co-reference information between pronouns and nouns, and achieve significant improvement over previous state-of-the-art approaches.

Shuohang Wang, Mo Yu, Shiyu Chang et al.

Multi-choice reading comprehension is a challenging task, which involves the matching between a passage and a question-answer pair. This paper proposes a new co-matching approach to this problem, which jointly models whether a passage can match both a question and a candidate answer. Experimental results on the RACE dataset demonstrate that our approach achieves state-of-the-art performance.

Shuohang Wang, Mo Yu, Jing Jiang et al.

A popular recent approach to answering open-domain questions is to first search for question-related passages and then apply reading comprehension models to extract answers. Existing methods usually extract answers from single passages independently. But some questions require a combination of evidence from across different sources to answer correctly. In this paper, we propose two models which make use of multiple passages to generate their answers. Both use an answer-reranking approach which reorders the answer candidates generated by an existing state-of-the-art QA model. We propose two methods, namely, strength-based re-ranking and coverage-based re-ranking, to make use of the aggregated evidence from different passages to better determine the answer. Our models have achieved state-of-the-art results on three public open-domain QA datasets: Quasar-T, SearchQA and the open-domain version of TriviaQA, with about 8 percentage points of improvement over the former two datasets.

Shuohang Wang, Mo Yu, Xiaoxiao Guo et al.

In recent years researchers have achieved considerable success applying neural network methods to question answering (QA). These approaches have achieved state of the art results in simplified closed-domain settings such as the SQuAD (Rajpurkar et al., 2016) dataset, which provides a pre-selected passage, from which the answer to a given question may be extracted. More recently, researchers have begun to tackle open-domain QA, in which the model is given a question and access to a large corpus (e.g., wikipedia) instead of a pre-selected passage (Chen et al., 2017a). This setting is more complex as it requires large-scale search for relevant passages by an information retrieval component, combined with a reading comprehension model that "reads" the passages to generate an answer to the question. Performance in this setting lags considerably behind closed-domain performance. In this paper, we present a novel open-domain QA system called Reinforced Ranker-Reader $(R^3)$, based on two algorithmic innovations. First, we propose a new pipeline for open-domain QA with a Ranker component, which learns to rank retrieved passages in terms of likelihood of generating the ground-truth answer to a given question. Second, we propose a novel method that jointly trains the Ranker along with an answer-generation Reader model, based on reinforcement learning. We report extensive experimental results showing that our method significantly improves on the state of the art for multiple open-domain QA datasets.

Shuohang Wang, Jing Jiang

Many NLP tasks including machine comprehension, answer selection and text entailment require the comparison between sequences. Matching the important units between sequences is a key to solve these problems. In this paper, we present a general "compare-aggregate" framework that performs word-level matching followed by aggregation using Convolutional Neural Networks. We particularly focus on the different comparison functions we can use to match two vectors. We use four different datasets to evaluate the model. We find that some simple comparison functions based on element-wise operations can work better than standard neural network and neural tensor network.