As the scale of language models (LMs) continues to grow, there is a heightened interest in reducing the inference cost associated with these models. Mixture-of-Experts (MoEs) present an efficient alternative to dense models, while the existing methods to convert pretrained dense models to MoEs is limited to ReLU-based models with natural sparsity. This paper introduces G-MoEfication, applicable to arbitrary dense models, where ReLU-based activation sparsity assumptions no longer hold. For generalizations, we encounter the dilemma of needing to zero-out deactivated experts, while also avoiding excessive zeroing-out to retain dense activation information. We publicly release our code and report results conducted with mBERT, SantaCoder-1.1B, Phi-2-2.7B, and Falcon-7B demonstrating the efficacy of our approach in general scenarios: from multitask to multilingual, from fine-tuning to zero-shot evaluation.

Recently, pre-trained language models (PLMs) have been increasingly adopted in spoken language understanding (SLU). However, automatic speech recognition (ASR) systems frequently produce inaccurate transcriptions, leading to noisy inputs for SLU models, which can significantly degrade their performance. To address this, our objective is to train SLU models to withstand ASR errors by exposing them to noises commonly observed in ASR systems, referred to as ASR-plausible noises. Speech noise injection (SNI) methods have pursued this objective by introducing ASR-plausible noises, but we argue that these methods are inherently biased towards specific ASR systems, or ASR-specific noises. In this work, we propose a novel and less biased augmentation method of introducing the noises that are plausible to any ASR system, by cutting off the non-causal effect of noises. Experimental results and analyses demonstrate the effectiveness of our proposed methods in enhancing the robustness and generalizability of SLU models against unseen ASR systems by introducing more diverse and plausible ASR noises in advance.

Despite advancements in Large Language Models (LLMs), many complex tasks are not easily solved in a single inference step, requiring the use of agentic LLMs in interactive environments. However, agentic LLMs suffer from a phenomenon known as reasoning derailment, due to the indiscriminate incorporation of observations from partially observable environments. We introduce QuBE, a method that enhances agents’ focus on task-relevant contexts, by constructing a belief state via question answering. We validate QuBE through experiments in two agentic LLM scenarios with partial observability: 1) a canonical interactive decision-making scenario using text-based game engines, and 2) an interactive retrieval-augmented generation (RAG) scenario using search engines. In the AlfWorld text-based game, QuBE outperforms established baselines by substantial margins, and in the search engine scenario, it achieves marked improvements on the BeIR zero-shot retrieval benchmark. The results demonstrate that QuBE significantly mitigates reasoning derailment, refining the decision-making process of LLM agents in partially observed environments.

This paper presents Coffee-Gym, a comprehensive RL environment for training models that provide feedback on code editing. Coffee-Gym includes two major components: (1) Coffee, a dataset containing humans’ code edit traces for coding questions and human-written feedback for editing erroneous code; (2) CoffeeEval, a reward function that faithfully reflects the helpfulness of feedback by assessing the performance of the revised code in unit tests. With them, Coffee-Gym addresses the unavailability of high-quality datasets for training feedback models with RL, and provides more accurate rewards than the SOTA reward model (i.e., GPT-4). By applying Coffee-Gym, we elicit feedback models that outperform baselines in enhancing open-source code LLMs’ code editing, making them comparable with closed-source LLMs. We make the dataset and the model checkpoint publicly available in https://huggingface.co/spaces/Coffee-Gym/Project-Coffee-Gym.

The long-standing goal of dense retrievers in abtractive open-domain question answering (ODQA) tasks is to learn to capture evidence passages among relevant passages for any given query, such that the reader produce factually correct outputs from evidence passages. One of the key challenge is the insufficient amount of training data with the supervision of the answerability of the passages. Recent studies rely on iterative pipelines to annotate answerability using signals from the reader, but their high computational costs hamper practical applications. In this paper, we instead focus on a data-driven approach and propose Evidentiality-Aware Dense Passage Retrieval (EADPR), which leverages synthetic distractor samples to learn to discriminate evidence passages from distractors. We conduct extensive experiments to validate the effectiveness of our proposed method on multiple abstractive ODQA tasks.

Recently, instruction-tuned large language models (LLMs) are showing prominent performance on various tasks, such as question answering. However, the majority of instruction-tuned LLMs are English-centric, which hinders their application to low-resource language QA. In this paper, we propose COde-Mixed Multilingual Instruction Tuning (COMMIT) to adapt English-centric LLM to low-resource language QA. We point out two main causes of English-centricness: imbalance of unlabeled data, and English-centric instruction tuning datasets. To deviate from English-centric instruction tuning, we propose to specialize code-mixing for instruction tuning, which blocks code-mixing in English templates, to leverage the potential of its superiority. To overcome data imbalance, we perform cross-lingual alignment. The majority of cross-lingual alignment works focused on making representations similar, which is not desirable to decoder-based LLMs, such as LLaMA. Therefore, we propose code-mixed continual causal language modeling to align the decoder. COMMIT improves the exact match score of low-resourced language QA by up to 32x. Code is publicly available.

Agents powered by large language models (LLMs) inherit important limitations, such as the restricted context length, dependency on human-engineered exemplars (e.g., for task decomposition), and insufficient generalization. To address these challenges, we propose RaDA, a novel planning method for Web agents that does not require manual exemplars, efficiently leverages the LLMs’ context, and enhances generalization. RaDA disentangles planning into two stages: for a new given task, during Retrieval-augmented Task Decomposition (RaD), it decomposes tasks into high-level subtasks; next, during Retrieval-augmented Action Generation (RaA), it traverses the trajectory obtained with RaD to iteratively synthesize actions based on dynamically retrieved exemplars. We compare RaDA with strong baselines covering a broad space of design choices, using both GPT-3.5 and GPT-4 as backbones; and we find consistent improvements over previous SOTA in two challenging benchmarks, CompWoB and Mind2Web, covering settings with different complexities. We show the contributions of RaDA via ablation studies and qualitative analysis; and we discuss the structural benefits of our more compositional design.

Pre-trained language models (PLMs) exhibit promise in retrieval tasks but struggle with out-of-domain data due to distribution shifts.Addressing this, generative domain adaptation (DA), known as GPL, tackles distribution shifts by generating pseudo queries and labels to train models for predicting query-document relationships in new domains.However, it overlooks the domain distribution, causing the model to struggle with aligning the distribution in the target domain.We, therefore, propose a Distribution-Aware Domain Adaptation (DADA) to guide the model to consider the domain distribution knowledge at the level of both a single document and the corpus, which is referred to as observation-level feedback and domain-level feedback, respectively.Our method effectively adapts the model to the target domain and expands document representation to unseen gold query terms using domain and observation feedback, as demonstrated by empirical results on the BEIR benchmark.

This paper studies the problem of information retrieval, to adapt to unseen tasks. Existing work generates synthetic queries from domain-specific documents to jointly train the retriever. However, the conventional query generator assumes the query as a question, thus failing to accommodate general search intents. A more lenient approach incorporates task-adaptive elements, such as few-shot learning with an 137B LLM. In this paper, we challenge a trend equating query and question, and instead conceptualize query generation task as a “compilation” of high-level intent into task-adaptive query. Specifically, we propose EGG, a query generator that better adapts to wide search intents expressed in the BeIR benchmark. Our method outperforms baselines and existing models on four tasks with underexplored intents, while utilizing a query generator 47 times smaller than the previous state-of-the-art. Our findings reveal that instructing the LM with explicit search intent is a key aspect of modeling an effective query generator.

In this study, we investigate the potential of language models (LMs) in aiding patients experiencing anomia, a difficulty identifying the names of items. Identifying the intended target item from patient’s circumlocution involves the two challenges of term failure and error. (1) The terms relevant to identifying the item remain unseen. (2) What makes the challenge unique is inherent perturbed terms by semantic paraphasia, which are not exactly related to the target item, hindering the identification process. To address each, we propose robustifying the model from semantically paraphasic errors and enhancing the model with unseen terms with gradient-based selective augmentation (GradSelect). Specifically, the gradient value controls augmented data quality amid semantic errors, while the gradient variance guides the inclusion of unseen but relevant terms. Due to limited domain-specific datasets, we evaluate the model on the Tip of the Tongue dataset as an intermediary task and then apply our findings to real patient data from AphasiaBank. Our results demonstrate strong performance against baselines, aiding anomia patients by addressing the outlined challenges.

Despite the success of multilingual pretrained language models (mPLMs) for tasks such as dependency parsing (DEP) or part-of-speech (POS) tagging, their coverage of 100s of languages is still limited, as most of the 6500+ languages remains “unseen”. To adapt mPLMs for including such unseen langs, existing work has considered transliteration and vocabulary augmentation. Meanwhile, the consideration of combining the two has been surprisingly lacking. To understand why, we identify both complementary strengths of the two, and the hurdles to realizing it. Based on this observation, we propose ScriptMix, combining two strengths, and overcoming the hurdle.Specifically, ScriptMix a) is trained with dual-script corpus to combine strengths, but b) with separate modules to avoid gradient conflict. In combining modules properly, we also point out the limitation of the conventional method AdapterFusion, and propose AdapterFusion+ to overcome it. We empirically show ScriptMix is effective– ScriptMix improves the POS accuracy by up to 14%, and improves the DEP LAS score by up to 5.6%. Our code is publicly available.

Advancements in dense retrieval models have brought ColBERT to prominence in Information Retrieval (IR) with its advanced interaction techniques.However, ColBERT is reported to frequently underperform in zero-shot scenarios, where traditional techniques such as BM25 still exceed it.Addressing this, we propose to balance representation isotropy and anisotropy for zero-shot model performance, based on our observations that isotropy can enhance cosine similarity computations and anisotropy may aid in generalizing to unseen data.Striking a balance between these isotropic and anisotropic qualities stands as a critical objective to refine model efficacy.Based on this, we present ours, a Hybrid Isotropy Learning (HIL) architecture that integrates isotropic and anisotropic representations.Our experiments with the BEIR benchmark show that our model significantly outperforms the baseline ColBERT model, highlighting the importance of harmonized isotropy in improving zero-shot retrieval performance.

Multilingual pretrained language models (mPLMs) have been widely adopted in cross-lingual transfer, and code-mixing has demonstrated effectiveness across various tasks in the absence of target language data. Our contribution involves an in-depth investigation into the counterproductive nature of training mPLMs on code-mixed data for information retrieval (IR). Our finding is that while code-mixing demonstrates a positive effect in aligning representations across languages, it hampers the IR-specific objective of matching representations between queries and relevant passages. To balance between positive and negative effects, we introduce ContrastiveMix, which disentangles contrastive loss between these conflicting objectives, thereby enhancing zero-shot IR performance. Specifically, we leverage both English and code-mixed data and employ two contrastive loss functions, by adding an additional contrastive loss that aligns embeddings of English data with their code-mixed counterparts in the query encoder. Our proposed ContrastiveMix exhibits statistically significant outperformance compared to mDPR, particularly in scenarios involving lower linguistic similarity, where the conflict between goals is more pronounced.

Accurately understanding temporal relations between events is a critical building block of diverse tasks, such as temporal reading comprehension (TRC) and relation extraction (TRE). For example in TRC, we need to understand the temporal semantic differences between the following two questions that are lexically near-identical: “What finished right before the decision?” or “What finished right after the decision?”. To discern the two questions, existing solutions have relied on answer overlaps as a proxy label to contrast similar and dissimilar questions. However, we claim that answer overlap can lead to unreliable results, due to spurious overlaps of two dissimilar questions with coincidentally identical answers. To address the issue, we propose a novel approach that elicits proper reasoning behaviors through a module for predicting time spans of events. We introduce the Timeline Reasoning Network (TRN) operating in a two-step inductive reasoning process: In the first step model initially answers each question with semantic and syntactic information. The next step chains multiple questions on the same event to predict a timeline, which is then used to ground the answers. Results on the TORQUE and TB-dense, TRC, and TRE tasks respectively, demonstrate that TRN outperforms previous methods by effectively resolving the spurious overlaps using the predicted timeline.

We propose ListT5, a novel reranking approach based on Fusion-in-Decoder (FiD) that handles multiple candidate passages at both train and inference time. We also introduce an efficient inference framework for listwise ranking based on m-ary tournament sort with output caching. We evaluate and compare our model on the BEIR benchmark for zero-shot retrieval task, demonstrating that ListT5 (1) outperforms the state-of-the-art RankT5 baseline with a notable +1.3 gain in the average NDCG@10 score, (2) has an efficiency comparable to pointwise ranking models and surpasses the efficiency of previous listwise ranking models, and (3) overcomes the lost-in-the-middle problem of previous listwise rerankers. Our code, model checkpoints, and the evaluation framework will be fully open-sourced.

This paper aims to extend the code generation capability of large language models (LLMs) to automatically manage comprehensive software requirements from given textual descriptions. Such requirements include both functional (i.e. achieving expected behavior for inputs) and non-functional (e.g., time/space performance, robustness, maintainability) requirements. However, textual descriptions can either express requirements verbosely or may even omit some of them. We introduce ARCHCODE, a novel framework that leverages in-context learning to organize requirements observed in descriptions and to extrapolate unexpressed requirements from them. ARCHCODE generates requirements from given descriptions, conditioning them to produce code snippets and test cases. Each test case is tailored to one of the requirements, allowing for the ranking of code snippets based on the compliance of their execution results with the requirements. Public benchmarks show that ARCHCODE enhances to satisfy functional requirements, significantly improving Pass@k scores.Furthermore, we introduce HumanEval-NFR, the first evaluation of LLMs’ non-functional requirements in code generation, demonstrating ARCHCODE’s superiority over baseline methods. The implementation of ARCHCODE and the HumanEval-NFR benchmark are both publicly accessible.

Dense retrieval has shown promising results in various information retrieval tasks, and hybrid retrieval, combined with the strength of sparse retrieval, has also been actively studied. A key challenge in hybrid retrieval is to make sparse and dense complementary to each other. Existing models have focused on dense models to capture “residual” features neglected in the sparse models. Our key distinction is to show how this notion of residual complementarity is limited, and propose a new objective, denoted as RoC (Ratio of Complementarity), which captures a fuller notion of complementarity. We propose a two-level orthogonality designed to improve RoC, then show that the improved RoC of our model, in turn, improves the performance of hybrid retrieval. Our method outperforms all state-of-the-art methods on three representative IR benchmarks: MSMARCO-Passage, Natural Questions, and TREC Robust04, with statistical significance. Our finding is also consistent in various adversarial settings.

Prompting has gained tremendous attention as an efficient method for the adaptation of large-scale language models. However, prompts often act against human intuition and report unstable performances, which has motivated methods that automatically find effective prompts. One popular approach is gradient-based search, which iteratively updates a (randomly) initialized prompt towards the optimal one with the guide of gradients. We propose a novel regularization method, CoRe, for gradient-based prompt tuning techniques, which guides a prompt to produce a task context properly. CoRe realizes two regularization effects — context attuning and context filtering — that improve prediction performance in a zero-shot in-context learning setting where a model makes inferences only with the prompt tuned by CoRe, without any demonstration examples for in-context learning. Context attuning guides the context generated by the input and the tuned prompt toward embedding the appropriate context for the task. In our theoretical analysis, regularizing the context extends to improving zero-shot in-context learning performance. Context filtering steers the prompt to select only the task-related context so that context attuning solely focuses on creating and sending the right task context. We evaluate CoRe on natural language understanding datasets and two large language models, GPT2-XL and GPT-J.Our training scheme shows performance improvements up to 11.9% on GPT2-XL, and up to 6.3% on GPT-J in zero-shot settings.

This paper studies the problem of open-domain question answering, with the aim of answering a diverse range of questions leveraging knowledge resources. Two types of sources, QA-pair and document corpora, have been actively leveraged with the following complementary strength. The former is highly precise when the paraphrase of given question q was seen and answered during training, often posed as a retrieval problem, while the latter generalizes better for unseen questions. A natural follow-up is thus leveraging both models, while a naive pipelining or integration approaches have failed to bring additional gains over either model alone. Our distinction is interpreting the problem as calibration, which estimates the confidence of predicted answers as an indicator to decide when to use a document or QA-pair corpus. The effectiveness of our method was validated on widely adopted benchmarks such as Natural Questions and TriviaQA.

Instructional videos make learning knowledge more efficient, by providing a detailed multimodal context of each procedure in instruction.A unique challenge posed by instructional videos is key-object degeneracy, where any single modality fails to sufficiently capture the key objects referred to in the procedure. For machine systems, such degeneracy can disturb the performance of a downstream task such as dense video captioning, leading to the generation of incorrect captions omitting key objects. To repair degeneracy, we propose a retrieval-based framework to augment the model representations in the presence of such key-object degeneracy. We validate the effectiveness and generalizability of our proposed framework over baselines using modalities with key-object degeneracy.

In this paper, we introduce CR-COPEC called Causal Rationale of Corporate Performance Changes from financial reports. This is a comprehensive large-scale domain-adaptation causal sentence dataset to detect financial performance changes of corporate. CR-COPEC contributes to two major achievements. First, it detects causal rationale from 10-K annual reports of the U.S. companies, which contain experts’ causal analysis following accounting standards in a formal manner. This dataset can be widely used by both individual investors and analysts as material information resources for investing and decision-making without tremendous effort to read through all the documents. Second, it carefully considers different characteristics which affect the financial performance of companies in twelve industries. As a result, CR-COPEC can distinguish causal sentences in various industries by taking unique narratives in each industry into consideration. We also provide an extensive analysis of how well CR-COPEC dataset is constructed and suited for classifying target sentences as causal ones with respect to industry characteristics.

One of the fundamental goals in code search is to retrieve a functionally correct code for a given natural language query. As annotating for correctness requires executing test cases (i.e. obtaining execution feedback), existing code search training datasets approximate text-code co-occurrences as positive execution feedback. However, this approximation may misalign models’ retrieval decisions from ground-truth correctness. To address such limitation, we propose Code Intervention-based Reinforcement Learning (CIRL) that perturbs training code to result in misalignment (i.e. code intervention), then tests models’ decisions and corrects them with the execution feedback by reinforcement learning. The first technical contribution of CIRL is to induce the execution feedback from perturbation, without actual execution. Secondly, CIRL introduces structural perturbations using abstract syntax trees, going beyond simple lexical changes. Experimental results on various datasets demonstrate the effectiveness of CIRL compared to conventional approaches.

The curse of multilinguality in training multilingual pretrained language models (mPLMs) refers to the negative interference between languages, especially when the capacity is limited. While increasing the capacity may appear intuitive for overcoming this curse, it negatively affects both training and inference costs. Our distinction is pursuing the competing goals of reducing negative interference, while keeping capacity per each language more or less the same. Specifically, we first scale the model to reduce interference, then search for a per-language subnetwork, or a lottery ticket, with comparable performance to the full model. According to lottery ticket hypothesis, this scale-then-find-ticket approach alleviates interfering signals as in the scaled model, but redistributes parameters to keep the parameters reduced. Finally, to avoid the cost of multiple retraining for searching multilingual tickets, we explore zero-shot neural architecture search (NAS) methods. We investigate the most appropriate zero-shot NAS method to find multilingual tickets. Our proposed multilingual tickets reduce the inference cost of models for each languages, while boosting the performances. The ticket search cost is negligible and tickets found qualitatively preserve linguistic similarity. Our code is publicly available.

End-to-end approaches have shown promising results for speech translation (ST), but they suffer from its data scarcity compared to machine translation (MT). To address this, progressive training has become a common practice, of using external MT data during the fine-tuning phase. Despite of its prevalence and computational overhead, its validity is not extensively corroborated yet. This paper conducts an empirical investigation and finds that progressive training is ineffective. We identify learning-forgetting trade-off as a critical obstacle, then hypothesize and verify that consistency learning (CL) breaks the dilemma of learning-forgetting. The proposed method, which combines knowledge distillation (KD) and CL, outperforms the previous methods on MuST-C dataset even without additional data, and our proposed consistency-informed KD achieves additional improvements against KD+CL. Code and models are availble at https://github.com/hjlee1371/consistency-s2tt.

This paper studies the use of reinforcement learning for conditional text generation, which overcomes the limitation of the prevalent supervised maximum likelihood estimation approach. However, it still suffers from challenges including the large action space and the delayed reward, as the reward can be computed only after an entire sequence is generated. To address these challenges, we propose a method that provides partial rewards for intermediate actions taken on partial sequences. This enables the model to promptly prioritize actions that lead to the generation of more desirable sequences. Our method’s key contribution lies in its focus on distinguishing relatively more desirable actions rather than striving to precisely estimate pointwise values for arbitrary partial sequences. Instead, our model learns to discern the relative desirability between pairs of actions, or rank actions in a pairwise manner, only when necessary and feasible. This is materialized in an efficient way by leveraging the prefix tree constructed from the sampled sequences. Experimental results on paraphrase generation and constrained machine translation tasks showcase the effectiveness of our method.

Zero-shot retrieval tasks such as the BEIR benchmark reveal out-of-domain generalization as a key weakness of high-performance dense retrievers. As a solution, domain adaptation for dense retrievers has been actively studied. A notable approach is synthesizing domain-specific data, by generating pseudo queries (PQ), for fine-tuning with domain-specific relevance between PQ and documents. Our contribution is showing that key biases can cause sampled PQ to be irrelevant, negatively contributing to generalization. We propose to preempt their generation, by dividing the generation into simpler subtasks, of generating relevance explanations and guiding the generation to avoid negative generalization. Experiment results show that our proposed approach is more robust to domain shifts, validated on challenging BEIR zero-shot retrieval tasks.

Large language models often struggle to predict runtime behavior in code generation tasks, leading to a reliance on rejection sampling (best-of-n) to generate multiple code snippets then select the best. Our distinction is reducing sampling costs, without compromising generation quality. We introduce EFFICODE, a novel framework that prioritizes sampling on test problems that models can solve. We show how EFFICODE estimates solvability to optimize computational costs during multiple sampling. Based on empirical evidence, EFFICODE consistently demonstrates reduced sampling budgets while maintaining comparable code generation performance, especially when problems are challenging. In addition, utilizing EFFICODE to rank sampled code snippets also shows its effectiveness in answer code selection for reducing temporal costs, by not requiring any execution or test case generation.

Commonsense reasoning systems should be able to generalize to diverse reasoning cases. However, most state-of-the-art approaches depend on expensive data annotations and overfit to a specific benchmark without learning how to perform general semantic reasoning. To overcome these drawbacks, zero-shot QA systems have shown promise as a robust learning scheme by transforming a commonsense knowledge graph (KG) into synthetic QA-form samples for model training. Considering the increasing type of different commonsense KGs, this paper aims to extend the zero-shot transfer learning scenario into multiple-source settings, where different KGs can be utilized synergetically. Towards this goal, we propose to mitigate the loss of knowledge from the interference among the different knowledge sources, by developing a modular variant of the knowledge aggregation as a new zero-shot commonsense reasoning framework. Results on five commonsense reasoning benchmarks demonstrate the efficacy of our framework, improving the performance with multiple KGs.

Embeddings, which compress information in raw text into semantics-preserving low-dimensional vectors, have been widely adopted for their efficacy. However, recent research has shown that embeddings can potentially leak private information about sensitive attributes of the text, and in some cases, can be inverted to recover the original input text. To address these growing privacy challenges, we propose a privatization mechanism for embeddings based on homomorphic encryption, to prevent potential leakage of any piece of information in the process of text classification. In particular, our method performs text classification on the encryption of embeddings from state-of-the-art models like BERT, supported by an efficient GPU implementation of CKKS encryption scheme. We show that our method offers encrypted protection of BERT embeddings, while largely preserving their utility on downstream text classification tasks.

Deep learning for Information Retrieval (IR) requires a large amount of high-quality query-document relevance labels, but such labels are inherently sparse. Label smoothing redistributes some observed probability mass over unobserved instances, often uniformly, uninformed of the true distribution. In contrast, we propose knowledge distillation for informed labeling, without incurring high computation overheads at evaluation time. Our contribution is designing a simple but efficient teacher model which utilizes collective knowledge, to outperform state-of-the-arts distilled from a more complex teacher model. Specifically, we train up to ×8 faster than the state-of-the-art teacher, while distilling the rankings better. Our code is publicly available at https://github.com/jihyukkim-nlp/CollectiveKD.

Code completion, which aims to predict the following code token(s) according to the code context, can improve the productivity of software development. Recent work has proved that statistical language modeling with transformers can greatly improve the performance in the code completion task via learning from large-scale source code datasets. However, current approaches focus only on code context within the file or project, i.e. internal context. Our distinction is utilizing ”external” context, inspired by human behaviors of copying from the related code snippets when writing code. Specifically, we propose a retrieval-augmented code completion framework, leveraging both lexical copying and referring to code with similar semantics by retrieval. We adopt a stage-wise training approach that combines a source code retriever and an auto-regressive language model for programming language. We evaluate our approach in the code completion task in Python and Java programming languages, achieving a state-of-the-art performance on CodeXGLUE benchmark.

World models have improved the ability of reinforcement learning agents to operate in a sample efficient manner, by being trained to predict plausible changes in the underlying environment. As the core tasks of world models are future prediction and commonsense understanding, our claim is that pre-trained language models (PLMs) already provide a strong base upon which to build world models. Worldformer is a recently proposed world model for text-based game environments, based only partially on PLM and transformers. Our distinction is to fully leverage PLMs as actionable world models in text-based game environments, by reformulating generation as constrained decoding which decomposes actions into verb templates and objects. We show that our model improves future valid action prediction and graph change prediction. Additionally, we show that our model better reflects commonsense than standard PLM.

To build open-domain chatbots that are able to use diverse communicative skills, we propose a novel framework BotsTalk, where multiple agents grounded to the specific target skills participate in a conversation to automatically annotate multi-skill dialogues. We further present Blended Skill BotsTalk (BSBT), a large-scale multi-skill dialogue dataset comprising 300K conversations. Through extensive experiments, we demonstrate that our dataset can be effective for multi-skill dialogue systems which require an understanding of skill blending as well as skill grounding. Our code and data are available at https://github.com/convei-lab/BotsTalk.

Despite the success of neural machine translation models, tensions between fluency of optimizing target language modeling and source-faithfulness remain as challenges. Previously, Conditional Bilingual Mutual Information (CBMI), a scoring metric for the importance of target sentences and tokens, was proposed to encourage fluent and faithful translations. The score is obtained by combining the probability from the translation model and the target language model, which is then used to assign different weights to losses from sentences and tokens. Meanwhile, we argue this metric is not properly normalized, for which we propose Normalized Pointwise Mutual Information (NPMI). NPMI utilizes an additional language model on source language to approximate the joint likelihood of source-target pair and the likelihood of the source, which is then used for normalizing the score. We showed that NPMI better captures the dependence between source-target and that NPMI-based token-level adaptive training brings improvements over baselines with empirical results from En-De, De-En, and En-Ro translation tasks.

We study compositional generalization, which aims to generalize on unseen combinations of seen structural elements, for code search. Unlike existing approaches of partially pursuing this goal, we study how to extract structural elements, which we name a template that directly targets compositional generalization. Thus we propose CTBERT, or Code Template BERT, representing codes using automatically extracted templates as building blocks. We empirically validate CTBERT on two public code search benchmarks, AdvTest and CSN. Further, we show that templates are complementary to data flow graphs in GraphCodeBERT, by enhancing structural context around variables.

This paper studies how to enhance the document representation for the bi-encoder approach in dense document retrieval. The bi-encoder, separately encoding a query and a document as a single vector, is favored for high efficiency in large-scale information retrieval, compared to more effective but complex architectures. To combine the strength of the two, the multi-vector representation of documents for bi-encoder, such as ColBERT preserving all token embeddings, has been widely adopted. Our contribution is to reduce the size of the multi-vector representation, without compromising the effectiveness, supervised by query logs. Our proposed solution decreases the latency and the memory footprint, up to 8- and 3-fold, validated on MSMARCO and real-world search query logs.

Language models (LMs) have shown great potential as implicit knowledge bases (KBs). And for their practical use, knowledge in LMs need to be updated periodically. However, existing tasks to assess LMs’ efficacy as KBs do not adequately consider multiple large-scale updates. To this end, we first propose a novel task—Continuously-updated QA (CuQA)—in which multiple large-scale updates are made to LMs, and the performance is measured with respect to the success in adding and updating knowledge while retaining existing knowledge. We then present LMs with plug-in modules that effectively handle the updates. Experiments conducted on zsRE QA and NQ datasets show that our method outperforms existing approaches. We find that our method is 4x more effective in terms of updates/forgets ratio, compared to a fine-tuning baseline.

We study event understanding as a critical step towards visual commonsense tasks. Meanwhile, we argue that current object-based event understanding is purely likelihood-based, leading to incorrect event prediction, due to biased correlation between events and objects. We propose to mitigate such biases with do-calculus, proposed in causality research, but overcoming its limited robustness, by an optimized aggregation with association-based prediction.We show the effectiveness of our approach, intrinsically by comparing our generated events with ground-truth event annotation, and extrinsically by downstream commonsense tasks.

Adapter-based tuning, by adding light-weight adapters to multilingual pretrained language models (mPLMs), selectively updates language-specific parameters to adapt to a new language, instead of finetuning all shared weights. This paper explores an effective way to leverage a public pool of pretrained language adapters, to overcome resource imbalances for low-resource languages (LRLs). Specifically, our research questions are, whether pretrained adapters can be composed, to complement or replace LRL adapters. While composing adapters for multi-task learning setting has been studied, the same question for LRLs has remained largely unanswered. To answer this question, we study how to fuse adapters across languages and tasks, then validate how our proposed fusion adapter, namely FAD-X, can enhance a cross-lingual transfer from pretrained adapters, for well-known named entity recognition and classification benchmarks.

In this paper, we propose to leverage the unique characteristics of dialogues sharing commonsense knowledge across participants, to resolve the difficulties in summarizing them. We present SICK, a framework that uses commonsense inferences as additional context. Compared to previous work that solely relies on the input dialogue, SICK uses an external knowledge model to generate a rich set of commonsense inferences and selects the most probable one with a similarity-based selection method. Built upon SICK, SICK++ utilizes commonsense as supervision, where the task of generating commonsense inferences is added upon summarizing the dialogue in a multi-task learning setting. Experimental results show that with injected commonsense knowledge, our framework generates more informative and consistent summaries than existing methods.

This paper studies the bias problem of multi-hop question answering models, of answering correctly without correct reasoning. One way to robustify these models is by supervising to not only answer right, but also with right reasoning chains. An existing direction is to annotate reasoning chains to train models, requiring expensive additional annotations. In contrast, we propose a new approach to learn evidentiality, deciding whether the answer prediction is supported by correct evidences, without such annotations. Instead, we compare counterfactual changes in answer confidence with and without evidence sentences, to generate “pseudo-evidentiality” annotations. We validate our proposed model on an original set and challenge set in HotpotQA, showing that our method is accurate and robust in multi-hop reasoning.

This paper studies the problem of generatinglikely queries for multimodal documents withimages. Our application scenario is enablingefficient “first-stage retrieval” of relevant doc-uments, by attaching generated queries to doc-uments before indexing. We can then indexthis expanded text to efficiently narrow downto candidate matches using inverted index, sothat expensive reranking can follow. Our eval-uation results show that our proposed multi-modal representation meaningfully improvesrelevance ranking. More importantly, ourframework can achieve the state of the art inthe first stage retrieval scenarios

This paper studies the keyphrase generation (KG) task for scenarios where structure plays an important role. For example, a scientific publication consists of a short title and a long body, where the title can be used for de-emphasizing unimportant details in the body. Similarly, for short social media posts (, tweets), scarce context can be augmented from titles, though often missing. Our contribution is generating/augmenting structure then injecting these information in the encoding, using existing keyphrases of other documents, complementing missing/incomplete titles. We propose novel structure-augmented document encoding approaches that consist of the following two phases: The first phase, generating structure, extends the given document with related but absent keyphrases, augmenting missing context. The second phase, encoding structure, builds a graph of keyphrases and the given document to obtain the structure-aware representation of the augmented text. Our empirical results validate that our proposed structure augmentation and augmentation-aware encoding/decoding can improve KG for both scenarios, outperforming the state-of-the-art.

This paper studies label augmentation for training dialogue response selection. The existing model is trained by “observational” annotation, where one observed response is annotated as gold. In this paper, we propose “counterfactual augmentation” of pseudo-positive labels. We validate that the effectiveness of augmented labels are comparable to positives, such that ours outperform state-of-the-arts without augmentation.

In this paper, we study review generation given a set of attribute identifiers which are user ID, product ID and rating. This is a difficult subtask of natural language generation since models are limited to the given identifiers, without any specific descriptive information regarding the inputs, when generating the text. The capacity of these models is thus confined and dependent to how well the models can capture vector representations of attributes. We thus propose to additionally leverage references, which are selected from a large pool of texts labeled with one of the attributes, as textual information that enriches inductive biases of given attributes. With these references, we can now pose the problem as an instance of text-to-text generation, which makes the task easier since texts that are syntactically, semantically similar with the output text are provided as input. Using this framework, we address issues such as selecting references from a large candidate set without textual context and improving the model complexity for generation. Our experiments show that our models improve over previous approaches on both automatic and human evaluation metrics.

Existing machine reading comprehension models are reported to be brittle for adversarially perturbed questions when optimizing only for accuracy, which led to the creation of new reading comprehension benchmarks, such as SQuAD 2.0 which contains such type of questions. However, despite the super-human accuracy of existing models on such datasets, it is still unclear how the model predicts the answerability of the question, potentially due to the absence of a shared annotation for the explanation. To address such absence, we release SQuAD2-CR dataset, which contains annotations on unanswerable questions from the SQuAD 2.0 dataset, to enable an explanatory analysis of the model prediction. Specifically, we annotate (1) explanation on why the most plausible answer span cannot be the answer and (2) which part of the question causes unanswerability. We share intuitions and experimental results that how this dataset can be used to analyze and improve the interpretability of existing reading comprehension model behavior.

We aim to leverage human and machine intelligence together for attention supervision. Specifically, we show that human annotation cost can be kept reasonably low, while its quality can be enhanced by machine self-supervision. Specifically, for this goal, we explore the advantage of counterfactual reasoning, over associative reasoning typically used in attention supervision. Our empirical results show that this machine-augmented human attention supervision is more effective than existing methods requiring a higher annotation cost, in text classification tasks, including sentiment analysis and news categorization.

Transfer learning is effective for improving the performance of tasks that are related, and Multi-task learning (MTL) and Cross-lingual learning (CLL) are important instances. This paper argues that hard-parameter sharing, of hard-coding layers shared across different tasks or languages, cannot generalize well, when sharing with a loosely related task. Such case, which we call sparse transfer, might actually hurt performance, a phenomenon known as negative transfer. Our contribution is using adversarial training across tasks, to “soft-code” shared and private spaces, to avoid the shared space gets too sparse. In CLL, our proposed architecture considers another challenge of dealing with low-quality input.

The performance of text classification has improved tremendously using intelligently engineered neural-based models, especially those injecting categorical metadata as additional information, e.g., using user/product information for sentiment classification. This information has been used to modify parts of the model (e.g., word embeddings, attention mechanisms) such that results can be customized according to the metadata. We observe that current representation methods for categorical metadata, which are devised for human consumption, are not as effective as claimed in popular classification methods, outperformed even by simple concatenation of categorical features in the final layer of the sentence encoder. We conjecture that categorical features are harder to represent for machine use, as available context only indirectly describes the category, and even such context is often scarce (for tail category). To this end, we propose using basis vectors to effectively incorporate categorical metadata on various parts of a neural-based model. This additionally decreases the number of parameters dramatically, especially when the number of categorical features is large. Extensive experiments on various data sets with different properties are performed and show that through our method, we can represent categorical metadata more effectively to customize parts of the model, including unexplored ones, and increase the performance of the model greatly.

Generating SQL codes from natural language questions (NL2SQL) is an emerging research area. Existing studies have mainly focused on clear scenarios where specified information is fully given to generate a SQL query. However, in developer forums such as Stack Overflow, questions cover more diverse tasks including table manipulation or performance issues, where a table is not specified. The SQL query posted in Stack Overflow, Pseudo-SQL (pSQL), does not usually contain table schemas and is not necessarily executable, is sufficient to guide developers. Here we describe a new NL2pSQL task to generate pSQL codes from natural language questions on under-specified database issues, NL2pSQL. In addition, we define two new metrics suitable for the proposed NL2pSQL task, Canonical-BLEU and SQL-BLEU, instead of the conventional BLEU. With a baseline model using sequence-to-sequence architecture integrated by denoising autoencoder, we confirm the validity of our task. Experiments show that the proposed NL2pSQL approach yields well-formed queries (up to 43% more than a standard Seq2Seq model). Our code and datasets will be publicly released.

This paper studies the problem of supporting question answering in a new language with limited training resources. As an extreme scenario, when no such resource exists, one can (1) transfer labels from another language, and (2) generate labels from unlabeled data, using translator and automatic labeling function respectively. However, these approaches inevitably introduce noises to the training data, due to translation or generation errors, which require a judicious use of data with varying confidence. To address this challenge, we propose a weakly-supervised framework that quantifies such noises from automatically generated labels, to deemphasize or fix noisy data in training. On reading comprehension task, we demonstrate the effectiveness of our model on low-resource languages with varying similarity to English, namely, Korean and French.

This paper studies the problem of non-factoid question answering, where the answer may span over multiple sentences. Existing solutions can be categorized into representation- and interaction-focused approaches. We combine their complementary strength, by a hybrid approach allowing multi-granular interactions, but represented at word level, enabling an easy integration with strong word-level signals. Specifically, we propose MICRON: Multigranular Interaction for Contextualizing RepresentatiON, a novel approach which derives contextualized uni-gram representation from n-grams. Our contributions are as follows: First, we enable multi-granular matches between question and answer n-grams. Second, by contextualizing word representation with surrounding n-grams, MICRON can naturally utilize word-based signals for query term weighting, known to be effective in information retrieval. We validate MICRON in two public non-factoid question answering datasets: WikiPassageQA and InsuranceQA, showing our model achieves the state of the art among baselines with reported performances on both datasets.

In this paper, we explore strategies to evaluate models for the task research paper novelty detection: Given all papers released at a given date, which of the papers discuss new ideas and influence future research? We find the novelty is not a singular concept, and thus inherently lacks of ground truth annotations with cross-annotator agreement, which is a major obstacle in evaluating these models. Test-of-time award is closest to such annotation, which can only be made retrospectively and is extremely scarce. We thus propose to compare and evaluate models using counterfactual simulations. First, we ask models if they can differentiate papers at time t and counterfactual paper from future time t+d. Second, we ask models if they can predict test-of-time award at t+d. These are proxies that can be agreed by human annotators and easily augmented by correlated signals, using which evaluation can be done through four tasks: classification, ranking, correlation and feature selection. We show these proxy evaluation methods complement each other regarding error handling, coverage, interpretability, and scope, and thus altogether contribute to the observation of the relative strength of existing models.

A major proportion of a text summary includes important entities found in the original text. These entities build up the topic of the summary. Moreover, they hold commonsense information once they are linked to a knowledge base. Based on these observations, this paper investigates the usage of linked entities to guide the decoder of a neural text summarizer to generate concise and better summaries. To this end, we leverage on an off-the-shelf entity linking system (ELS) to extract linked entities and propose Entity2Topic (E2T), a module easily attachable to a sequence-to-sequence model that transforms a list of entities into a vector representation of the topic of the summary. Current available ELS’s are still not sufficiently effective, possibly introducing unresolved ambiguities and irrelevant entities. We resolve the imperfections of the ELS by (a) encoding entities with selective disambiguation, and (b) pooling entity vectors using firm attention. By applying E2T to a simple sequenceto-sequence model with attention mechanism as base model, we see significant improvements of the performance in the Gigaword (sentence to title) and CNN (long document to multi-sentence highlights) summarization datasets by at least 2 ROUGE points.

Cross-cultural differences and similarities are common in cross-lingual natural language understanding, especially for research in social media. For instance, people of distinct cultures often hold different opinions on a single named entity. Also, understanding slang terms across languages requires knowledge of cross-cultural similarities. In this paper, we study the problem of computing such cross-cultural differences and similarities. We present a lightweight yet effective approach, and evaluate it on two novel tasks: 1) mining cross-cultural differences of named entities and 2) finding similar terms for slang across languages. Experimental results show that our framework substantially outperforms a number of baseline methods on both tasks. The framework could be useful for machine translation applications and research in computational social science.

The use of user/product information in sentiment analysis is important, especially for cold-start users/products, whose number of reviews are very limited. However, current models do not deal with the cold-start problem which is typical in review websites. In this paper, we present Hybrid Contextualized Sentiment Classifier (HCSC), which contains two modules: (1) a fast word encoder that returns word vectors embedded with short and long range dependency features; and (2) Cold-Start Aware Attention (CSAA), an attention mechanism that considers the existence of cold-start problem when attentively pooling the encoded word vectors. HCSC introduces shared vectors that are constructed from similar users/products, and are used when the original distinct vectors do not have sufficient information (i.e. cold-start). This is decided by a frequency-guided selective gate vector. Our experiments show that in terms of RMSE, HCSC performs significantly better when compared with on famous datasets, despite having less complexity, and thus can be trained much faster. More importantly, our model performs significantly better than previous models when the training data is sparse and has cold-start problems.

Besides providing the relevant information, amusing users has been an important role of the web. Many web sites provide serendipitous (unexpected but relevant) information to draw user traffic. In this paper, we study the representative scenario of mining an amusing quiz. An existing approach leverages a knowledge base to mine an unexpected property then find quiz questions on such property, based on prototype theory in cognitive science. However, existing deterministic model is vulnerable to noise in the knowledge base. Therefore, we instead propose to leverage probabilistic approach to build a prototype that can overcome noise. Our extensive empirical study shows that our approach not only significantly outperforms baselines by 0.06 in accuracy, and 0.11 in serendipity but also shows higher relevance than the traditional relevance-pursuing baseline using TF-IDF.