EMNLP 2023 Papers

We introduce PersonaLM - Domain-distributed Span-Aggregated K-nearest N-gram retrieval augmentation to improve language modeling for Automatic Speech Recognition (ASR) personalization. PersonaLM leverages contextually similar n-gram word frequencies for recognizing rare word patterns associated with unseen domains. It aggregates the next-word probability distribution based on the relative importance of different domains to the input query. To achieve this, we propose a Span Aggregated Group-Contrastive Neural (SCAN) retriever that learns to rank external domains/users by utilizing a group-wise contrastive span loss that pulls together span representations belonging to the same group while pushing away spans from unrelated groups in the semantic space. We propose ASAP benchmark for ASR LM personalization that consists of three user-specific speech-to-text tasks for meetings, TED talks, and financial earnings calls. Extensive experiments show that PersonaLM significantly outperforms strong baselines with a 10-16% improvement in perplexity and a 5-8% reduction in Word Error Rates on popular Wikitext-103, UserLibri, and our ASAP dataset. We further demonstrate the usefulness of the SCAN retriever for improving user-personalized text generation and classification by retrieving relevant context for zero-shot prompting and few-shot fine-tuning of LLMs by 7-12% on the LAMP benchmark.

The field of natural language processing (NLP) has made significant strides in recent years, particularly in the development of large-scale vision-language models (VLMs). These models aim to bridge the gap between text and visual information, enabling a more comprehensive understanding of multimedia data. However, as these models become larger and more complex, they also become more challenging to train and deploy. One approach to addressing this challenge is the use of sparsely-gated mixture-of-experts (MoE) techniques, which divide the model into smaller, specialized sub-models that can jointly solve a task. In this paper, we explore the effectiveness of MoE in scaling vision-language models, demonstrating its potential to achieve state-of-the-art performance on a range of benchmarks over dense models of equivalent computational cost. Our research offers valuable insights into stabilizing the training of MoE models, understanding the impact of MoE on model interpretability, and balancing the trade-offs between compute performance when scaling VLMs. We hope our work will inspire further research into the use of MoE for scaling large-scale vision-language models and other multimodal machine learning applications.

Aspect-based sentiment analysis (ABSA) has been widely studied since the explosive growth of social networking services. However, the recognition of implicit sentiments that do not contain obvious opinion words remains less explored. In this paper, we propose aspect-category enhanced learning with a neural coherence model (ELCoM). It captures document-level coherence by using contrastive learning, and sentence-level by a hypergraph to mine opinions from explicit sentences to aid implicit sentiment classification. To address the issue of sentences with different sentiment polarities in the same category, we perform cross-category enhancement to offset the impact of anomalous nodes in the hypergraph and obtain sentence representations with enhanced aspect-category. Extensive experiments on benchmark datasets show that the ELCoM achieves state-of-the-art performance. Our source codes and data are released at \url{https://github.com/cuijin-23/ELCoM}.

Most existing stylistic text rewriting methods and evaluation metrics operate on a sentence level, but ignoring the broader context of the text can lead to preferring generic, ambiguous, and incoherent rewrites. In this paper, we investigate integrating the preceding textual context into both the $\textit{rewriting}$ and $\textit{evaluation}$ stages of stylistic text rewriting, and introduce a new composite contextual evaluation metric $\texttt{CtxSimFit}$ that combines similarity to the original sentence with contextual cohesiveness. We comparatively evaluate non-contextual and contextual rewrites in formality, toxicity, and sentiment transfer tasks. Our experiments show that humans significantly prefer contextual rewrites as more fitting and natural over non-contextual ones, yet existing sentence-level automatic metrics (e.g., ROUGE, SBERT) correlate poorly with human preferences ($\rho$=0--0.3). In contrast, human preferences are much better reflected by both our novel $\texttt{CtxSimFit}$ ($\rho$=0.7--0.9) as well as proposed context-infused versions of common metrics ($\rho$=0.4--0.7). Overall, our findings highlight the importance of integrating context into the generation and especially the evaluation stages of stylistic text rewriting.

The remarkable abilities of large language models (LLMs) like ChatGPT and GPT-4 partially stem from the post-training processes involving human preferences encoded within a reward model as part of a Reinforcement Learning from Human Feedback (RLHF) regimen. These reward models (RMs) often lack direct knowledge of why, or under what principles, the preferences annotations were made. In this study, we identify principles that guide RMs to better align with human preferences, and then develop an axiomatic framework to generate a rich variety of preference signals to uphold them. We use these axiomatic signals to train a model for the scoring answers to longform questions. Our approach yields a \textbf{Preference Model} with only about 220M parameters that agrees with gold human-annotated preference labels more often than GPT-4. The contributions of this work include: training a standalone preference model that can score human- and LLM-generated answers on the same scale; developing an axiomatic framework for generating training data pairs tailored to certain principles; and showing that a small amount of axiomatic signals can help small models outperform GPT-4 in preference scoring. We intend to release our axiomatic data and model.

Adversarial samples pose a significant challenge to neural inference models. In this paper, we propose a novel enhancing approach A3 for the robustness of the neural NLP models, which combines the adversarial training and data augmentation. We propose an adversarial sample generator that consists of a conditioned paraphrasing model and a condition generator. The latter aims to generate conditions which guides the paraphrasing model to generate adversarial samples. A pretrained discriminator is introduced to help the adversarial sample generator adapt to the data characteristics for different tasks. We adopt a weighted loss to incorporate the generated adversarial samples with the original samples for augmented training. Compared to existing methods, our approach is much efficient since the generation process is independent to the target model and the generated samples are reusable for different models. Experimental results on several tasks show that our approach improves the overall performance of the trained model. Specially, the enhanced model is robust for various attacking techniques.

The proliferation of misinformation on social media platforms (SMPs) poses a significant danger to public health, social cohesion and ultimately democracy. Previous research has shown how social correction can be an effective way to curb misinformation, by engaging directly in a constructive dialogue with users who spread – often in good faith – misleading messages. Although professional fact-checkers are crucial to debunking viral claims, they usually do not engage in conversations on social media. Thereby, significant effort has been made to automate the use of fact-checker material in social correction; however, no previous work has tried to integrate it with the style and pragmatics that are commonly employed in social media communication. To fill this gap, we present VerMouth, the first large-scale dataset comprising roughly 12 thousand claim-response pairs (linked to debunking articles), accounting for both SMP-style and basic emotions, two factors which have a significant role in misinformation credibility and spreading. To collect this dataset we used a technique based on an author-reviewer pipeline, which efficiently combines LLMs and human annotators to obtain high-quality data. We also provide comprehensive experiments showing how models trained on our proposed dataset have significant improvements in terms of output quality and generalization capabilities.

Existing works on coreference resolution suggest that task-specific models are necessary to achieve state-of-the-art performance. In this work, we present compelling evidence that such models are not necessary. We finetune a pretrained seq2seq transformer to map an input document to a tagged sequence encoding the coreference annotation. Despite the extreme simplicity, our model outperforms or closely matches the best coreference systems in the literature on an array of datasets. We consider an even simpler version of seq2seq that generates only the tagged spans and find it highly performant. Our analysis shows that the model size, the amount of supervision, and the choice of sequence representations are key factors in performance.

Complex Query Answering (CQA) is a challenge task of Knowledge Graph (KG). Due to the incompleteness of KGs, query embedding (QE) methods have been proposed to encode queries and entities into the same embedding space, and treat logical operators as neural set operators to obtain answers. However, these methods train KG embeddings and neural set operators concurrently on both simple (one-hop) and complex (multi-hop and logical) queries, which causes performance degradation on simple queries and low training efficiency. In this paper, we propose Query to Triple (Q2T), a novel approach that decouples the training for simple and complex queries. Q2T divides the training into two stages: (1) Pre-training the neural link predictor on simple queries to predict tail entities based on the head entity and relation. (2) Training the query encoder on complex queries to encode diverse complex queries into a unified triple form that can be efficiently solved by the pretrained link predictor. Our proposed Q2T is not only efficient to train, but also modular, thus easily adaptable to various neural link predictors that have been studied well. Extensive experiments demonstrate that, even without explicit modeling for neural set operators, Q2T still achieves state-of-the-art performance on diverse complex queries over three public benchmarks.

When translating words referring to the speaker, speech translation (ST) systems should not resort to default masculine generics nor rely on potentially misleading vocal traits. Rather, they should assign gender according to the speakers' preference. The existing solutions to do so, though effective, are hardly feasible in practice as they involve dedicated model re-training on gender-labeled ST data. To overcome these limitations, we propose the first inference-time solution to control speaker-related gender inflections in ST. Our approach partially replaces the (biased) internal language model (LM) implicitly learned by the ST decoder with gender-specific external LMs. Experiments on en→es/fr/it show that our solution outperforms the base models and the best training-time mitigation strategy by up to 31.0 and 1.6 points in gender accuracy, respectively, for feminine forms. The gains are even larger (up to 32.0 and 3.4) in the challenging condition where speakers' vocal traits conflict with their gender.

Analogy-making between narratives is crucial for human reasoning. In this paper, we evaluate the ability to identify and generate analogies by constructing a first-of-its-kind large-scale story-level analogy corpus, \textsc{StoryAnalogy}, which contains 24K story pairs from diverse domains with human annotations on two similarities from the extended Structure-Mapping Theory. We design a set of tests on \textsc{StoryAnalogy}, presenting the first evaluation of story-level analogy identification and generation. Interestingly, we find that the analogy identification tasks are incredibly difficult not only for sentence embedding models but also for the recent large language models (LLMs) such as ChatGPT and LLaMa. ChatGPT, for example, only achieved around 30\% accuracy in multiple-choice questions (compared to over 85\% accuracy for humans). Furthermore, we observe that the data in \textsc{StoryAnalogy} can improve the quality of analogy generation in LLMs, where a fine-tuned FlanT5-xxl model achieves comparable performance to zero-shot ChatGPT.

To enhance the multi-step reasoning capabilities of large language models, researchers have extensively explored prompting methods, notably the Chain-of-Thought (CoT) method which explicitly elicits human-like rationales. However, they have inadvertently overlooked the potential of enhancing model reasoning performance by formulating higher-quality problems \footnote{A reasoning problem often consists of two parts: the context and the final question \cite{DBLP:journals/corr/abs-2205-09712}.}. In this work, we start from the problem side and propose Self-Polish (SP), a novel method that facilitates the model's reasoning by guiding it to progressively refine the given problems to be more comprehensible and solvable. We also explore several automatic prompting varients and propose the Self-Polish prompt bank for the community. SP is orthogonal to all other prompting methods of answer/reasoning side like CoT, allowing for seamless integration with state-of-the-art techniques for further improvement. Thorough experiments show that the proposed method attains notable and consistent effectiveness on five reasoning benchmarks across different models. Furthermore, our method also showcases impressive performance on robustness evaluation. Codes and prompts are available at https://github.com/WooooDyy/Self-Polish.

As the impact of social media gradually escalates, people are more likely to be exposed to indistinguishable fake news. Therefore, numerous studies have attempted to detect rumors on social media by analyzing the textual content and propagation paths. However, fewer works on rumor detection tasks consider the malicious attacks commonly observed at response level. Moreover, existing detection models have poor interpretability. To address these issues, we propose a novel framework named **D**efend-**A**nd-**S**ummarize (DAS) based on the concept that responses sharing similar opinions should exhibit similar features. Specifically, DAS filters out the attack responses and summarizes the responsive posts of each conversation thread in both extractive and abstractive ways to provide multi-perspective prediction explanations. Furthermore, we enhance our detection architecture with the transformer and Bi-directional Graph Convolutional Networks. Experiments on three public datasets, *i.e.*, RumorEval2019, Twitter15, and Twitter16, demonstrate that our DAS defends against malicious attacks and provides prediction explanations, and the proposed detection model achieves state-of-the-art.

It is widely acknowledged that large and sparse models have higher accuracy than small and dense models under the same model size constraints. This motivates us to train a large model and then remove its redundant neurons or weights by pruning. Most existing works pruned the networks in a deterministic way, the performance of which solely depends on a single pruning criterion and thus lacks variety. Instead, in this paper, we propose a model pruning strategy that first generates several pruning masks in a designed random way. Subsequently, along with an effective mask-selection rule, the optimal mask is chosen from the pool of mask candidates. To further enhance efficiency, we introduce an early mask evaluation strategy, mitigating the overhead associated with training multiple masks. Our extensive experiments demonstrate that this approach achieves state-of-the-art performance across eight datasets from GLUE, particularly excelling at high levels of sparsity.

Extensive work has shown that the performance and interpretability of commonsense reasoning can be improved via knowledge-augmented reasoning methods, where the knowledge that underpins the reasoning process is explicitly verbalized and utilized. However, existing implementations, including "chain-of-thought" and its variants, fall short in capturing the *introspective* nature of knowledge required in commonsense reasoning, and in accounting for the mutual adaptation between the generation and utilization of knowledge. We propose a novel method to develop an introspective commonsense reasoner, **Crystal**. To tackle commonsense problems, it first introspects for knowledge statements related to the given question, and subsequently makes an informed prediction that is grounded in the previously introspected knowledge. The knowledge introspection and knowledge-grounded reasoning modes of the model are tuned via reinforcement learning to mutually adapt, where the reward derives from the feedback given by the model itself. Experiments show that Crystal significantly outperforms both the standard supervised finetuning and chain-of-thought distilled methods, and enhances the transparency of the commonsense reasoning process. Our work ultimately validates the feasibility and potential of reinforcing a neural model with self-feedback.

Detecting hallucinations in natural language processing (NLP) is a critical undertaking that demands a deep understanding of both the semantic and pragmatic aspects of languages. Cognitive approaches that leverage users’ behavioural signals, such as gaze, have demonstrated effectiveness in addressing NLP tasks with similar linguistic complexities. However, their potential in the context of hallucination detection remains largely unexplored. In this paper, we propose a novel cognitive approach for hallucination detection that leverages gaze signals from humans. We first collect and introduce an eye tracking corpus (IITB-HGC: IITB-Hallucination Gaze corpus) consisting of 500 instances, annotated by five annotators for hallucination detection. Our analysis reveals that humans selectively attend to relevant parts of the text based on distributional similarity, similar to the attention bias phenomenon in psychology. We identify two attention strategies employed by humans: global attention, which focuses on the most informative sentence, and local attention, which focuses on important words within a sentence. Leveraging these insights, we propose a novel cognitive framework for hallucination detection that incorporates these attention biases. Experimental evaluations on the FactCC dataset demonstrate the efficacy of our approach, obtaining a balanced accuracy of 87.1%. Our study highlights the potential of gaze-based approaches in addressing the task of hallucination detection and sheds light on the cognitive processes employed by humans in identifying inconsistencies.

Most dense retrieval models contain an implicit assumption: the training query-document pairs are exactly matched. Since it is expensive to annotate the corpus manually, training pairs in real-world applications are usually collected automatically, which inevitably introduces mismatched-pair noise. In this paper, we explore an interesting and challenging problem in dense retrieval, how to train an effective model with mismatched-pair noise. To solve this problem, we propose a novel approach called Noisy Pair Corrector (NPC), which consists of a detection module and a correction module. The detection module estimates noise pairs by calculating the perplexity between annotated positive and easy negative documents. The correction module utilizes an exponential moving average (EMA) model to provide a soft supervised signal, aiding in mitigating the effects of noise. We conduct experiments on text-retrieval benchmarks Natural Question and TriviaQA, code-search benchmarks StaQC and SO-DS. Experimental results show that NPC achieves excellent performance in handling both synthetic and realistic noise.

While Diffusion Generative Models have achieved great success on image generation tasks, how to efficiently and effectively incorporate them into speech generation especially translation tasks remains a non-trivial problem. Specifically, due to the low information density of speech data, the transformed discrete speech unit sequence is much longer than the corresponding text transcription, posing significant challenges to existing auto-regressive models. Furthermore, it is not optimal to brutally apply discrete diffusion on the speech unit sequence while disregarding the continuous space structure, which will degrade the generation performance significantly. In this paper, we propose a novel diffusion model by applying the diffusion forward process in the continuous speech representation space, while employing the diffusion backward process in the discrete speech unit space. In this way, we preserve the semantic structure of the continuous speech representation space in the diffusion process and integrate the continuous and discrete diffusion models. We conduct extensive experiments on the textless direct speech-to-speech translation task, where the proposed method achieves comparable results to the computationally intensive auto-regressive baselines (500 steps on average) with significantly fewer decoding steps (50 steps).

Portuguese Sign Language (LGP) is the official language in deaf education in Portugal. Current approaches in developing a translation system between European Portuguese and LGP rely on hand-crafted rules. In this paper, we present a fully automatic corpora-driven rule-based machine translation system between European Portuguese and LGP glosses, and also two neural machine translation models. We also contribute with the LGP-5-Domain corpus, composed of five different text domains, built with the help of our rule-based system, and used to train the neural models. In addition, we provide a gold collection, annotated by LGP experts, that can be used for future evaluations. Compared with the only similar available translation system, PE2LGP, results are always improved with the new rule-based model, which competes for the highest scores with one of the neural models.