NAACL 2024 Papers

The field of cross-lingual sentence embeddings has recently experienced significant advancements, but research concerning low-resource languages has lagged due to the scarcity of parallel corpora. This paper shows that cross-lingual word representation in low-resource languages is notably under-aligned with that in high-resource languages in current models. To address this, we introduce a novel framework that explicitly aligns words between English and eight low-resource languages, utilizing off-the-shelf word alignment models. This framework incorporates three primary training objectives: aligned word prediction and word translation ranking, along with the widely used translation ranking. We evaluate our approach through experiments on the bitext retrieval task, which demonstrate substantial improvements on sentence embeddings in low-resource languages. In addition, the competitive performance of the proposed model across a broader range of tasks in high-resource languages underscores its practicality.

Large language models (LLMs) tend to inadequately integrate input context during text generation, relying excessively on encoded prior knowledge in model parameters, potentially resulting in generated text with factual inconsistencies or contextually unfaithful content. LLMs utilize two primary knowledge sources: 1) prior (parametric) knowledge from pretraining, and 2) contextual (non-parametric) knowledge from input prompts. The study addresses the open question of how LLMs effectively balance these knowledge sources during the generation process, specifically in the context of open-domain question answering. To address this issue, we introduce a novel approach integrating contrastive decoding with adversarial irrelevant passages as negative samples to enhance robust context grounding during generation. Notably, our method operates at inference time without requiring further training. We conduct comprehensive experiments to demonstrate its applicability and effectiveness, providing empirical evidence showcasing its superiority over existing methodologies.

Large language models (LLMs) can achieve impressive performance on various reasoning tasks by incorporating chain-of-thought (CoT) prompting, where step-by-step reasoning is provided to guide LLMs to generate answers to questions, and the question-rationale-answer triplets are utilized as demonstration exemplars. However, the reasoning chains of demonstrations generated by LLMs are observed to be prone to errors, which can subsequently lead to incorrect reasoning during inference. Furthermore, inappropriate exemplars, e.g., overly simplistic or complex exemplars depending on the question’s difficulty level, can affect the LLM’s performance. To address these issues, we introduce Iter-CoT (Iterative bootstrapping in Chain-of-Thoughts prompting). Iter-CoT has two advantages: (1) it adopts iterative bootstrapping that enables LLMs to rectify errors autonomously, resulting in more precise and comprehensive reasoning chains. (2) it selects exemplars of challenging yet answerable (i.e., the LLM has the potential to answer correctly) questions, enhancing the LLMs’ generalizability to answer questions with varying difficulty levels. Experimental results exhibit Iter-CoT superior performance on three distinct reasoning tasks on ten datasets.

The proliferation of social media platforms has given rise to the amount of online debates and arguments. Consequently, the need for automatic summarization methods for such debates is imperative, however this area of summarization is rather understudied. The Key Point Analysis (KPA) task formulates argument summarization as representing the summary of a large collection of arguments in the form of concise sentences in bullet-style format, called key points. A sub-task of KPA, called Key Point Generation (KPG), focuses on generating these key points given the arguments. This paper introduces a novel extractive approach for key point generation, that outperforms previous state-of-the-art methods for the task. Our method utilizes an extractive clustering based approach that offers concise, high quality generated key points with higher coverage of reference summaries, and less redundant outputs. In addition, we show that the existing evaluation metrics for summarization such as ROUGE are incapable of differentiating between generated key points of different qualities. To this end, we propose a new evaluation metric for assessing the generated key points by their coverage. Our code can be accessed online.

Multi-hop question answering (QA) involves finding multiple relevant passages and step-by-step reasoning to answer complex questions, indicating a retrieve-and-read paradigm. However, previous retrievers were customized for two-hop questions, and most of them were trained separately across different hops, resulting in a lack of supervision over the entire multi-hop retrieval process and leading to poor performance in complicated scenarios beyond two hops. In this work, we introduce Beam Retrieval, an end-to-end beam retrieval framework for multi-hop QA. This approach models the multi-hop retrieval process in an end-to-end manner by jointly optimizing an encoder and two classification heads across all hops. Moreover, Beam Retrieval maintains multiple partial hypotheses of relevant passages at each step, expanding the search space and reducing the risk of missing relevant passages. To establish a complete QA system, we incorporate a supervised reader or a large language model (LLM). Experimental results demonstrate that Beam Retrieval achieves a nearly 50% improvement compared with baselines on challenging MuSiQue-Ans, and it also surpasses all previous retrievers on HotpotQA and achieves 99.9% precision on 2WikiMultiHopQA. Providing high-quality context, Beam Retrieval helps our supervised reader achieve new state-of-the-art performance and substantially improves the few-shot QA performance of LLMs.

Interpretability research has shown that self-supervised Spoken LanguageModels (SLMs) encode a wide variety of features in human speech from theacoustic, phonetic, phonological, syntactic and semantic levels, to speakercharacteristics. The bulk of prior research on representations of phonologyhas focused on segmental features such as phonemes; the encoding ofsuprasegmental phonology (such as tone and stress patterns) in SLMs is not yetwell understood. Tone is a suprasegmental feature that is present in more thanhalf of the world’s languages. This paper aims to analyze the tone encodingcapabilities of SLMs, using Mandarin and Vietnamese as case studies. We showthat SLMs encode lexical tone to a significant degree even when they aretrained on data from non-tonal languages. We further find that SLMs behavesimilarly to native and non-native human participants in tone and consonantperception studies, but they do not follow the same developmental trajectory.

Diffusion models have achieved state-of-the-art synthesis quality on both visual and audio tasks, and recent works further adapt them to textual data by diffusing on the embedding space. In this paper, we conduct systematic studies of the optimization challenges encountered with both the embedding space and the denoising model, which have not been carefully explored. Firstly, the data distribution is learnable for embeddings, which may lead to the collapse of the embedding space and unstable training. To alleviate this problem, we propose a new objective called the anchor loss which is more efficient than previous methods. Secondly, we find the noise levels of conventional schedules are insufficient for training a desirable denoising model while introducing varying degrees of degeneration in consequence. To address this challenge, we propose a novel framework called noise rescaling. Based on the above analysis, we propose Difformer, an embedding diffusion model based on Transformer. Experiments on varieties of seminal text generation tasks show the effectiveness of the proposed methods and the superiority of Difformer over previous state-of-the-art embedding diffusion baselines.

Emotion Recognition in Conversation (ERC) involves detecting the underlying emotion behind each utterance within a conversation. Effectively generating representations for utterances remains a significant challenge in this task. Recent works propose various models to address this issue, but they still struggle with differentiating similar emotions such as excitement and happiness. To alleviate this problem, We propose an Emotion-Anchored Contrastive Learning (EACL) framework that can generate more distinguishable utterance representations for similar emotions. To achieve this, we utilize label encodings as anchors to guide the learning of utterance representations and design an auxiliary loss to ensure the effective separation of anchors for similar emotions. Moreover, an additional adaptation process is proposed to adapt anchors to serve as effective classifiers to improve classification performance. Across extensive experiments, our proposed EACL achieves state-of-the-art emotion recognition performance and exhibits superior performance on similar emotions. Our code is available at https://github.com/Yu-Fangxu/EACL.

Large language models can solve new tasks without task-specific fine-tuning. This ability, also known as in-context learning (ICL), is considered an emergent ability and is primarily seen in large language models with billions of parameters. This study investigates if such emergent properties are strictly tied to model size or can be demonstrated by smaller models trained on reduced-scale data. To explore this, we simplify pre-training data and pre-train 36 causal language models with parameters varying from 1 million to 165 million parameters. We show that models trained on this simplified pre-training data demonstrate enhanced zero-shot capabilities across various tasks in simplified language, achieving performance comparable to that of pre-trained models six times larger on unrestricted language. This suggests that downscaling the language allows zero-shot learning capabilities to emerge in models with limited size.Additionally, we find that these smaller models pre-trained on simplified data demonstrate a power law relationship between the evaluation loss and the three scaling factors: compute, dataset size, and model size.

Previous research in multi-document news summarization has typically concentrated on collating information that all sources agree upon. However, the summarization of diverse information dispersed across multiple articles about an event remains underexplored. In this paper, we propose a new task of summarizing diverse information encountered in multiple news articles encompassing the same event. To facilitate this task, we outlined a data collection schema for identifying diverse information and curated a dataset named DiverseSumm. The dataset includes 245 news stories, with each story comprising 10 news articles and paired with a human-validated reference. Next, to enable consistent automatic evaluation, we conducted a comprehensive analysis to pinpoint the position and verbosity biases when utilizing Large Language Model (LLM)-based metrics for evaluating the coverage and faithfulness of summaries. Through correlation analyses, we outline the best practices for effectively using automatic LLM-based metrics on the DiverseSumm dataset. Finally, we study how LLMs summarize multiple news articles by analyzing which type of diverse information LLMs are capable of identifying. Our analyses suggest that despite the extraordinary capabilities of LLMs in single-document summarization, the proposed task remains a complex challenge for them mainly due to their limited coverage, with GPT-4 only able to cover under 40% of the diverse information on average.

Large Language models (LLMs) have shown remarkable success in assisting robot learning tasks, i.e., complex household planning.However, the performance of pretrained LLMs heavily relies on domain-specific templated text data, which may be infeasible in real-world robot learning tasks with image-based observations. Moreover, existing LLMs with text inputs lack the capability to evolve with non-expert interactions with environments.In this work, we introduce a novel learning paradigm that generates robots’ executable actions in the form of text, derived solely from visual observations. Our proposed paradigm stands apart from previous works, which utilized either language instructions or a combination of language and visual data as inputs. We demonstrate that our proposed method can employ two fine-tuning strategies, including imitation learning and reinforcement learning approaches, to adapt to the target test tasks effectively.We conduct extensive experiments involving various model selections, environments, and tasks across 7 house layouts in the VirtualHome environment. Our experimental results demonstrate that our method surpasses existing baselines, confirming the effectiveness of this novel learning paradigm.

Spanish is one of the most widespread languages: the official language in 20 countries and the second most-spoken native language. Its contact with other languages across different regions and the rich regional and cultural diversity has produced varieties which divert from each other, particularly in terms of lexicon. Still, available corpora, and models trained upon them, generally treat Spanish as one monolithic language, which dampers prediction and generation power when dealing with different varieties. To alleviate the situation, we compile and curate datasets in the different varieties of Spanish around the world at an unprecedented scale and create the CEREAL corpus. With such a resource at hand, we perform a stylistic analysis to identify and characterise varietal differences. We implement a classifier specially designed to deal with long documents and identify Spanish varieties (and therefore expand CEREAL further). We produce varietal-specific embeddings, and analyse the cultural differences that they encode. We make data, code and models publicly available.

Generating factual responses is a crucial requirement for dialogue systems. To promotemore factual responses, a common strategyis to ground their responses in relevant documents that inform response generation. However, common dialogue models still often hallucinate information that was not containedin these documents and is therefore unfaithful. In this work, we propose to alleviate suchhallucinations by ‘subtracting’ the parametersof a model trained to hallucinate from a dialogue response generation model in order to‘negate’ the contribution of such hallucinatedexamples from it. Extensive automatic and human evaluation shows favourable results whencompared to state-of-the-art methods that combine the distributions of multiple models, suchas DExperts (Liu et al., 2021), and others thatchange the training procedure, such as Quark(Lu et al., 2022a). Finally, we show how wecan not only reduce hallucinations but also discourage extractive responses, which are oftena consequence of reducing hallucinations byencouraging copy-pasting of document spans.We publicly release our code for reproducibilityand facilitating further research.

Supernet training of LLMs is of great interest in industrial applications as it confers the ability to produce a palette of smaller models at constant cost, regardless of the number of models (of different size / latency) produced. We propose a new method called Multistage Low-rank Fine-tuning of Super-transformers (MLFS) for parameter-efficient supernet training. We show that it is possible to obtain high-quality encoder models that are suitable for commercial edge applications, and that while decoder-only models are resistant to a comparable degree of compression, decoders can be effectively sliced for a significant reduction in training time.

Encoder-decoder foundation models have displayed state-of-the-art performance on a range of autoregressive sequence tasks. This paper proposes a simple and lightweight modification to such systems to control the behaviour according to a specific attribute of interest. This paper proposes a novel inference-efficient approach to modifying the behaviour of an encoder-decoder system according to a specific attribute of interest. Specifically, we show that a small proxy network can be used to find a sample-by-sample perturbation of the encoder output of a frozen foundation model to trigger the decoder to generate improved decodings. This work explores a specific realization of this framework focused on improving the COMET performance of Flan-T5 on Machine Translation and the WER of Whisper foundation models on Speech Recognition. Results display consistent improvements in performance evaluated through COMET and WER respectively. Furthermore, experiments also show that the proxies are robust to the exact nature of the data used to train them and can extend to other domains.

Differentiating relationships between entity pairs with limited labeled instances poses a significant challenge in few-shot relation classification. Representations of textual data extract rich information spanning the domain, entities, and relations. In this paper, we introduce a novel approach to enhance information extraction combining multiple sentence representations and contrastive learning. While representations in relation classification are commonly extracted using entity marker tokens, we argue that substantial information within the internal model representations remains untapped. To address this, we propose aligning multiple sentence representations, such as the CLS] token, the [MASK] token used in prompting, and entity marker tokens. Our method employs contrastive learning to extract complementary discriminative information from these individual representations. This is particularly relevant in low-resource settings where information is scarce. Leveraging multiple sentence representations is especially effective in distilling discriminative information for relation classification when additional information, like relation descriptions, are not available. We validate the adaptability of our approach, maintaining robust performance in scenarios that include relation descriptions, and showcasing its flexibility to adapt to different resource constraints.

Understanding visually situated language requires interpreting complex layouts of textual and visual elements. Pre-processing tools, such as optical character recognition (OCR), can map document image inputs to textual tokens, then large language models (LLMs) can reason over text.However, such methods have high computational and engineering complexity. Can small pretrained image-to-text models accurately understand visual documents through similar recognition and reasoning steps instead?We propose Rationale Distillation (RD), which incorporates the outputs of OCR tools, LLMs, and larger multimodal models as intermediate “rationales”, and trains a small student model to predict both rationales and answers. On three visual document understanding benchmarks representing infographics, scanned documents, and figures, our Pix2Struct (282M parameters) student model finetuned with RD outperforms the base model by 4-5% absolute accuracy with only 1% higher computational cost.

In the context of computational models of dependency syntax, most dependency treebanks have the restriction that any valid dependency tree must have exactly one edge coming out of the root node in addition to respecting the spanning tree constraints. Many algorithms for dependency tree sampling were recently proposed, both for sampling with and without replacement.In this paper we propose a new algorithm called Wilson Reject SWOR for the case of sampling without replacement by adapting the Wilson Reject algorithm originally created for sampling with replacement and combining it with a Trie data structure. Experimental results indicate the efficiency of our approach in the scenario of sampling without replacement from dependency graphs with random weights.

In recent years, there has been a growing interest in utilizing external knowledge to reduce hallucinations in large language models (LLMs) and provide them with updated information. Despite this improvement, a major challenge lies in the lack of explicit citations, which hampers the ability to verify the information generated by these models.This paper focuses on providing models with citation capabilities efficiently. By constructing a dataset of citations, we train two model architectures: an FID-style FLAN-T5 model for efficient answer composition and a 13B model known for its success in instruction following after tuning. Evaluation on fluency, correctness, and citation quality is conducted through human assessment and the newly introduced Automatic LLMs’ Citation Evaluation (ALCE) benchmark.Results demonstrate significant improvements in answer quality and efficiency, surpassing the performance of the popular ChatGPT on some of the metrics. The models exhibit exceptional out-of-domain generalization in both human and automatic evaluation. Notably, the FID-style FLAN-T5 model with only 3B parameters performs impressively compared to the 13B model.