NAACL 2024 Papers

Aspect-based sentiment analysis (ABSA) is a task that aims to determine the sentiment polarity of aspects by identifying opinion words. Recent advancements have predominantly been rooted either in semantic or syntactic methods. However, both of them tend to interference from local factors such as irrelevant words and edges, hindering the precise identification of opinion words. In this paper, we present Distance-based and Aspect-oriented Graph Convolutional Network (DAGCN) to address the aforementioned issue. Firstly, we introduce the Distance-based Syntactic Weight (DSW). It focuses on the local scope of aspects in the pruned dependency trees, thereby reducing the candidate pool of opinion words. Additionally, we propose Aspect-Fusion Attention (AF) to further filter opinion words within the local context and consider cases where opinion words are distant from the aspect. With the combination of DSW and AF, we achieve precise identification of corresponding opinion words. Extensive experiments on three public datasets demonstrate that the proposed model outperforms state-of-the-art models and verify the effectiveness of the proposed architecture.

Many leading methods in Vision and language (V+L) pretraining utilize masked language modeling (MLM) as a standard pretraining component, with the expectation that reconstruction of masked text tokens would necessitate reference to corresponding image context via cross/self attention and thus promote representation fusion. However, we observe that the minimization of MLM loss in earlier training stages can depend disproportionately on local text signals, leading to poor training efficiency and inconsistency with the goal of representation fusion. The extent of this lack of cross modal interaction depends strongly which token(s) are masked. To address this issue, we propose a curriculum masking scheme as a replacement for random masking. Tokens are selected to be masked at a frequency proportional to the expected level of cross modal interaction necessary to reconstruct them. This is achieved using a parallel mask selection agent that measures the cross modal flow of information and treats it as a reward to be maximized. By additionally masking contiguous spans that include key objects and their relations, we also achieve better relational understanding, which has been shown to be lacking in many SOTA models. Our experiments on a wide range of V+L tasks show that we trail closely behind state-of-the-art methods despite pretraining on 300x to 1000x less data and we also achieve either top or runner-up performance on tasks from the ARO benchmark which tests compositional relationships. Finally, we demonstrate the potential of our method to scale to larger pretraining data.

The Mayan languages comprise a language family with an ancient history, millions of speakers, and immense cultural value, that, nevertheless, remains severely underrepresented in terms of resources and global exposure. In this paper we develop, curate, and publicly release a set of corpora in several Mayan languages spoken in Guatemala and Southern Mexico, which we call MayanV. The datasets are parallel with Spanish, the dominant language of the region, and are taken from official native sources focused on representing informal, day-to-day, and non-domain-specific language. As such, and according to our dialectometric analysis, they differ in register from most other available resources. Additionally, we present neural machine translation models, trained on as many resources and Mayan languages as possible, and evaluated exclusively on our datasets. We observe lexical divergences between the dialects of Spanish in our resources and the more widespread written standard of Spanish, and that resources other than the ones we present do not seem to improve translation performance, indicating that many such resources may not accurately capture common, real-life language usage. The MayanV dataset is available at https://github.com/transducens/mayanv.

Recent advancements in large language models (LLMs) have underscored their importance in the evolution of artificial intelligence. However, despite extensive pretraining on multilingual datasets, available open-sourced LLMs exhibit limited effectiveness in processing Vietnamese. The challenge is exacerbated by the absence of systematic benchmark datasets and metrics tailored for Vietnamese LLM evaluation. To mitigate these issues, we have finetuned LLMs specifically for Vietnamese and developed a comprehensive evaluation framework encompassing 10 tasks and 31 metrics. We observe that finetuning can help LLMs transfer knowledge across languages, serving as an efficient way to bolster their capabilities in non-English languages. Moreover, our analysis indicates that larger models can introduce more biases and uncalibrated outputs and the key factor influencing LLM performance is the quality of the training or finetuning datasets. These insights underscore the significance of meticulous finetuning with high-quality datasets in enhancing LLM performance.

Cross-Lingual Summarization (XLS) aims to summarize a document in the source language into a condensed version in the target language, effectively removing language barriers for non-native readers. Previous approaches, however, have the same limitation that only a single reference (gold summary) is exploited during model training, making the base model exposed to an underrepresented hypothesis space since the actual number of possible hypotheses is exponentially large. To alleviate this problem, we present a study adopting pseudo-labels in regularizing standard cross-lingual summarization training. We investigate several components leading to the gains in regularization training with verified experiments involving 8 diverse languages from different families. Conclusively, we show that pseudo-labeling is a simple and effective approach that significantly improves over standard gold reference training in XLS.

We propose Label Creative Generation (LCG), a new paradigm in multi-label data augmentation. Beyond repeating data points with fixed labels, LCG creates new data by exploring innovative label combinations. Within LCG, we introduce Tail-Driven Conditional Augmentation (TDCA), combining tail-driven label sampling and label-conditioned text generation for balanced, consistent data augmentation. Our approach has demonstrated a **100.21%** increase in PSP@1 across three datasets, successfully mitigating the long-tail effect in MLTC and markedly enhancing model performance.

In-context learning can improve the performances of knowledge-rich tasks such as question answering. In such scenarios, in-context examples trigger a language model (LM) to surface information stored in its parametric knowledge. We study how to better construct in-context example sets, based on whether the model is aware of the in-context examples. We identify ‘known’ examples, where models can correctly answer from their parametric knowledge, and ‘unknown’ ones. Our experiments show that prompting with ‘unknown’ examples decreases the performance, potentially as it encourages hallucination rather than searching for its parametric knowledge. Constructing an in-context example set that presents both known and unknown information performs the best across diverse settings. We perform analysis on three multi-answer question answering datasets, which allows us to further study answer set ordering strategies based on the LM’s knowledge of each answer. Together, our study sheds light on how to best construct in-context example sets for knowledge-rich tasks.

Recent trends in natural language processing research and annotation tasks affirm a paradigm shift from the traditional reliance on a single ground truth to a focus on individual perspectives, particularly in subjective tasks. In scenarios where annotation tasks are meant to encompass diversity, models that solely rely on the majority class labels may inadvertently disregard valuable minority perspectives. This oversight could result in the omission of crucial information and, in a broader context, risk disrupting the balance within larger ecosystems. As the landscape of annotator modeling unfolds with diverse representation techniques, it becomes imperative to investigate their effectiveness with the fine-grained features of the datasets in view. This study systematically explores various annotator modeling techniques and compares their performance across seven corpora. From our findings, we show that the commonly used user token model consistently outperforms more complex models. We introduce a composite embedding approach and show distinct differences in which model performs best as a function of the agreement with a given dataset. Our findings shed light on the relationship between corpus statistics and annotator modeling performance, which informs future work on corpus construction and perspectivist NLP.

The Directed Acyclic Transformer is a fast non-autoregressive (NAR) model that performs well in Neural Machine Translation. Two issues prevent its application to general Natural Language Generation (NLG) tasks: frequent Out-Of-Vocabulary (OOV) errors and the inability to faithfully generate entity names. We introduce Control-DAG, a constrained decoding algorithm for our Directed Acyclic T5 (DA-T5) model which offers lexical, vocabulary and length control. We show that Control-DAG significantly enhances DA-T5 on the Schema Guided Dialogue and the DART datasets, establishing strong NAR results for Task-Oriented Dialogue and Data-to-Text NLG.

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.

Recently, deep end-to-end learning has been studied for intent classification in Spoken Language Understanding (SLU). However, end-to-end models require a large amount of speech data with intent labels, and highly optimized models are generally sensitive to the inconsistency between the training and evaluation conditions. Therefore, a natural language understanding approach based on Automatic Speech Recognition (ASR) remains attractive because it can utilize a pre-trained general language model and adapt to the mismatch of the speech input environment. Using this module-based approach, we improve a noisy-channel model to handle transcription inconsistencies caused by ASR errors. We propose a two-stage method, Contrastive and Consistency Learning (CCL), that correlates error patterns between clean and noisy ASR transcripts and emphasizes the consistency of the latent features of the two transcripts. Experiments on four benchmark datasets show that CCL outperforms existing methods and improves the ASR robustness in various noisy environments. Code is available at https://github.com/syoung7388/CCL

There exists a discrepancy between the token-level objective during training and the overall sequence-level quality that is expected from the model. This discrepancy leads to issues like exposure bias.To align the model with human expectations, sequence-level objectives are often used to fine-tune pre-trained models.In this paper, we introduce a contrastive preference model that enhances the traditional Plackett-Luce model by incorporating an indicator function. Building upon this novel preference model, we propose Contrastive Preference Learning (CPL), which uses offline samples with list-wise preferences to fine-tune a pre-trained model in Neural Machine Translation. Our experiments, conducted on three language pairs, demonstrate that CPL outperforms not only the vanilla Transformer model but also other token-level and sequence-level baselines. Furthermore, the ablation study highlights the essential role of the proposed indicator function in achieving this improvement.

Recently, language models have accelerated the improvement in natural language processing. However, recent studies have highlighted a significant issue: social biases inherent in training data can lead models to learn and propagate these biases. In this study, we propose a contrastive learning method for bias mitigation, utilizing anchor points to push further negatives and pull closer positives within the representation space. This approach employs stereotypical data as negatives and stereotype-free data as positives, enhancing debiasing performance. Our model attained state-of-the-art performance in the ICAT score on the StereoSet, a benchmark for measuring bias in models. In addition, we observed that effective debiasing is achieved through an awareness of biases, as evidenced by improved hate speech detection scores. The implementation code and trained models are available at https://github.com/HUFS-NLP/CL_Polarizer.git.

Recent work has compared neural network representations via similarity-based analyses to improve model interpretation. The quality of a similarity measure is typically evaluated by its success in assigning a high score to representations that are expected to be matched. However, existing similarity measures perform mediocrely on standard benchmarks. In this work, we develop a new similarity measure, dubbed ContraSim, based on contrastive learning. In contrast to common closed-form similarity measures, ContraSim learns a parameterized measure by using both similar and dissimilar examples. We perform an extensive experimental evaluation of our method, with both language and vision models, on the standard layer prediction benchmark and two new benchmarks that we introduce: the multilingual benchmark and the image–caption benchmark. In all cases, ContraSim achieves much higher accuracy than previous similarity measures, even when presented with challenging examples. Finally, ContraSim is more suitable for the analysis of neural networks, revealing new insights not captured by previous measures.

In recent times, large language models (LLMs) have shown impressive performance on various document-level tasks such as document classification, summarization, and question-answering. However, research on understanding their capabilities on the task of self-contradictions in long documents has been very limited. In this work, we introduce ContraDoc, the first human-annotated dataset to study self-contradictions in long documents across multiple domains, varying document lengths, self-contradiction types, and appearance scope. We then analyze the current capabilities of four state-of-the-art open-source and commercially available LLMs: GPT3.5, GPT4, PaLM2, and LLaMAv2 on this dataset. While GPT4 performs the best and can outperform humans on this task, we find that it is still unreliable and struggles with self-contradictions that require more nuance and context. We release the dataset and all the code associated with the experiments.

Automatic assessment of the quality of arguments has been recognized as a challenging task with significant implications for misinformation and targeted speech. While real-world arguments are tightly anchored in context, existing computational methods analyze their quality in isolation, which affects their accuracy and generalizability. We propose SPARK: a novel method for scoring argument quality based on contextualization via relevant knowledge. We devise four augmentations that leverage large language models to provide feedback, infer hidden assumptions, supply a similar-quality argument, or give a counter-argument. SPARK uses a dual-encoder Transformer architecture to enable the original argument and its augmentation to be considered jointly. Our experiments in both in-domain and zero-shot setups show that SPARK consistently outperforms existing techniques across multiple metrics

Large language models (LLMs) have demonstrated considerable success in various natural language processing tasks, but open-source LLMs have yet to attain state-of-the-art performance in Neural Machine Translation (NMT). Nevertheless, their significant performance in tasks demanding a broad understanding and contextual processing shows their potential for translation. To exploit these abilities, we investigate using LLMs for MT and explore recent parameter-efficient fine-tuning techniques. Surprisingly, our initial experiments found that fine-tuning with Q-LoRA for translation purposes led to performance improvements in terms of BLEU but degradation in COMET compared to in-context learning. To overcome this, we propose an alternative approach: adapting LLMs as Automatic Post-Editors (APE) rather than direct translators. Building on the ability of the LLM to handle long sequences, we also propose extending our approach to document-level translation. We show that leveraging Low-Rank-Adapter fine-tuning for APE can yield significant improvements across both sentence and document-level metrics while generalizing to out-of-domain data. Most notably, we achieve a state-of-the-art accuracy rate of 88.7% on the ContraPro test set, which assesses the model’s ability to resolve pronoun ambiguities when translating from English to German. Lastly, during manual post-editing for document-level translation, the source sentences are iteratively annotated, which can be used to refine further translations in the document. Here, we demonstrate that leveraging human corrections can significantly reduce the number of edits required for subsequent translations.

Label projection, which involves obtaining translated labels and texts jointly, is essential for leveraging machine translation to facilitate cross-lingual transfer in structured prediction tasks. Prior research exploring label projection often compromise translation accuracy by favoring simplified label translation or relying solely on word-level alignments. In this paper, we introduce a novel label projection approach, CLaP, which translates text to the target language and performs *contextual translation* on the labels using the translated text as the context, ensuring better accuracy for the translated labels. We leverage instruction-tuned language models with multilingual capabilities as our contextual translator, imposing the constraint of the presence of translated labels in the translated text via instructions. We benchmark CLaP with other label projection techniques on zero-shot cross-lingual transfer across 39 languages on two representative structured prediction tasks - event argument extraction (EAE) and named entity recognition (NER), showing over 2.4 F1 improvement for EAE and 1.4 F1 improvement for NER. We further explore the applicability of CLaP on ten extremely low-resource languages to showcase its potential for cross-lingual structured prediction.

Crowdsourced labels play a crucial role in evaluating task-oriented dialogue systems (TDSs). Obtaining high-quality and consistent ground-truth labels from annotators presents challenges. When evaluating a TDS, annotators must fully comprehend the dialogue before providing judgments. Previous studies suggest using only a portion of the dialogue context in the annotation process. However, the impact of this limitation on label quality remains unexplored. This study investigates the influence of dialogue context on annotation quality, considering the truncated context for relevance and usefulness labeling. We further propose to use large language models ( LLMs) to summarize the dialogue context to provide a rich and short description of the dialogue context and study the impact of doing so on the annotator’s performance. Reducing context leads to more positive ratings. Conversely, providing the entire dialogue context yields higher-quality relevance ratings but introduces ambiguity in usefulness ratings. Using the first user utterance as context leads to consistent ratings, akin to those obtained using the entire dialogue, with significantly reduced annotation effort. Our findings show how task design, particularly the availability of dialogue context, affects the quality and consistency of crowdsourced evaluation labels.