EMNLP 2023 Papers

Social media has become a major driver of social change, by facilitating the formation of online social movements. Automatically understanding the perspectives driving the movement and the voices opposing it, is a challenging task as annotated data is difficult to obtain. We propose a weakly supervised graph-based approach that explicitly models perspectives in \#BackLivesMatter-related tweets. Our proposed approach utilizes a social-linguistic representation of the data. We convert the text to a graph by breaking it into structured elements and connect it with the social network of authors, then structured prediction is done over the elements for identifying perspectives. Our approach uses a small seed set of labeled examples. We experiment with large language models for generating artificial training examples, compare them to manual annotation, and find that it achieves comparable performance. We perform quantitative and qualitative analyses using a human-annotated test set. Our model outperforms multitask baselines by a large margin, successfully characterizing the perspectives supporting and opposing \#BLM.

Unsupervised clustering is widely used to explore large corpora, but existing formulations neither consider the users' goals nor explain clusters' meanings. We propose a new task formulation, "Goal-Driven Clustering with Explanations (GoalEx), which represents both the goal and the explanations as free-form language descriptions. For example, to categorize the errors made by a summarization system, the input to GoalEx is a corpus of annotator-written comments for system-generated summaries and a goal description "cluster the comments based on why the annotators think the summary is imperfect."; the outputs are text clusters each with an explanation ("this cluster mentions that the summary misses important context information."), which relates to the goal and accurately explains which comments should (not) belong to a cluster. To tackle GoalEx, we prompt a language model with "[corpus subset] + [goal] + Brainstorm a list of explanations each representing a cluster."; then we classify whether each sample belongs to a cluster based on its explanation; finally, we use integer linear programming to select a subset of candidate clusters to cover most samples while minimizing overlaps. Under both automatic and human evaluation on corpora with or without labels, our method produces more accurate and goal-related explanations than prior methods.

Multilingual large-scale Pretrained Language Models (PLMs) have been shown to store considerable amounts of factual knowledge, but large variations are observed across languages. With the ultimate goal of ensuring that users with different language backgrounds obtain consistent feedback from the same model, we study the cross-lingual consistency (CLC) of factual knowledge in various multilingual PLMs. To this end, we propose a Ranking-based Consistency (RankC) metric to evaluate knowledge consistency across languages independently from accuracy. Using this metric, we conduct an in-depth analysis of the determining factors for CLC, both at model level and at language-pair level. Among other results, we find that increasing model size leads to higher factual probing accuracy in most languages, but does not improve cross-lingual consistency. Finally, we conduct a case study on CLC when new factual associations are inserted in the PLMs via model editing. Results on a small sample of facts inserted in English reveal a clear pattern whereby the new piece of knowledge transfers only to languages with which English has a high RankC score. All code and data are released at https://github.com/Betswish/Cross-Lingual-Consistency.

The discovery of new intent categories from user utterances is a crucial task in expanding agent skills. The key lies in how to efficiently solicit semantic evidence from utterances and properly transfer knowledge from existing intents to new intents. However, previous methods laid too much emphasis on relations among utterances or clusters for transfer learning, while paying less attention to the usage of semantics. As a result, these methods suffer from in-domain over-fitting and often generate meaningless new intent clusters due to data distortion. In this paper, we present a novel approach called Cluster Semantic Enhanced Prompt Learning (CsePL) for discovering new intents. Our method leverages two-level contrastive learning with label semantic alignment to learn meaningful representations of intent clusters. These learned intent representations are then utilized as soft prompt initializations for discriminating new intents, reducing the dominance of existing intents. Extensive experiments conducted on three public datasets demonstrate the superiority of our proposed method. It not only outperforms existing methods but also suggests meaningful intent labels and enables early detection of new intents.

We explore the connection between presupposition, discourse and sarcasm and propose to leverage that connection in a transfer learning scenario with the goal of improving the performance of NLI models on cases involving presupposition. We exploit advances in training transformer-based models that show that pre-finetuning—--i.e., finetuning the model on an additional task or dataset before the actual finetuning phase—--can help these models, in some cases, achieve a higher performance on a given downstream task. Building on those advances and that aforementioned connection, we propose pre-finetuning NLI models on carefully chosen tasks in an attempt to improve their performance on NLI cases involving presupposition. We notice that, indeed, pre-finetuning on those tasks leads to performance improvements. Furthermore, we run several diagnostic tests to understand whether these gains are merely a byproduct of additional training data. The results show that, while additional training data seems to be helping on its own in some cases, the choice of the tasks plays a role in the performance improvements.

Large language models (LLMs) have demonstrated their potential to refine their generation based on their own feedback. However, the feedback from LLM itself is often inaccurate, thereby limiting its benefits. In this paper, we propose Study Assistant for Large LAnguage Model (SALAM), a novel framework with an auxiliary agent to assist the main LLM in learning from mistakes through interactive cooperation. In the gathering phase, the student assistant agent probes the main LLM, analyzes its errors, and collects the interaction in a mistake memory. During the examination phase, the study assistant provides guidelines by retrieving relevant cases to help the main LLM anticipate and avoid similar errors. We first investigate the effectiveness of a general study assistant and then customize it to provide LLM-specific guidance through imitation learning from successful guidance experiences. Our experiments on three LLMs using two challenging frameworks demonstrate that SALAM can significantly boost LLMs by an accuracy margin of up to 6.6 on BBH and 12.6 on BBQ.

Despite the impressive performance of current AI models reported across various tasks, performance reports often do not include evaluations of how these models perform on the specific groups that will be impacted by these technologies. Among the minority groups under-represented in AI, data from low-income households are often overlooked in data collection and model evaluation. We evaluate the performance of a state-of-the-art vision-language model (CLIP) on a geo-diverse dataset containing household images associated with different income values (DollarStreet) and show that performance inequality exists among households of different income levels. Our results indicate that performance for the poorer groups is consistently lower than the wealthier groups across various topics and countries. We highlight insights that can help mitigate these issues and propose actionable steps for economic-level inclusive AI development.

Pre-trained large language models (LLMs) reflect the inherent social biases of their training corpus. Many methods have been proposed to mitigate this issue, but they often fail to debias or they sacrifice model accuracy. We use *conceptors*--a soft projection method--to identify and remove the bias subspace in LLMs such as BERT and GPT. We propose two methods of applying conceptors (1) bias subspace projection by post-processing by the conceptor NOT operation; and (2) a new architecture, conceptor-intervened BERT (CI-BERT), which explicitly incorporates the conceptor projection into all layers during training. We find that conceptor post-processing achieves state-of-the-art (SoTA) debiasing results while maintaining LLMs' performance on the GLUE benchmark. Further, it is robust in various scenarios and can mitigate intersectional bias efficiently by its AND operation on the existing bias subspaces. Although CI-BERT's training takes all layers' bias into account and can beat its post-processing counterpart in bias mitigation, CI-BERT reduces the language model accuracy. We also show the importance of carefully constructing the bias subspace. The best results are obtained by removing outliers from the list of biased words, combining them (via the OR operation), and computing their embeddings using the sentences from a cleaner corpus.

With the rise of non-autoregressive approach, some non-autoregressive models for joint multiple intent detection and slot filling have obtained the promising inference speed. However, most existing SLU models (1) suffer from the multi-modality problem that leads to reference intents and slots may not be suitable for training; (2) lack of alignment between the correct predictions of the two tasks, which extremely limits the overall accuracy. Therefore, in this paper, we propose $\textbf{M}$odifying the $\textbf{R}$eference via $\textbf{R}$einforcement $\textbf{L}$earning (MRRL), a novel method for multiple intent detection and slot filling, which introduces a modifier module and employs reinforcement learning. Specifically, we try to provide the better training target for the non-autoregressive SLU model via modifying the reference based on the output of the non-autoregressive SLU model, and propose a suitability reward to ensure that the output of the modifier module could fit well with the output of the non-autoregressive SLU model and does not deviate too far from the reference. In addition, we also propose a compromise reward to realize a flexible trade-off between the two subtasks. Experiments on two multi-intent datasets and non-autoregressive baselines demonstrate that our MRRL could consistently improve the performance of baselines. More encouragingly, our best variant achieves new state-of-the-art results, outperforming the previous best approach by 3.6 overall accuracy on MixATIS dataset.

We present the Granular AMR Parsing Evaluation Suite (GrAPES), a challenge set for Abstract Meaning Representation (AMR) parsing with accompanying evaluation metrics. AMR parsers now obtain high scores on the standard AMR evaluation metric Smatch, close to or even above reported inter-annotator agreement. But that does not mean that AMR parsing is solved; in fact, human evaluation in previous work indicates that current parsers still quite frequently make errors on node labels or graph structure that substantially distort sentence meaning. Here, we provide an evaluation suite that tests AMR parsers on a range of phenomena of practical, technical, and linguistic interest. Our 36 categories range from seen and unseen labels, to structural generalization, to coreference. GrAPES reveals in depth the abilities and shortcomings of current AMR parsers.

Significant improvements in end-to-end speech translation (ST) have been achieved through the application of multi-task learning. However, the extent to which auxiliary tasks are highly consistent with the ST task, and how much this approach truly helps, have not been thoroughly studied. In this paper, we investigate the consistency between different tasks, considering different times and modules. We find that the textual encoder primarily facilitates cross-modal conversion, but the presence of noise in speech impedes the consistency between text and speech representations. Furthermore, we propose an improved multi-task learning (IMTL) approach for the ST task, which bridges the modal gap by mitigating the difference in length and representation. We conduct experiments on the MuST-C dataset. The results demonstrate that our method attains state-of-the-art results. Moreover, when additional data is used, we achieve the new SOTA result on MuST-C English to Spanish task with 20.8\% of the training time required by the current SOTA method.

Recently, prompt-based generative frameworks have shown impressive capabilities in sequence labeling tasks. However, in practical dialogue scenarios, relying solely on simplistic templates and traditional corpora presents a challenge for these methods in generalizing to unknown input perturbations. To address this gap, we propose a multi-task demonstration-based generative framework for noisy slot filling, named DemoNSF. Specifically, we introduce three noisy auxiliary tasks, namely noisy recovery (NR), random mask (RM), and hybrid discrimination (HD), to implicitly capture semantic structural information of input perturbations at different granularities. In the downstream main task, we design a noisy demonstration construction strategy for the generative framework, which explicitly incorporates task-specific information and perturbed distribution during training and inference. Experiments on two benchmarks demonstrate that DemoNSF outperforms all baseline methods and achieves strong generalization. Further analysis provides empirical guidance for the practical application of generative frameworks. Our code is released at https://github.com/dongguanting/Demo-NSF.

We introduce SHARCS for adaptive inference that takes into account the hardness of input samples. SHARCS can train a router on any transformer network, enabling the model to direct different samples to sub-networks with varying widths. Our experiments demonstrate that: (1) SHARCS outperforms or complements existing per-sample adaptive inference methods across various classification tasks in terms of accuracy vs. FLOPs; (2) SHARCS generalizes across different architectures and can be even applied to compressed and efficient transformer encoders to further improve their efficiency; (3) SHARCS can provide a 2 times inference speed up at an insignificant drop in accuracy.

Cross-domain Relation Extraction aims to transfer knowledge from a source domain to a different target domain to address low-resource challenges. However, the semantic gap caused by data bias between domains is a major challenge, especially in few-shot scenarios. Previous work has mainly focused on transferring knowledge between domains through shared feature representations without analyzing the impact of each factor that may produce data bias based on the characteristics of each domain. This work takes a causal perspective and proposes a new framework CausalGF. By constructing a unified structural causal model, we estimating the causal effects of factors such as syntactic structure, label distribution,and entities on the outcome. CausalGF calculates the causal effects among the factors and adjusts them dynamically based on domain characteristics, enabling adaptive gap filling. Our experiments show that our approach better fills the domain gap, yielding significantly better results on the cross-domain few-shot relation extraction task.

Text classification is essential for organizing unstructured text. Traditional methods rely on human annotations or, more recently, a set of class seed words for supervision, which can be costly, particularly for specialized or emerging domains. To address this, using class surface names alone as extremely weak supervision has been proposed. However, existing approaches treat different levels of text granularity (documents, sentences, or words) independently, disregarding inter-granularity class disagreements and the context identifiable exclusively through joint extraction. In order to tackle these issues, we introduce MEGClass, an extremely weakly-supervised text classification method that leverages Mutually-Enhancing Text Granularities. MEGClass utilizes coarse- and fine-grained context signals obtained by jointly considering a document's most class-indicative words and sentences. This approach enables the learning of a contextualized document representation that captures the most discriminative class indicators. By preserving the heterogeneity of potential classes, MEGClass can select the most informative class-indicative documents as iterative feedback to enhance the initial word-based class representations and ultimately fine-tune a pre-trained text classifier. Extensive experiments on seven benchmark datasets demonstrate that MEGClass outperforms other weakly and extremely weakly supervised methods.

Deep learning models have reignited the interest in Anomaly Detection research in recent years. Methods for Anomaly Detection in text have shown strong empirical results on ad-hoc anomaly setups that are usually made by downsampling some classes of a labeled dataset. This can lead to reproducibility issues and models that are biased toward detecting particular anomalies while failing to recognize them in more sophisticated scenarios. In the present work, we provide a unified benchmark for detecting various types of anomalies, focusing on problems that can be naturally formulated as Anomaly Detection in text, ranging from syntax to stylistics. In this way, we are hoping to facilitate research in Text Anomaly Detection. We also evaluate and analyze two strong shallow baselines, as well as two of the current state-of-the-art neural approaches, providing insights into the knowledge the neural models are learning when performing the anomaly detection task. We provide the code for evaluation, downloading, and preprocessing the dataset at https://github.com/mateibejan1/ad-nlp/.

Adjusting for latent covariates is crucial for estimating causal effects from observational textual data. Most existing methods only account for confounding covariates that affect both treatment and outcome, potentially leading to biased causal effects. This bias arises from insufficient consideration of non-confounding covariates, which are relevant only to either the treatment or the outcome. In this work, we aim to mitigate the bias by unveiling interactions between different variables to disentangle the non-confounding covariates when estimating causal effects from text. The disentangling process ensures covariates only contribute to their respective objectives, enabling independence between variables. Additionally, we impose a constraint to balance representations from the treated group and control group to alleviate selection bias. We conduct experiments on two different treatment factors under various scenarios, and the proposed model significantly outperforms recent strong baselines. Furthermore, our thorough analysis on earnings call transcripts demonstrates that our model can effectively disentangle the variables, and further investigations into real-world scenarios provide guidance for investors to make informed decisions.

Sparse annotation poses persistent challenges to training dense retrieval models; for example, it distorts the training signal when unlabeled relevant documents are used spuriously as negatives in contrastive learning. To alleviate this problem, we introduce evidence-based label smoothing, a novel, computationally efficient method that prevents penalizing the model for assigning high relevance to false negatives. To compute the target relevance distribution over candidate documents within the ranking context of a given query, we assign a non-zero relevance probability to those candidates most similar to the ground truth based on the degree of their similarity to the ground-truth document(s). To estimate relevance we leverage an improved similarity metric based on reciprocal nearest neighbors, which can also be used independently to rerank candidates in post-processing. Through extensive experiments on two large-scale ad hoc text retrieval datasets, we demonstrate that reciprocal nearest neighbors can improve the ranking effectiveness of dense retrieval models, both when used for label smoothing, as well as for reranking. This indicates that by considering relationships between documents and queries beyond simple geometric distance we can effectively enhance the ranking context.

Stance detection aims to identify the user’s attitude toward specific \textit{targets} from text, which is an important research area in text mining and benefits a variety of application domains. Existing studies on stance detection were conducted mainly in English. Due to the low-resource problem in most non-English languages, cross-lingual stance detection was proposed to transfer knowledge from high-resource (source) language to low-resource (target) language. However, previous research has ignored the practical issue of no labeled training data available in target language. Moreover, target inconsistency in cross-lingual stance detection brings about the additional issue of unseen targets in target language, which in essence requires the transfer of both language and target-oriented knowledge from source to target language. To tackle these challenging issues, in this paper, we propose the new task of cross-lingual cross-target stance detection and develop the first computational work with dual knowledge distillation. Our proposed framework designs a cross-lingual teacher and a cross-target teacher using the source language data and a dual distillation process that transfers the two types of knowledge to target language. To bridge the target discrepancy between languages, cross-target teacher mines target category information and generalizes it to the unseen targets in target language via category-oriented learning. Experimental results on multilingual stance datasets demonstrate the effectiveness of our method compared to the competitive baselines.