EMNLP 2023 Papers

Abstract grammatical knowledge—of parts of speech and grammatical patterns—is key to the capacity for linguistic generalization in humans. But how abstract is grammatical knowledge in large language models? In the human literature, compelling evidence for grammatical abstraction comes from structural priming. A sentence that shares the same grammatical structure as a preceding sentence is processed and produced more readily. Because confounds exist when using stimuli in a single language, evidence of abstraction is even more compelling from crosslingual structural priming, where use of a syntactic structure in one language primes an analogous structure in another language. We measure crosslingual structural priming in large language models, comparing model behavior to human experimental results from eight crosslingual experiments covering six languages, and four monolingual structural priming experiments in three non-English languages. We find evidence for abstract monolingual and crosslingual grammatical representations in the models that function similarly to those found in humans. These results demonstrate that grammatical representations in multilingual language models are not only similar across languages, but they can causally influence text produced in different languages.

In this paper, we investigate the impact of objects on gender bias in image captioning systems. Our results show that only gender-specific objects have a strong gender bias (e.g., women-lipstick). In addition, we propose a visual semantic-based gender score that measures the degree of bias and can be used as a plug-in for any image captioning system. Our experiments demonstrate the utility of the gender score, since we observe that our score can measure the bias relation between a caption and its related gender; therefore, our score can be used as an additional metric to the existing Object Gender Co-Occ approach.

Question Answering over Knowledge Graph (KGQA) aims to seek answer entities for the natural language question from a large-scale Knowledge Graph (KG). To better perform reasoning on KG, recent work typically adopts a pre-trained language model (PLM) to model the question, and a graph neural network (GNN) based module to perform multi-hop reasoning on the KG. Despite the effectiveness, due to the divergence in model architecture, the PLM and GNN are not closely integrated, limiting the knowledge sharing and fine-grained feature interactions. To solve it, we aim to simplify the above two-module approach, and develop a more capable PLM that can directly support subgraph reasoning for KGQA, namely ReasoningLM. In our approach, we propose a subgraph-aware self-attention mechanism to imitate the GNN for performing structured reasoning, and also adopt an adaptation tuning strategy to adapt the model parameters with 20,000 subgraphs with synthesized questions. After adaptation, the PLM can be parameter-efficient fine-tuned on downstream tasks. Experiments show that ReasoningLM surpasses state-of-the-art models by a large margin, even with fewer updated parameters and less training data. Our codes and data are publicly available at https://github.com/RUCAIBox/ReasoningLM.

Recent advances in deep learning, and especially the invention of encoder-decoder architectures, have significantly improved the performance of abstractive summarization systems. While the majority of research has focused on written documents, we have observed an increasing interest in the summarization of dialogues and multi-party conversations over the past few years. In this paper, we present a dataset of French political debates for the purpose of enhancing resources for multi-lingual dialogue summarization. Our dataset consists of manually transcribed and annotated political debates, covering a range of topics and perspectives. We highlight the importance of high-quality transcription and annotations for training accurate and effective dialogue summarization models, and emphasize the need for multilingual resources to support dialogue summarization in non-English languages. We also provide baseline experiments using state-of-the-art methods, and encourage further research in this area to advance the field of dialogue summarization. Our dataset will be made publicly available for use by the research community, enabling further advances in multilingual dialogue summarization.

We propose a general method to break down a main complex task into a set of intermediary easier sub-tasks, which are formulated in natural language as binary questions related to the final target task. Our method allows for representing each example by a vector consisting of the answers to these questions. We call this representation Natural Language Learned Features (NLLF). NLLF is generated by a small transformer language model (e.g., BERT) that has been trained in a Natural Language Inference (NLI) fashion, using weak labels automatically obtained from a Large Language Model (LLM). We show that the LLM normally struggles for the main task using in-context learning, but can handle these easiest subtasks and produce useful weak labels to train a BERT. The NLI-like training of the BERT allows for tackling zero-shot inference with any binary question, and not necessarily the ones seen during the training. We show that this NLLF vector not only helps to reach better performances by enhancing any classifier, but that it can be used as input of an easy-to-interpret machine learning model like a decision tree. This decision tree is interpretable but also reaches high performances, surpassing those of a pre-trained transformer in some cases. We have successfully applied this method to two completely different tasks: detecting incoherence in students' answers to open-ended mathematics exam questions, and screening abstracts for a systematic literature review of scientific papers on climate change and agroecology.

The rapid growth of web pages and the increasing complexity of their structure poses a challenge for web mining models. Web mining models are required to understand semi-structured web pages, particularly when little is known about the subject or template of a new page. Current methods migrate language models to web mining by embedding the XML source code into the transformer or encoding the rendered layout with graph neural networks. However, these approaches do not take into account the relationships between text nodes within and across pages. In this paper, we propose a new approach, ReXMiner, for zero-shot relation extraction in web mining. ReXMiner encodes the shortest relative paths in the Document Object Model (DOM) tree of the web page which is a more accurate and efficient signal for key-value pair extraction within a web page. It also incorporates the popularity of each text node by counting the occurrence of the same text node across different web pages. We use contrastive learning to address the issue of sparsity in relation extraction. Extensive experiments on public benchmarks show that our method, ReXMiner, outperforms the state-of-the-art baselines in the task of zero-shot relation extraction in web mining.

Using a narrative style is an effective way to communicate health information both on and off social media. Given the amount of misinformation being spread online and its potential negative effects, it is crucial to investigate the interplay between narrative communication style and misinformative health content on user engagement on social media platforms. To explore this in the context of Twitter, we start with previously annotated health misinformation tweets (n ≈15,000) and annotate a subset of the data (n=3,000) for the presence of narrative style. We then use these manually assigned labels to train text classifiers, experimenting with supervised fine-tuning and in-context learning for automatic narrative detection. We use our best model to label remaining portion of the dataset, then statistically analyze the relationship between narrative style, misinformation, and user-level features on engagement, finding that narrative use is connected to increased tweet engagement and can, in some cases, lead to increased engagement with misinformation. Finally, we analyze the general categories of language used in narratives and health misinformation in our dataset.

Pre-trained language models~(LMs) have brought remarkable performance on numerous NLP tasks. However, they require significant resources and entail high computational costs for inference, making them challenging to deploy in real-world and real-time systems. Existing early exiting methods aim to reduce computational complexity by selecting the layer at which to exit, but suffer from the limitation that they have to sequentially traverse through all layers prior to the selected exit layer, which lacks flexibility and degrades their performance. To solve this problem, we propose a \textbf{h}omotopic and \textbf{ad}aptive layer \textbf{skip}ping fine-tuning method named HadSkip. HadSkip adaptively selects the layers to skip based on a predefined budget. Specifically, we introduce a learnable gate before each layer of the LM to determine whether the current layer should be skipped. To tackle various challenges in training such as discrete gates and the budget constraint, we propose a fine-grained initialization strategy and homotopic optimization strategy. We conduct extensive experiments on the GLUE benchmark, and experimental results demonstrate the proposed HadSkip outperforms all state-of-the-art baselines significantly.

Mental illness remains one of the most critical public health issues of our time, due to the severe scarcity and accessibility limit of professionals. Psychotherapy requires high-level expertise to conduct deep, complex reasoning and analysis on the cognition modeling of the patients. In the era of Large Language Models, we believe it is the right time to develop AI assistance for computational psychotherapy. We study the task of cognitive distortion detection and propose the Diagnosis of Thought (DoT) prompting. DoT performs diagnosis on the patient's speech via three stages: subjectivity assessment to separate the facts and the thoughts; contrastive reasoning to elicit the reasoning processes supporting and contradicting the thoughts; and schema analysis to summarize the cognition schemas. The generated diagnosis rationales through the three stages are essential for assisting the professionals. Experiments demonstrate that DoT obtains significant improvements over ChatGPT for cognitive distortion detection, while generating high-quality rationales approved by human experts.

While pre-trained language models (PLMs) have shown evidence of acquiring vast amounts of knowledge, it remains unclear how much of this parametric knowledge is actually usable in performing downstream tasks. We propose a systematic framework to measure parametric knowledge utilization in PLMs. Our framework first extracts knowledge from a PLM's parameters and subsequently constructs a downstream task around this extracted knowledge. Performance on this task thus depends exclusively on utilizing the model's possessed knowledge, avoiding confounding factors like insufficient signal. As an instantiation, we study factual knowledge of PLMs and measure utilization across 125M to 13B parameter PLMs. We observe that: (1) PLMs exhibit two gaps - in acquired vs. utilized knowledge, (2) they show limited robustness in utilizing knowledge under distribution shifts, and (3) larger models close the acquired knowledge gap but the utilized knowledge gap remains. Overall, our study provides insights into PLMs' capabilities beyond their acquired knowledge.

Non-compositional expressions, by virtue of their non-compositionality, are a classic `pain in the neck' for NLP systems. Different from the general language modeling and generation tasks that are primarily compositional, generating non-compositional expressions is more challenging for current neural models, including large pre-trained language models. The main reasons are 1) their non-compositionality, and 2) the limited data resources. Therefore, to make the best use of available data for modeling non-compositionality, we propose a dynamic curriculum learning framework, which learns training examples from easy ones to harder ones thus optimizing the learning step by step but suffers from the forgetting problem. To alleviate the forgetting problem brought by the arrangement of training examples, we also apply a continual learning method into our curriculum learning framework. Our proposed method combined curriculum and continual learning, to gradually improve the model's performance on the task of non-compositional expression generation. Experiments on idiomatic expression generation and metaphor generation affirm the effectiveness of our proposed curriculum learning framework and the application of continual learning. Our codes are available at https://github.com/zhjjn/CL2Gen.git.

As generative large language models (LLMs) grow more performant and prevalent, we must develop comprehensive enough tools to measure and improve their fairness. Different prompt-based datasets can be used to measure social bias across multiple text domains and demographic axes, meaning that testing LLMs on more datasets can potentially help us characterize their biases more fully, and better ensure equal and equitable treatment of marginalized demographic groups. In this work, our focus is two-fold: (1) Benchmarking: a comparison of 6 different prompt-based bias and toxicity metrics across 12 demographic axes and 5 families of generative LLMs. Out of those 6 metrics, AdvPromptSet and HolisticBiasR are novel datasets proposed in the paper. The comparison of those benchmarks gives us insights about the bias and toxicity of the compared models. Therefore, we explore the frequency of demographic terms in common LLM pre-training corpora and how this may relate to model biases. (2) Mitigation: we conduct a comprehensive study of how well 3 bias/toxicity mitigation techniques perform across our suite of measurements. ROBBIE aims to provide insights for practitioners while deploying a model, emphasizing the need to not only measure potential harms, but also understand how they arise by characterizing the data, mitigate harms once found, and balance any trade-offs. We open-source our analysis code in hopes of encouraging broader measurements of bias in future LLMs.

Recently, post-processing networks (PPNs), which modify the outputs of arbitrary modules including non-differentiable ones in task-oriented dialogue systems, have been proposed. PPNs have successfully improved the dialogue performance by post-processing natural language understanding (NLU), dialogue state tracking (DST), and dialogue policy (Policy) modules with a classification-based approach. However, they cannot be applied to natural language generation (NLG) modules because the post-processing of the utterance output by the NLG module requires a generative approach. In this study, we propose a new post-processing component for NLG, generative post-processing networks (GenPPNs). For optimizing GenPPNs via reinforcement learning, the reward function incorporates dialogue act contribution, a new measure to evaluate the contribution of GenPPN-generated utterances with regard to task completion in dialogue. Through simulation and human evaluation experiments based on the MultiWOZ dataset, we confirmed that GenPPNs improve the task completion performance of task-oriented dialogue systems.

This work presents a manually annotated dataset for Information Extraction (IE) from legal wills, and relevant in-context learning experiments on the dataset. The dataset consists of entities, binary relations between the entities (e.g., relations between testator and beneficiary), and n-ary events (e.g., bequest) extracted from 45 legal wills from two US states. This dataset can serve as a foundation for downstream tasks in the legal domain. Another use case of this dataset is evaluating the performance of large language models (LLMs) on this IE task. We evaluated GPT-4 with our dataset to investigate its ability to extract information from legal wills. Our evaluation result demonstrates that the model is capable of handling the task reasonably well. When given instructions and examples as a prompt, GPT-4 shows decent performance for both entity extraction and relation extraction tasks. Nevertheless, the evaluation result also reveals that the model is not perfect. We observed inconsistent outputs (given a prompt) as well as prompt over-generalization.

Transformer-based language models (LMs) are powerful and widely-applicable tools, but their usefulness is constrained by a finite context window and the expensive computational cost of processing long text documents. We propose to adapt pre-trained LMs into AutoCompressors. These language models are capable of compressing long contexts into summary vectors, which are then accessible to the model as soft prompts. Summary vectors are trained with an unsupervised objective, whereby long documents are processed in segments, and summary vectors from all previous segments are used in language modeling. We fine-tune OPT and Llama-2 models on sequences of up to 30,720 tokens and show that AutoCompressors can utilize long contexts to improve perplexity. We evaluate AutoCompressors on in-context learning by compressing task demonstrations and find that summary vectors are good substitutes for plain-text demonstrations, increasing accuracy while reducing inference costs. Finally, we explore the benefits of pre-computing summary vectors for large corpora by applying summary vectors to retrieval-augmented language modeling and a passage re-ranking task. Overall, AutoCompressors emerge as a simple and inexpensive solution to extend the context window of LMs while speeding up inference over long contexts.

We present a setup for training, evaluating and interpreting neural language models, that uses artificial, language-like data. The data is generated using a massive probabilistic grammar (based on state-split PCFGs), that is itself derived from a large natural language corpus, but also provides us complete control over the generative process. We describe and release both grammar and corpus, and test for the naturalness of our generated data. This approach allows us define closed-form expressions to efficiently compute exact lower bounds on obtainable perplexity using both causal and masked language modelling. Our results show striking differences between neural language modelling architectures and training objectives in how closely they allow approximating the lower bound on perplexity. Our approach also allows us to directly compare learned representations to symbolic rules in the underlying source. We experiment with various techniques for interpreting model behaviour and learning dynamics. With access to the underlying true source, our results show striking differences and outcomes in learning dynamics between different classes of words.

Building automatic extraction models for visually rich documents like invoices, receipts, bills, tax forms, etc. has received significant attention lately. A key bottleneck in developing extraction models for new document types is the cost of acquiring the several thousand high-quality labeled documents that are needed to train a model with acceptable accuracy. In this paper, we propose selective labeling as a solution to this problem. The key insight is to simplify the labeling task to provide “yes/no” labels for candidate extractions predicted by a model trained on partially labeled documents. We combine this with a custom active learning strategy to find the predictions that the model is most uncertain about. We show through experiments on document types drawn from 3 different domains that selective labeling can reduce the cost of acquiring labeled data by 10× with a negligible loss in accuracy.

Efficiently retrieving FAQ questions that match users' intent is essential for online customer service. Existing methods aim to fully utilize the dynamic conversation context to enhance the semantic association between the user query and FAQ questions. However, the conversation context contains noise, e.g., users may click questions they don't like, leading to inaccurate semantics modeling. To tackle this, we introduce tags of FAQ questions, which can help us eliminate irrelevant information. We later integrate them into a reinforcement learning framework and minimize the negative impact of irrelevant information in the dynamic conversation context. We experimentally demonstrate our efficiency and effectiveness on conversational FAQ retrieval compared to other baselines.

Dialogue systems are frequently updated to accommodate new services, but naively updating them by continually training with data for new services in diminishing performance on previously learnt services. Motivated by the insight that dialogue state tracking (DST), a crucial component of dialogue systems that estimates the user's goal as a conversation proceeds, is a simple natural language understanding task, we propose reformulating it as a bundle of granular example-guided question answering tasks to minimize the task shift between services and thus benefit continual learning. Our approach alleviates service-specific memorization and teaches a model to contextualize the given question and example to extract the necessary information from the conversation. We find that a model with just 60M parameters can achieve a significant boost by learning to learn from in-context examples retrieved by a retriever trained to identify turns with similar dialogue state changes. Combining our method with dialogue-level memory replay, our approach attains state of the art performance on DST continual learning metrics without relying on any complex regularization or parameter expansion methods.