EMNLP 2024 Papers

In the ten years since the development of the Abstract Meaning Representation (AMR) formalism, substantial progress has been made on AMR-related tasks such as parsing and alignment. Still, the engineering applications of AMR are not fully understood. In this survey, we categorize and characterize more than 100 papers which use AMR for downstream tasks— the first survey of this kind for AMR. Specifically, we highlight (1) the range of applications for which AMR has been harnessed, and (2) the techniques for incorporating AMR into those applications. We also detect broader AMR engineering patterns and outline areas of future work that seem ripe for AMR incorporation. We hope that this survey will be useful to those interested in using AMR and that it sparks discussion on the role of symbolic representations in the age of neural-focused NLP research.

The growing size of large language models (LLMs) requires parameter-efficient fine-tuning (PEFT) methods for their adaptation to new tasks. Existing methods, such as Low-Rank Adaptation (LoRA), typically involve model adaptation by training the PEFT parameters. One open problem required to be solved to effectively employ these methods is the identification of PEFT parameters. More precisely, related works identify PEFT parameters by projecting high dimensional parameters of LLMs onto low dimensional parameter manifolds with predefined projections, or identifying PEFT parameters as projections themselves. To study this problem, we propose a new approach called Learning to Efficiently Fine-tune (LEFT) where we aim to learn spaces of PEFT parameters from data. In order to learn how to generate the PEFT parameters on a learned parameter space while fine-tuning the LLMs, we propose the Parameter Generation (PG) method. In the experimental analyses, we examine the effectiveness of our solutions exploring accuracy of fine-tuned LLMs and characteristics of PEFT parameters on benchmark GLUE tasks.

In this paper, we apply a method to quantify biases associated with named entities from various countries. We create counterfactual examples with small perturbations on target-domain data instead of relying on templates or specific datasets for bias detection. On widely used classifiers for subjectivity analysis, including sentiment, emotion, hate speech, and offensive text using Twitter data, our results demonstrate positive biases related to the language spoken in a country across all classifiers studied. Notably, the presence of certain country names in a sentence can strongly influence predictions, up to a 23% change in hate speech detection and up to a 60% change in the prediction of negative emotions such as anger. We hypothesize that these biases stem from the training data of pre-trained language models (PLMs) and find correlations between affect predictions and PLMs likelihood in English and unknown languages like Basque and Maori, revealing distinct patterns with exacerbate correlations. Further, we followed these correlations in-between counterfactual examples from a same sentence to remove the syntactical component, uncovering interesting results suggesting the impact of the pre-training data was more important for English-speaking-country names.

Recently, Large Language Models (LLMs) have demonstrated a superior ability to serve as ranking models. However, concerns have arisen as LLMs will exhibit discriminatory ranking behaviors based on users’ sensitive attributes (gender). Worse still, in this paper, we identify a subtler form of discrimination in LLMs, termed implicit ranking unfairness, where LLMs exhibit discriminatory ranking patterns based solely on non-sensitive user profiles, such as user names. Such implicit unfairness is more widespread but less noticeable, threatening the ethical foundation. To comprehensively explore such unfairness, our analysis will focus on three research aspects: (1) We propose an evaluation method to investigate the severity of implicit ranking unfairness. (2) We uncover the reasons for causing such unfairness. (3) To mitigate such unfairness effectively, we utilize a pair-wise regression method to conduct fair-aware data augmentation for LLM fine-tuning. The experiment demonstrates that our method outperforms existing approaches in ranking fairness, achieving this with only a small reduction in accuracy. Lastly, we emphasize the need for the community to identify and mitigate the implicit unfairness, aiming to avert the potential deterioration in the reinforced human-LLMs ecosystem deterioration.

This work proposes a simple training-free prompt-free approach to leverage large language models (LLMs) for the Chinese spelling correction (CSC) task, which is totally different from all previous CSC approaches. The key idea is to use an LLM as a pure language model in a conventional manner. The LLM goes through the input sentence from the beginning, and at each inference step, produces a distribution over its vocabulary for deciding the next token, given a partial sentence. To ensure that the output sentence remains faithful to the input sentence, we design a minimal distortion model that utilizes pronunciation or shape similarities between the original and replaced characters. Furthermore, we propose two useful reward strategies to address practical challenges specific to the CSC task. Experiments on five public datasets demonstrate that our approach significantly improves LLM performance, enabling them to compete with state-of-the-art domain-general CSC models.

Large language models (LLMs) have demonstrated impressive abilities in generating unstructured natural language according to instructions. However, their performance can be inconsistent when tasked with producing text that adheres to specific structured formats, which is crucial in applications like named entity recognition (NER) or relation extraction (RE). To address this issue, this paper introduces an efficient method, G&O, to enhance their structured text generation capabilities. It breaks the generation into a two-step pipeline: initially, LLMs generate answers in natural language as intermediate responses. Subsequently, LLMs are asked to organize the output into the desired structure, using the intermediate responses as context. G&O effectively separates the generation of content from the structuring process, reducing the pressure of completing two orthogonal tasks simultaneously. Tested on zero-shot NER and RE, the results indicate a significant improvement in LLM performance with minimal additional efforts. This straightforward and adaptable prompting technique can also be combined with other strategies, like self-consistency, to further elevate LLM capabilities in various structured text generation tasks.

The deployment of large language models (LLMs) is often hindered by the extensive memory requirements of the Key-Value (KV) cache, especially as context lengths increase. Existing approaches to reduce the KV cache size involve either fine-tuning the model to learn a compression strategy or leveraging attention scores to reduce the sequence length. We analyse the attention distributions in decoder-only Transformers-based models and observe that attention allocation patterns stay consistent across most layers. Surprisingly, we find a clear correlation between the L2 norm and the attention scores over cached KV pairs, where a low L2 norm of a key embedding usually leads to a high attention score during decoding. This finding indicates that the influence of a KV pair is potentially determined by the key embedding itself before being queried. Based on this observation, we compress the KV cache based on the L2 norm of key embeddings. Our experimental results show that this simple strategy can reduce the KV cache size by 50% on language modelling and needle-in-a-haystack tasks and 90% on passkey retrieval tasks without losing accuracy. Moreover, without relying on the attention scores, this approach remains compatible with FlashAttention, enabling broader applicability.

We present LLoVi, a simple yet effective **L**anguage-based **Lo**ng-range **Vi**deo question-answering (LVQA) framework. Our method decomposes the short- and long-range modeling aspects of LVQA into two stages. First, we use a short-term visual captioner to generate textual descriptions of short video clips (0.5-8 seconds in length) densely sampled from a long input video. Afterward, an LLM aggregates the densely extracted short-term captions to answer a given question. Furthermore, we propose a novel multi-round summarization prompt that asks the LLM first to summarize the noisy short-term visual captions and then answer a given input question. To analyze what makes our simple framework so effective, we thoroughly evaluate various components of our framework. Our empirical analysis reveals that the choice of the visual captioner and LLM is critical for good LVQA performance. The proposed multi-round summarization prompt also leads to a significant LVQA performance boost. Our method achieves the best-reported results on the EgoSchema dataset, best known for very long-form video question-answering. LLoVi also outperforms the previous state-of-the-art by **10.2%** and **6.2%** on NExT-QA and IntentQA for LVQA. Finally, we extend LLoVi to grounded VideoQA, which requires both QA and temporal localization, and show that it outperforms all prior methods on NExT-GQA. Code is available at https://github.com/CeeZh/LLoVi.

Contrastive learning has been successfully adopted in VRL (visual representation learning) by constructing effective contrastive pairs. A promising baseline SimCSE has made notable breakthroughs in unsupervised SRL (sentence representation learning) following the success of contrastive learning. However, considering the difference between VRL and SRL, there is still room for designing a novel contrastive framework specially targeted for SRL. We pro- pose a novel angle-based similarity function for contrastive objective. By examining the gra- dient of our contrastive objective, we show that an angle-based similarity function incites better training dynamics on SRL than the off-the-shelf cosine similarity: (1) effectively pulling a posi- tive instance toward an anchor instance in the early stage of training and (2) not excessively repelling a false negative instance during the middle of training. Our experimental results on widely-utilized benchmarks demonstrate the ef- fectiveness and extensibility of our novel angle- based approach. Subsequent analyses establish its improved sentence representation power.

The interactive theorem prover Lean enables the verification of formal mathematical proofs and is backed by an expanding community. Central to this ecosystem is its mathematical library, mathlib4, which lays the groundwork for the formalization of an expanding range of mathematical theories. However, searching for theorems in mathlib4 can be challenging. To successfully search in mathlib4, users often need to be familiar with its naming conventions or documentation strings. Therefore, creating a semantic search engine that can be used easily by individuals with varying familiarity with mathlib4 is very important. In this paper, we present a semantic search engine for mathlib4 that accepts informal queries and finds the relevant theorems. We also establish a benchmark for assessing the performance of various search engines for mathlib4.

Keyword mnemonics are memorable explanations that link new terms to simpler keywords.Prior work generates mnemonics for students, but they do not train models using mnemonics students prefer and aid learning.We build SMART, a mnemonic generator trained on feedback from real students learning new terms.To train SMART, we first fine-tune LLaMA-2 on a curated set of user-written mnemonics.We then use LLM alignment to enhance SMART: we deploy mnemonics generated by SMART in a flashcard app to find preferences on mnemonics students favor.We gather 2684 preferences from 45 students across two types: **expressed** (inferred from ratings) and **observed** (inferred from student learning), yielding three key findings.First, expressed and observed preferences disagree; what students *think* is helpful does not always capture what is *truly* helpful.Second, Bayesian models can synthesize complementary data from multiple preference types into a single effectiveness signal.SMART is tuned via Direct Preference Optimization on this signal, which resolves ties and missing labels in the typical method of pairwise comparisons, augmenting data for LLM output quality gains. Third, mnemonic experts assess SMART as matching GPT-4 at much lower deployment costs, showing the utility of capturing diverse student feedback to align LLMs in education.

Traditional VQA models are inherently vulnerable to language bias, resulting in a significant performance drop when encountering out-of-distribution datasets. The conventional VQA models suffer from language bias that indicates a spurious correlation between textual questions and answers. Given the outstanding effectiveness of counterfactual causal inference in eliminating bias, we propose a model agnostic dual-debiasing framework based on Counterfactual Causal Effect (DCCE), which explicitly models two types of language bias(i.e., shortcut and distribution bias) by separate branches under the counterfactual inference framework. The effects of both types ofbias on answer prediction can be effectively mitigated by subtracting direct effect of textual questions on answers from total effect ofvisual questions on answers. Experimental results demonstrate that our proposed DCCE framework significantly reduces language biasand achieves state-of-the-art performance on the benchmark datasets without requiring additional augmented data. Our code is available inhttps://github.com/sxycyck/dcce.

In task-oriented dialogue systems, intent detection is crucial for interpreting user queries and providing appropriate responses. Existing research primarily addresses simple queries with a single intent, lacking effective systems for handling complex queries with multiple intents and extracting different intent spans. Additionally, there is a notable absence of multilingual, multi-intent datasets. This study addresses three critical tasks: extracting multiple intent spans from queries, detecting multiple intents, and developing a multilingual multi-label intent dataset. We introduce a novel multi-label multi-class intent detection dataset (MLMCID-dataset) curated from existing benchmark datasets. We also propose a pointer network-based architecture (MLMCID) to extract intent spans and detect multiple intents with coarse and fine-grained labels in the form of sextuplets. Comprehensive analysis demonstrates the superiority of our pointer network based system over baseline approaches in terms of accuracy and F1-score across various datasets.

This study introduces a hypothesis-testing framework to assess whether large language models (LLMs) possess genuine reasoning abilities or primarily depend on token bias. We go beyond evaluating LLMs on accuracy; rather, we aim to investigate their token bias in solving logical reasoning tasks. Specifically, we develop carefully controlled synthetic datasets, featuring conjunction fallacy and syllogistic problems. Our framework outlines a list of hypotheses where token biases are readily identifiable, with all null hypotheses assuming genuine reasoning capabilities of LLMs. The findings in this study suggest, with statistical guarantee, that most LLMs still struggle with logical reasoning. While they may perform well on classic problems, their success largely depends on recognizing superficial patterns with strong token bias, thereby raising concerns about their actual reasoning and generalization abilities.

We evaluate the robustness of several large language models on multiple datasets. Robustness here refers to the relative insensitivity of the model’s answers to meaning-preserving variants of their input. Benchmark datasets are constructed by introducing naturally-occurring, non-malicious perturbations, or by generating semantically equivalent paraphrases of input questions or statements. We further propose a novel metric for assessing a model robustness, and demonstrate its benefits in the non-adversarial scenario by empirical evaluation of several models on the created datasets.

Theory of Mind (ToM) can be used to assess the capabilities of Large Language Models (LLMs) in complex scenarios where social reasoning is required. While the research community has proposed many ToM benchmarks, their hardness varies greatly, and their complexity is not well defined. This work proposes a framework inspired by cognitive load theory to measure the complexity of ToM tasks. We quantify a problem’s complexity as the number of states necessary to solve it correctly. Our complexity measure also accounts for spurious states of a ToM problem designed to make it apparently harder. We use our method to assess the complexity of five widely adopted ToM benchmarks. On top of this framework, we design a prompting technique that augments the information available to a model with a description of how the environment changes with the agents’ interactions. We name this technique Discrete World Models (DWM) and show how it elicits superior performance on ToM tasks.

Conventional Knowledge Graph Reasoning (KGR) models learn the embeddings of KG components over the structure of KGs, but their performances are limited when the KGs are severely incomplete. Recent LLM-enhanced KGR models input KG structural information into LLMs. However, they require fine-tuning on open-source LLMs and are not applicable to closed-source LLMs. Therefore, in this paper, to leverage the knowledge in LLMs without fine-tuning to assist and enhance conventional KGR models, we propose a new three-stage pipeline, including knowledge alignment, KG reasoning and entity reranking. Specifically, in the alignment stage, we propose three strategies to align the knowledge in LLMs to the KG schema by explicitly associating unconnected nodes with semantic relations. Based on the enriched KGs, we train structure-aware KGR models to integrate aligned knowledge to original knowledge existing in KGs. In the reranking stage, after obtaining the results of KGR models, we rerank the top-scored entities with LLMs to recall correct answers further. Experiments show our pipeline can enhance the KGR performance in both incomplete and general situations. Code and datasets are available.

Large Language Models (LLMs) can generate the same sequences contained in the pre-train corpora, known as memorization.Previous research studied it at a macro level, leaving micro yet important questions under-explored, e.g., what makes sentences memorized, the dynamics when generating memorized sequence, its connection to unmemorized sequence, and its predictability.We answer the above questions by analyzing the relationship of memorization with outputs from LLM, namely, embeddings, probability distributions, and generated tokens.A memorization score is calculated as the overlap between generated tokens and actual continuations when the LLM is prompted with a context sequence from the pre-train corpora.Our findings reveal:(1) The inter-correlation between memorized/unmemorized sentences, model size, continuation size, and context size, as well as the transition dynamics between sentences of different memorization scores,(2) A sudden drop and increase in the frequency of input tokens when generating memorized/unmemorized sequences (boundary effect),(3) Cluster of sentences with different memorization scores in the embedding space,(4) An inverse boundary effect in the entropy of probability distributions for generated memorized/unmemorized sequences,(5) The predictability of memorization is related to model size and continuation length. In addition, we show a Transformer model trained by the hidden states of LLM can predict unmemorized tokens.

Contemporary deep learning models effectively handle languages with diverse morphology despite not being directly integrated into them. Morphology and word order are closely linked, with the latter incorporated into transformer-based models through positional encodings. This prompts a fundamental inquiry: Is there a correlation between the morphological complexity of a language and the utilization of positional encoding in pre-trained language models? In pursuit of an answer, we present the first study addressing this question, encompassing 22 languages and 5 downstream tasks. Our findings reveal that the importance of positional encoding diminishes with increasing morphological complexity in languages. Our study motivates the need for a deeper understanding of positional encoding, augmenting them to better reflect the different languages under consideration.