Singapore

While Large Language Models (LLMs) have achieved remarkable performance in many tasks, much about their inner workings remains unclear. In this study, we present novel experimental insights into the resilience of LLMs, particularly GPT-4, when subjected to extensive character-level permutations. To investigate this, we first propose the Scrambled Bench, a suite designed to measure the capacity of LLMs to handle scrambled input, in terms of both recovering scrambled sentences and answering questions given scrambled context. The experimental results indicate that multiple advanced LLMs demonstrate the capability akin to typoglycemia, a phenomenon where humans can understand the meaning of words even when the letters within those words are scrambled, as long as the first and last letters remain in place. More surprisingly, we found that only GPT-4 nearly flawlessly processes inputs with unnatural errors, a task significantly challenging for other LLMs and even humans. Specifically, GPT-4 can almost perfectly reconstruct the original sentences from scrambled ones, decreasing the edit distance by 95%, even when all letters within each word are entirely scrambled. It is counter-intuitive that LLMs can exhibit such resilience despite severe disruption to input tokenization caused by scrambled text.

EMNLP 2023

Unnatural Error Correction: GPT-4 Can Almost Perfectly Handle Unnatural Scrambled Text

scrambled text

emergent ability

gpt-4

large language models

poster

**Welcome to EMNLP 2023!**

On behalf of the EMNLP 2023 Organizing Committee, I extend a warm and heartfelt welcome to all of you to the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). It is with immense pleasure and excitement that we gather here in Singapore, a vibrant hub of innovation and technological advancement.

The conference program is packed with insightful presentations, thought-provoking workshops, and engaging networking opportunities. In addition to the technical sessions, we have also planned several social events that will provide you with the opportunity to connect with your colleagues.

The conference is held in person at the Resorts World Convention Centre in Singapore, and available on-line with the help of Underline.

For those in Singapore, I hope that you will find time to explore the exciting Sentosa Island, Gardens by the Bay, Singapore Botanic Gardens and the many other attractions unique to Singapore.

Of course, we are grateful to our sponsors and partners for their generous support of EMNLP 2023, whose contributions make it possible for us to host this world-class event.

Yuji Matsumoto (RIKEN AIP) 
EMNLP 2023 General Chair

To access the **EMNLP 2023** event page on Underline, you need to register for the Conference. 
Please follow **[this link](https://2023.emnlp.org/registration/)** for more details.

EMNLP 2023 took place in Singapore from Dec 6th to Dec 10th, 2023.

A popular approach to streaming speech translation is to employ a single offline model with a wait-k policy to support different latency requirements, which is simpler than training multiple online models with different latency constraints. However, there is a mismatch problem in using a model trained with complete utterances for streaming inference with partial input. We demonstrate that speech representations extracted at the end of a streaming input are significantly different from those extracted from a complete utterance. To address this issue, we propose a new approach called Future-Aware Streaming Translation (FAST) that adapts an offline ST model for streaming input. FAST includes a Future-Aware Inference (FAI) strategy that incorporates future context through a trainable masked embedding, and a Future-Aware Distillation (FAD) framework that transfers future context from an approximation of full speech to streaming input. Our experiments on the MuST-C EnDe, EnEs, and EnFr benchmarks show that FAST achieves better trade-offs between translation quality and latency than strong baselines. Extensive analyses suggest that our methods effectively alleviate the aforementioned mismatch problem between offline training and online inference.

Adapting Offline Speech Translation Models for Streaming with Future-Aware Distillation and Inference | VIDEO

Over-generation errors occur when a keyphrase extraction model correctly determines a candidate keyphrase as a keyphrase because it contains a word that frequently appears in the document but at the same time erroneously outputs other candidates as keyphrases because they contain the same word. To mitigate this issue, we propose a new heterogeneous centrality detection approach (CentralityRank), which extracts keyphrases by simultaneously identifying both implicit and explicit centrality within a heterogeneous graph as the importance score of each candidate. More specifically, CentralityRank detects centrality by taking full advantage of the content within the input document to construct graphs that encompass semantic nodes of varying granularity levels, not limited to just phrases. These additional nodes act as intermediaries between candidate keyphrases, enhancing cross-phrase relations. Furthermore, we introduce a novel adaptive boundary-aware regularization that can leverage the position information of candidate keyphrases, thus influencing the importance of candidate keyphrases. Extensive experimental results demonstrate the superiority of CentralityRank over recent state-of-the-art unsupervised keyphrase extraction baselines across three benchmark datasets.

Mitigating Over-Generation for Unsupervised Keyphrase Extraction with Heterogeneous Centrality Detection

Sarcasm is a complex linguistic construct with incongruity at its very core. Detecting sarcasm depends on the actual content spoken and tonality, facial expressions, the context of an utterance, and personal traits like language proficiency and cognitive capabilities. In this paper, we propose the utilization of synthetic gaze data to improve the task performance for multimodal sarcasm detection in a conversational setting. We enrich an existing multimodal conversational dataset, i.e., MUStARD++ with gaze features. With the help of human participants, we collect gaze features for 20% of data instances, and we investigate various methods for gaze feature prediction for the rest of the dataset. We perform extrinsic and intrinsic evaluations to assess the quality of the predicted gaze features. We observe a performance gain of up to 6.6% points by adding a new modality, i.e., collected gaze features. When both collected and predicted data are used, we observe a performance gain of 2.3% points on the complete dataset. Interestingly, with only predicted gaze features, too, we observe a gain in performance (1.9% points). We retain and use the feature prediction model, which maximally correlates with collected gaze features. Our model trained on combining collected and synthetic gaze data achieves SoTA performance on the MUStARD++ dataset. To the best of our knowledge, ours is the first predict-and-use model for sarcasm detection. We publicly release the code, gaze data, and our best models for further research.

Predict and Use: Harnessing Predicted Gaze to Improve Multimodal Sarcasm Detection

Artificial learners often behave differently from human learners in the context of neural agent-based simulations of language emergence and change. A common explanation is the lack of appropriate cognitive biases in these learners. However, it has also been proposed that more naturalistic settings of language learning and use could lead to more human-like results. We investigate this latter account focusing on the word-order/case-marking trade-off, a widely attested language universal that has proven particularly hard to simulate. We propose a new Neural-agent Language Learning and Communication framework (NeLLCom) where pairs of speaking and listening agents first learn a miniature language via supervised learning, and then optimize it for communication via reinforcement learning. Following closely the setup of earlier human experiments, we succeed in replicating the trade-off with the new framework without hard-coding specific biases in the agents. We see this as an essential step towards the investigation of language universals with neural learners.

Communication Drives the Emergence of Language Universals in Neural Agents: Evidence from the Word-order/Case-marking Trade-off" | VIDEO

Hallucinated translations can severely undermine and raise safety issues when machine translation systems are deployed in the wild. Previous research on the topic focused on small bilingual models trained on high-resource languages, leaving a gap in our understanding of hallucinations in multilingual models across diverse translation scenarios. In this work, we fill this gap by conducting a comprehensive analysis -- over 100 language pairs across various resource levels and going beyond English-centric directions -- on both the M2M neural machine translation (NMT) models and GPT large language models (LLMs). Among several insights, we highlight that models struggle with hallucinations primarily in low-resource directions and when translating out of English, where, critically, they may reveal toxic patterns that can be traced back to the training data. We also find that LLMs produce qualitatively different hallucinations to those of NMT models. Finally, we show that hallucinations are hard to reverse by merely scaling models trained with the same data. However, employing more diverse models, trained on different data or with different procedures, as fallback systems can improve translation quality and virtually eliminate certain pathologies.

Hallucinations in Large Multilingual Translation Models | VIDEO

Recent researches reveal that pre-trained models (PTMs) are vulnerable to backdoor attacks before the fine-tuning stage. The attackers can implant transferable task-agnostic backdoors in PTMs, and control model outputs on any downstream task, which poses severe security threats to all downstream applications. Existing backdoor-removal defenses focus on task-specific classification models and they are not suitable for defending PTMs against task-agnostic backdoor attacks. To this end, we propose the first task-agnostic backdoor removal method for PTMs. Based on the selective activation phenomenon in backdoored PTMs, we design a simple and effective backdoor eraser, which continually pre-trains the backdoored PTMs with a regularization term in an end-to-end approach. The regularization term removes backdoor functionalities from PTMs while the continual pre-training maintains the normal functionalities of PTMs. We conduct extensive experiments on pre-trained models across different modalities and architectures. The experimental results show that our method can effectively remove backdoors inside PTMs and preserve benign functionalities of PTMs with a few downstream-task-irrelevant auxiliary data, e.g., unlabeled plain texts. The average attack success rate on three downstream datasets is reduced from 99.88% to 8.10% after our defense on the backdoored BERT. The codes are publicly available at https://github.com/thunlp/RECIPE.

Removing Backdoors in Pre-trained Models by Regularized Continual Pre-training

Linguistic variation across a region of interest can be captured by partitioning the region into areas and using social media data to train embeddings that represent language use in those areas. Recent work has focused on larger areas, such as cities or counties, to ensure that enough social media data is available in each area, but larger areas have a limited ability to find fine-grained distinctions, such as intracity differences in language use. We demonstrate that it is possible to embed smaller areas, which can provide higher resolution analyses of language variation. We embed voting precincts, which are tiny, evenly sized political divisions for the administration of elections. The issue with modeling language use in small areas is that the data becomes incredibly sparse, with many areas having scant social media data. We propose a novel embedding approach that alternates training with smoothing, which mitigates these sparsity issues. We focus on linguistic variation across Texas as it is relatively understudied. We developed two novel quantitative evaluations that measure how well the embeddings can be used to capture linguistic variation. The first evaluation measures how well a model can map a dialect given terms specific to that dialect. The second evaluation measures how well a model can map preference of lexical variants. These evaluations show how embedding models could be used directly by sociolinguists and measure how much sociolinguistic information is contained within the embeddings. We complement this second evaluation with a methodology for using embeddings as a kind of genetic code where we identify “genes” that correspond to a sociological variable and connect those “genes” to a linguistic phenomenon thereby connecting sociological phenomena to linguistic ones. Finally, we explore approaches for inferring isoglosses using embeddings.


Capturing Fine-Grained Regional Differences in Language Use through Voting Precinct Embeddings

The utilization of monolingual data has been shown to be a promising strategy for addressing low-resource machine translation problems. Previous studies have demonstrated the effectiveness of techniques such as Back-Translation and self-supervised objectives, including Masked Language Modeling, Causal Language Modeling, and Denoise Autoencoding, in improving the performance of machine translation models. However, the manner in which these methods contribute to the success of machine translation tasks and how they can be effectively combined remains an under-researched area. In this study, we carry out a systematic investigation of the effects of these techniques on linguistic properties through the use of probing tasks, including source language comprehension, bilingual word alignment, and translation fluency. We further evaluate the impact of Pre-Training, Back-Translation, and Multi-Task Learning on bitexts of varying sizes. Our findings inform the design of more effective pipelines for leveraging monolingual data in extremely low-resource and low-resource machine translation tasks. Experiment results show consistent performance gains in seven translation directions, which provide further support for our conclusions and understanding of the role of monolingual data in machine translation.

Rethinking the Exploitation of Monolingual Data for Low-Resource Neural Machine Translation

Figures of speech such as metaphors, similes, and idioms are integral parts of human communication. They are ubiquitous in many forms of discourse, allowing people to convey complex, abstract ideas and evoke emotion. As figurative forms are often conveyed through multiple modalities (e.g., both text and images), understanding multimodal figurative language is an important AI challenge, weaving together profound vision, language, commonsense and cultural knowledge. 

In this work, we develop the Image Recognition of Figurative Language (IRFL) dataset. We leverage human annotation and an automatic pipeline we created to generate a multimodal dataset, and introduce two novel tasks as a benchmark for multimodal figurative language understanding. We experimented with state-of-the-art vision and language models and found that the best (22%) performed substantially worse than humans (97%). We release our dataset, benchmark, and code in hopes of driving the development of models that can better understand figurative language.

IRFL: Image Recognition of Figurative Language

Recent vision-language (VL) models are powerful, but can they reliably distinguish "right" from "left"? We curate three new corpora to quantify model comprehension of such basic spatial relations. These tests isolate spatial reasoning more precisely than existing datasets like VQAv2, e.g., our What'sUp benchmark contains sets of photographs varying only the spatial relations of objects, keeping their identity fixed (see Figure 1: models must comprehend not only the usual case of a dog under a table, but also, the same dog on top of the same table). We evaluate 18 VL models, finding that all perform poorly, e.g., BLIP finetuned on VQAv2, which nears human parity on VQAv2, achieves 56% accuracy on our benchmarks vs. humans at 99%. We conclude by studying causes of this surprising behavior, finding: 1) that popular vision-language pretraining corpora like LAION-2B contain little reliable data for learning spatial relationships; and 2) that basic modeling interventions like up-weighting preposition-containing instances or fine-tuning on our corpora are not sufficient to address the challenges our benchmarks pose. We are hopeful that these corpora will facilitate further research, and we release our data and code at https://github.com/amitakamath/whatsup_vlms.

Premium content

Unnatural Error Correction: GPT-4 Can Almost Perfectly Handle Unnatural Scrambled Text

Downloads

Next from EMNLP 2023

Adapting Offline Speech Translation Models for Streaming with Future-Aware Distillation and Inference | VIDEO

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES