Thailand

Tokenisation is a core part of language models (LMs). It involves splitting a character sequence into subwords which are assigned random indices before being served to the LM. However, this process---while typically lossless---may lead to less efficient LM training, because it removes character-level information, thereby making it more difficult to generalise across similar subwords, such as *now* and *Now*. We refer to such subwords as **near duplicates**. In this paper, we study the impact of near duplicate subwords on LM training efficiency. First, we design an experiment that gives us an upper bound to how much we should expect a model to improve if we could perfectly generalise across near duplicates. We do this, by duplicating each token in our LM&#39;s vocabulary, creating perfectly equivalent classes of subwords. Experimentally, we find that LMs need roughly 17\% more data when trained in a fully duplicated setting. Second, we investigate the impact of naturally occurring near duplicates on LMs. Here, we see that deduplicating them considerably hurts LM performance; but that this loss in performance can be easily mitigated.

ACL 2024

On the Effect of (Near) Duplicate Subwords in Language Modelling

duplicate

language modelling

tokenization

vocabulary

Tokenisation is a core part of language models (LMs). It involves splitting a character sequence into subwords which are assigned random indices before being served to the LM. However, this process---while typically lossless---may lead to less efficient LM training, because it removes character-level information, thereby making it more difficult to generalise across similar subwords, such as *now* and *Now*. We refer to such subwords as **near duplicates**. In this paper, we study the impact of near duplicate subwords on LM training efficiency. First, we design an experiment that gives us an upper bound to how much we should expect a model to improve if we could perfectly generalise across near duplicates. We do this, by duplicating each token in our LM's vocabulary, creating perfectly equivalent classes of subwords. Experimentally, we find that LMs need roughly 17\% more data when trained in a fully duplicated setting. Second, we investigate the impact of naturally occurring near duplicates on LMs. Here, we see that deduplicating them considerably hurts LM performance; but that this loss in performance can be easily mitigated.

poster

### Welcome!
The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. Our Virtual Poster Sessions will take place online Thursday, August 22, 2024.

You are required to register for this event. **Please register [here](https://2024.aclweb.org/registration). **

If you have already registered, please check your inbox for an email from Underline granting you access to ACL 2024 content.

Please register!

The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. More information will be announced soon.

LLMs have marked a revolutonary shift, yet they falter when faced with compositional reasoning tasks. Our research embarks on a quest to uncover the root causes of compositional reasoning failures of LLMs, uncovering that most of them stem from the improperly generated or leveraged implicit reasoning results. Inspired by our empirical findings, we resort to Logit Lens and an intervention experiment to dissect the inner hidden states of LLMs. This deep dive reveals that implicit reasoning results indeed surface within middle layers and play a causative role in shaping the final explicit reasoning results. Our exploration further locates multi-head self-attention (MHSA) modules within these layers, which emerge as the linchpins in accurate generation and leveraing of implicit reasoning results. Grounded on the above findings, we develop CREME, a lightweight method to patch errors in compositional reasoning via editing the located MHSA modules. Our empirical evidence stands testament to CREME's effectiveness, paving the way for autonomously and continuously enhancing compositional reasoning capabilities in language models.

Understanding and Patching Compositional Reasoning in LLMs

Discourse parsing is a crucial task in natural language processing that aims to reveal the higher-level relations in a text. Despite growing interest in cross-lingual discourse parsing, challenges persist due to limited parallel data and inconsistencies in the Rhetorical Structure Theory (RST) application across languages and corpora. To address this, we introduce a parallel Russian annotation for the large and diverse English GUM RST corpus. Leveraging recent advances, our end-to-end RST parser achieves state-of-the-art results on both English and Russian corpora. It demonstrates effectiveness in both monolingual and bilingual settings, successfully transferring even with limited second-language annotation. To the best of our knowledge, this work is the first to evaluate the potential of cross-lingual end-to-end RST parsing on a manually annotated parallel corpus.

Bilingual Rhetorical Structure Parsing with Large Parallel Annotations

Integrated Speech and Large Language Models (SLMs) that can follow speech instructions and generate relevant text responses have gained popularity lately. However, the safety and robustness of these models remains largely unclear. In this work, we investigate the potential vulnerabilities of such instruction-following speech-language models to adversarial attacks and jailbreaking. Specifically, we design algorithms that can generate adversarial examples to jailbreak SLMs in both white-box and black-box attack settings without human involvement. Additionally, we propose countermeasures to thwart such jailbreaking attacks. Our models, trained on dialog data with speech instructions, achieve state-of-the-art performance on spoken question-answering task, scoring over 80% on both safety and helpfulness metrics. Despite safety guardrails, experiments on jailbreaking demonstrate the vulnerability of SLMs to adversarial perturbations and transfer attacks, with average attack success rates of 90% and 10% respectively when evaluated on a dataset of carefully designed harmful questions spanning 12 different toxic categories. However, we demonstrate that our proposed countermeasures reduce the attack success significantly.

SpeechGuard: Exploring the Adversarial Robustness of Multi-modal Large Language Models

The recent advances in natural language processing have predominantly favored well-resourced English-centric models, resulting in a significant gap with low-resource languages. In this work, we introduce TURNA, a language model developed for the low-resource language Turkish and is capable of both natural language understanding and generation tasks.TURNA is pretrained with an encoder-decoder architecture based on the unified framework UL2 with a diverse corpus that we specifically curated for this purpose. We evaluated TURNA with three generation tasks and five understanding tasks for Turkish. The results show that TURNA outperforms several multilingual models in both understanding and generation tasks and competes with monolingual Turkish models in understanding tasks.

TURNA: A Turkish Encoder-Decoder Language Model for Enhanced Understanding and Generation

Pretrained language models (PLMs) have shown remarkable few-shot learning capabilities when provided with properly formatted examples. However, selecting the "best'' examples remains an open challenge. We propose a complexity-based prompt selection approach for sequence tagging tasks. This approach avoids the training of a dedicated model for selection of examples, and instead uses certain metrics to align the syntactico-semantic complexity of test sentences and examples. We use both sentence- and word-level metrics to match the complexity of examples to the (test) sentence being considered. Our results demonstrate that our approach extracts greater performance from PLMs: it achieves state-of-the-art performance on few-shot NER, achieving a 5\% absolute improvement in F1 score on the CoNLL2003 dataset for GPT-4. We also see large gains of upto 28.85 points (F1/Acc.) in smaller models like GPT-j-6B.

Designing Informative Metrics for Few-Shot Example Selection

Chain-of-thought (COT) prompting can help large language models (LLMs) reason toward correct answers, but its efficacy in reasoning toward incorrect answers is unexplored. This process of elimination (PoE), when used with COT, can enhance self-consistency, interpretability, and tasks such as medical diagnoses of exclusion. Thus, we propose PoE with COT, where LLMs must reason toward incorrect options on multiple-choice questions. We evaluate the ability of GPT-3.5, LLaMA-2, and Falcon to perform PoE with COT on a total of four commonsense and scientific reasoning datasets. We find that the strategy of PoE always underperforms the strategy of choosing the correct answer. The agreement of these strategies is also lower than the self-consistency of each strategy. To study these issues further, we conduct error analyses and give suggestions for future work.

It's Not Easy Being Wrong: Large Language Models Struggle with Process of Elimination Reasoning

Aspect-based Sentiment Analysis (ABSA) is extensively researched in the NLP community, yet related models face challenges due to data sparsity when shifting to a new domain. Hence, data augmentation for cross-domain ABSA has attracted increasing attention in recent years. However, two key points have been neglected in prior studies: First, target domain unlabeled data are labeled with pseudo labels by the model trained in the source domain with little quality control, leading to inaccuracy and error propagation. Second, the label and text patterns of generated labeled data are monotonous, thus limiting the robustness and generalization ability of trained ABSA models. In this paper, we aim to design a simple yet effective framework to address the above shortages in ABSA data augmentation, called Refining and Synthesis Data Augmentation (RSDA). Our framework roughly includes two steps: First, it refines generated labeled data using a natural language inference (NLI) filter to control data quality. Second, it synthesizes diverse labeled data via novel label composition and paraphrase approaches. We conduct experiments on 4 kinds of ABSA subtasks, and our framework outperforms 7 strong baselines, demonstrating its effectiveness.

Refining and Synthesis: A Simple yet Effective Data Augmentation Framework for Cross-Domain Aspect-based Sentiment Analysis

The sound codec's dual roles in minimizing data transmission latency and serving as tokenizers underscore its critical importance.
Recent years have witnessed significant developments in codec models.
The ideal sound codec should preserve content, paralinguistics, speakers, and audio information.
However, the question of which codec achieves optimal sound information preservation remains unanswered, as in different papers, models are evaluated on their selected experimental settings.
This study introduces Codec-SUPERB, an acronym for Codec sound processing Universal PERformance Benchmark.
It is an ecosystem designed to assess codec models across representative sound applications and signal-level metrics rooted in sound domain knowledge.
Codec-SUPERB simplifies result sharing through an online leaderboard, promoting collaboration within a community-driven benchmark database, thereby stimulating new development cycles for codecs.
Furthermore, we undertake an in-depth analysis to offer insights into codec models from both application and signal perspectives, diverging from previous codec papers mainly concentrating on signal-level comparisons.
Finally, we will release codes, the leaderboard, and data to accelerate progress within the community.

Codec-SUPERB: An In-Depth Analysis of Sound Codec Models

Charts provide visual representations of data and are widely used for analyzing information, addressing queries, and conveying insights to others. Various chart-related downstream tasks have emerged recently, such as question-answering and summarization. A common strategy to solve these tasks is to fine-tune various models originally trained on vision tasks language. However, such task-specific models are not capable of solving a wide range of chart-related tasks, constraining their real-world applicability. To overcome these challenges, we introduce ChartInsruct: a novel chart-specific vision-language Instruction-following dataset comprising 191K instructions generated with 71K charts. We then present two distinct systems for instruction tuning on such datasets: (1) an end-to-end model that connects a vision encoder for chart understanding with a LLM; and (2) a pipeline model that employs a two-step approach to extract chart data tables and input them into the LLM. In experiments on four downstream tasks, we first show the effectiveness of our model--achieving a new set of state-of-the-art results. Further evaluation shows that our instruction-tuning approach supports a wide array of real-world chart comprehension and reasoning scenarios, thereby expanding the scope and applicability of our models to new kinds of tasks.

ChartInstruct: Instruction Tuning for Chart Comprehension and Reasoning

Given a source and its edited version performed based on human instructions in natural language, how do we extract the underlying edit operations, to automatically replicate similar edits on other images? This is the problem of reverse designing, and we present TAME-RD, a model to solve this problem. TAME-RD automatically learns from the complex interplay of image editing operations and the natural language instructions to learn fully specified edit operations. It predicts both the underlying image edit operations as discrete categories and their corresponding parameter values in the continuous space.
We accomplish this by mapping together the contextual information from the natural language text and the structural differences between the corresponding source and edited images using the concept of pre-post effect. We demonstrate the efficiency of our network through quantitative evaluations on multiple datasets. We observe improvements of 6-10% on various accuracy metrics and 1.01X-4X on the RMSE score and the concordance correlation coefficient for the corresponding parameter values on the benchmark GIER dataset. We also introduce I-MAD, a new two-part dataset: I-MAD-Dense, a collection of approximately 100K source and edited images, together with automatically generated text instructions and annotated edit operations, and I-MAD-Pro, consisting of about 1.6K source and edited images, together with text instructions and annotated edit operations provided by professional editors. On our dataset, we observe absolute improvements of 1-10% on the accuracy metrics and 1.14X--5X on the RMSE score.

Premium content

On the Effect of (Near) Duplicate Subwords in Language Modelling

Downloads

Next from ACL 2024

Understanding and Patching Compositional Reasoning in LLMs

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES