Understanding textual description to generate code seems to be an achieved capability of instruction-following Large Language Models (LLMs) in zero-shot scenario. However, there is a severe possibility that this translation ability may be influenced by having seen target textual descriptions and the related code. This effect is known as Data Contamination. In this study, we investigate the impact of Data Contamination on the performance of GPT-3.5 in the Text-to-SQL code-generating tasks. Hence, we introduce a novel method to detect Data Contamination in GPTs and examine GPT-3.5's Text-to-SQL performances using the known Spider Dataset and our new unfamiliar dataset Termite. Furthermore, we analyze GPT-3.5's efficacy on databases with modified information via an adversarial table disconnection (ATD) approach, complicating Text-to-SQL tasks by removing structural pieces of information from the database. Our results indicate a significant performance drop in GPT-3.5 on the unfamiliar Termite dataset, even with ATD modifications, highlighting the effect of Data Contamination on LLMs in Text-to-SQL translation tasks.
Large Language Models (LLMs) have been demonstrating the ability to solve complex tasks by delivering answers that are positively evaluated by humans due in part to the intensive use of human feedback that refines responses. However, the suggestibility transmitted through human feedback increases the inclination to produce responses that correspond to the user's beliefs or misleading prompts as opposed to true facts, a behaviour known as sycophancy. This phenomenon decreases the bias, robustness, and, consequently, their reliability. In this paper, we shed light on the suggestibility of LLMs to sycophantic behaviour, demonstrating these tendencies via human-influenced prompts over different tasks. Our investigation reveals that LLMs show sycophantic tendencies when responding to queries involving subjective opinions and statements that should elicit a contrary response based on facts, demonstrating a lack of robustness.
Chain-of-Thought (CoT) prompting empowers the reasoning abilities of Large Language Models (LLMs), eliciting them to solve complex reasoning tasks step-by-step. However, with the success of CoT methods, the ability to deliver multi-step reasoning remains limited to English due to the imbalance in the distribution of the pre-training data, making the other languages a barrier. In this work, we propose a Cross-lingual multi-step reasoning approach, aiming to align reasoning processes across different languages. In particular, our method, through a Self-consistent Cross-lingual prompting mechanism inspired by the Tree-of-Thoughts approach, delivers multi-step reasoning paths in different languages that, during the steps, lead to the final solution. Our experimental evaluations show that our method significantly outperforms existing prompting methods, reducing the number of interactions and achieving state-of-the-art performance.
Instruction-tuned Large Language Models (It-LLMs) have been exhibiting outstanding abilities to reason around cognitive states, intentions, and reactions of all people involved, letting humans guide and comprehend day-to-day social interactions effectively. In fact, several multiple-choice questions (MCQ) benchmarks have been proposed to construct solid assessments of the models' abilities. However, earlier works are demonstrating the presence of inherent "order bias" in It-LLMs, posing challenges to the appropriate evaluation. In this paper, we investigate It-LLMs' resilience abilities towards a series of probing tests using four MCQ benchmarks. Introducing adversarial examples, we show a significant performance gap, mainly when varying the order of the choices, which reveals a selection bias and brings into discussion reasoning abilities. Following a correlation between first positions and model choices due to positional bias, we hypothesized the presence of structural heuristics in the decision-making process of the It-LLMs, strengthened by including significant examples in few-shot scenarios. Finally, by using the Chain-of-Thought (CoT) technique, we elicit the model to reason and mitigate the bias by obtaining more robust models.
The language ability of Large Language Models (LLMs) is often unbalanced towards English because of the imbalance in the distribution of the pre-training data. This disparity is demanded in further fine-tuning and affecting the cross-lingual abilities of LLMs. In this paper, we propose to empower Instructiontuned LLMs (It-LLMs) in languages other than English by building semantic alignment between them. Hence, we propose CrossAlpaca, an It-LLM with cross-lingual instruction-following and Translation-following demonstrations to improve semantic alignment between languages. We validate our approach on the multilingual Question Answering (QA) benchmarks XQUAD and MLQA and adapted versions of MMLU and BBH. Our models, tested over six different languages, outperform the It-LLMs tuned on monolingual data. The final results show that instruction tuning on non-English data is not enough and that semantic alignment can be further improved by Translation-following demonstrations.
An outbreak in the popularity of transformer-based Language Models (such as GPT (Brown et al., 2020) and PaLM (Chowdhery et al., 2022)) has opened the doors to new Machine Learning applications. In particular, in Natural Language Processing and how pre-training from large text, corpora is essential in achieving remarkable results in downstream tasks. However, these Language Models seem to have inherent biases toward certain demographics reflected in their training data. While research has attempted to mitigate this problem, existing methods either fail to remove bias altogether, degrade performance, or are expensive. This paper examines the bias produced by promising Language Models when varying parameters and pre-training data. Finally, we propose a de-biasing technique that produces robust de-bias models that maintain performance on downstream tasks.
Pre-trained Language Models such as BERT are impressive machines with the ability to memorize, possibly generalized learning examples. We present here a small, focused contribution to the analysis of the interplay between memorization and performance of BERT in downstream tasks. We propose PreCog, a measure for evaluating memorization from pre-training, and we analyze its correlation with the BERT's performance. Our experiments show that highly memorized examples are better classified, suggesting memorization is an essential key to success for BERT.
The overwhelming success of transformers is a real conundrum stimulating a compelling question: are these machines replicating some traditional linguistic models or discovering radically new theories? In this paper, we propose a novel standpoint to investigate this important question. Using typological similarities among languages, we aim to layer-wise compare transformers for different languages to observe whether these similarities emerge for particular layers. For this investigation, we propose to use Centered kernel alignment to measure similarity among weight matrices. We discovered that syntactic typological similarity is consistent with the similarity among weights in the middle layers. This finding confirms results obtained by syntactically probing BERT and, thus, gives an important confirmation that BERT is replicating traditional linguistic models.
Pre-trained Transformers are challenging human performances in many natural language processing tasks. The gigantic datasets used for pre-training seem to be the key for their success on existing tasks. In this paper, we explore how a range of pre-trained natural language understanding models perform on truly novel and unexplored data, provided by classification tasks over a DarkNet corpus. Surprisingly, results show that syntactic and lexical neural networks largely outperform pre-trained Transformers. This seems to suggest that pre-trained Transformers have serious difficulties in adapting to radically novel texts.
Word embeddings are powerful dictionaries, which may easily capture language variations. However, these dictionaries fail to give sense to rare words, which are surprisingly often covered by traditional dictionaries. In this paper, we propose to use definitions retrieved in traditional dictionaries to produce word embeddings for rare words. For this purpose, we introduce two methods: Definition Neural Network (DefiNNet) and Define BERT (DefBERT). In our experiments, DefiNNet and DefBERT significantly outperform state-of-the-art as well as baseline methods devised for producing embeddings of unknown words. In fact, DefiNNet significantly outperforms FastText, which implements a method for the same task-based on n-grams, and DefBERT significantly outperforms the BERT method for OOV words. Then, definitions in traditional dictionaries are useful to build word embeddings for rare words.