Instruction-tuned Large Language Models (LLMs) excel at many tasks and will even explain their reasoning, so-called self-explanations. However, convincing and wrong self-explanations can lead to unsupported confidence in LLMs, thus increasing risk. Therefore, it’s important to measure if self-explanations truly reflect the model’s behavior. Such a measure is called interpretability-faithfulness and is challenging to perform since the ground truth is inaccessible, and many LLMs only have an inference API. To address this, we propose employing self-consistency checks to measure faithfulness. For example, if an LLM says a set of words is important for making a prediction, then it should not be able to make its prediction without these words. While self-consistency checks are a common approach to faithfulness, they have not previously been successfully applied to LLM self-explanations for counterfactual, feature attribution, and redaction explanations. Our results demonstrate that faithfulness is explanation, model, and task-dependent, showing self-explanations should not be trusted in general. For example, with sentiment classification, counterfactuals are more faithful for Llama2, feature attribution for Mistral, and redaction for Falcon 40B.

A recent renewal in interest in long text understanding has sparked the emergence of high-quality long text benchmarks, as well as new models demonstrating significant performance improvements on these benchmarks. However, gauging the implication of these advancements based solely on the length of the input text offers limited insight. Such benchmarks may require models to parse long-range dependencies or merely to locate and comprehend the relevant paragraph within a longer text. This work introduces the Minimal Viable Phrase (MVP), a novel metric that determines, through perturbations to the input text, the shortest average text length that needs to be preserved to execute the task with limited performance degradation. Our evaluation of the popular SCROLLS benchmark reveals that only one of its seven tasks necessitates an MVP of over 512 tokens–the maximum text length manageable by the previous generation of pre-trained models. We highlight the limited need for understanding long-range dependencies in resolving these tasks, discuss the specific design decisions that seem to have led to the QuALITY task requiring reliance on long-range dependencies to be solved, and point out specific modeling choices that seem to outperform on the QuALITY task.

Large language models are first pre-trained on trillions of tokens and then instruction-tuned or aligned to specific preferences. While pre-training remains out of reach for most researchers due to the compute required, fine-tuning has become affordable thanks to parameter-efficient methods such as LoRA and QLoRA. Alignment is known to be sensitive to the many factors involved, including the quantity and quality of data, the alignment method, and the adapter rank. However, there has not yet been an extensive study of their effect on downstream performance. To address this gap, we conduct an in-depth investigation of the impact of popular choices for three crucial axes: (i) the alignment dataset (HH-RLHF and BeaverTails), (ii) the alignment technique (SFT and DPO), and (iii) the model (LLaMA-1, Vicuna-v1.3, Mistral-7b, and Mistral-7b-Instruct). Our extensive setup spanning over 300 experiments reveals consistent trends and unexpected findings. We observe how more informative data helps with preference alignment, cases where supervised fine-tuning outperforms preference optimization, and how aligning to a distinct preference boosts performance on downstream tasks. Through our in-depth analyses, we put forward key guidelines to help researchers perform more effective parameter-efficient LLM alignment.

The widespread use of large language models has brought up essential questions about the potential biases these models might learn. This led to the development of several metrics aimed at evaluating and mitigating these biases. In this paper, we first demonstrate that prompt-based fairness metrics exhibit poor agreement, as measured by correlation, raising important questions about the reliability of fairness assessment using prompts. Then, we outline six relevant reasons why such a low correlation is observed across existing metrics. Based on these insights, we propose a method called Correlated Fairness Output (CAIRO) to enhance the correlation between fairness metrics. CAIRO augments the original prompts of a given fairness metric by using several pre-trained language models and then selects the combination of the augmented prompts that achieves the highest correlation across metrics. We show a significant improvement in Pearson correlation from 0.3 and 0.18 to 0.90 and 0.98 across metrics for gender and religion biases, respectively. Our code is available at https://github.com/chandar-lab/CAIRO.

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.

Language Models (LMs) pre-trained with selfsupervision on large text corpora have become the default starting point for developing models for various NLP tasks. Once the pre-training corpus has been assembled, all data samples in the corpus are treated with equal importance during LM pre-training. However, due to varying levels of relevance and quality of data, equal importance to all the data samples may not be the optimal choice. While data reweighting has been explored in the context of task-specific supervised learning and LM fine-tuning, model-driven reweighting for pretraining data has not been explored. We fill this important gap and propose PRESENCE, a method for jointly reweighting samples by leveraging self-influence (SI) scores as an indicator of sample importance and pre-training. PRESENCE promotes novelty and stability for model pre-training. Through extensive analysis spanning multiple model sizes, datasets, and tasks, we present PRESENCE as an important first step in the research direction of sample reweighting for pre-training language models.

In the age of artificial intelligence, the role of large language models (LLMs) is becoming increasingly central. Despite their growing prevalence, their capacity to consolidate knowledge from different training documents—a crucial ability in numerous applications—remains unexplored. This paper presents the first study examining the capability of LLMs to effectively combine such information within their parameter space. We introduce EpiK-Eval, a novel question-answering benchmark tailored to evaluate LLMs’ proficiency in formulating a coherent and consistent knowledge representation from segmented narratives. Evaluations across various LLMs reveal significant weaknesses in this domain. We contend that these shortcomings stem from the intrinsic nature of prevailing training objectives. Consequently, we advocate for refining the approach towards knowledge consolidation, as it harbors the potential to dramatically improve their overall effectiveness and performance. The findings from this study offer insights for developing more robust and reliable LLMs. Our code and benchmark are available at https://github.com/chandar-lab/EpiK-Eval

Many recent perturbation studies have found unintuitive results on what does and does not matter when performing Natural Language Understanding (NLU) tasks in English. Coding properties, such as the order of words, can often be removed through shuffling without impacting downstream performances. Such insight may be used to direct future research into English NLP models. As many improvements in multilingual settings consist of wholesale adaptation of English approaches, it is important to verify whether those studies replicate or not in multilingual settings. In this work, we replicate a study on the importance of local structure, and the relative unimportance of global structure, in a multilingual setting. We find that the phenomenon observed on the English language broadly translates to over 120 languages, with a few caveats.

Recent research analyzing the sensitivity of natural language understanding models to word-order perturbations has shown that neural models are surprisingly insensitive to the order of words. In this paper, we investigate this phenomenon by developing order-altering perturbations on the order of words, subwords, and characters to analyze their effect on neural models’ performance on language understanding tasks. We experiment with measuring the impact of perturbations to the local neighborhood of characters and global position of characters in the perturbed texts and observe that perturbation functions found in prior literature only affect the global ordering while the local ordering remains relatively unperturbed. We empirically show that neural models, invariant of their inductive biases, pretraining scheme, or the choice of tokenization, mostly rely on the local structure of text to build understanding and make limited use of the global structure.

Providing better language tools for low-resource and endangered languages is imperative for equitable growth.Recent progress with massively multilingual pretrained models has proven surprisingly effective at performing zero-shot transfer to a wide variety of languages.However, this transfer is not universal, with many languages not currently understood by multilingual approaches.It is estimated that only 72 languages possess a “small set of labeled datasets” on which we could test a model’s performance, the vast majority of languages not having the resources available to simply evaluate performances on.In this work, we attempt to clarify which languages do and do not currently benefit from such transfer.To that end, we develop a general approach that requires only unlabelled text to detect which languages are not well understood by a cross-lingual model.Our approach is derived from the hypothesis that if a model’s understanding is insensitive to perturbations to text in a language, it is likely to have a limited understanding of that language.We construct a cross-lingual sentence similarity task to evaluate our approach empirically on 350, primarily low-resource, languages.

Neural models trained for next utterance generation in dialogue task learn to mimic the n-gram sequences in the training set with training objectives like negative log-likelihood (NLL) or cross-entropy. Such commonly used training objectives do not foster generating alternate responses to a context. But, the effects of minimizing an alternate training objective that fosters a model to generate alternate response and score it on semantic similarity has not been well studied. We hypothesize that a language generation model can improve on its diversity by learning to generate alternate text during training and minimizing a semantic loss as an auxiliary objective. We explore this idea on two different sized data sets on the task of next utterance generation in goal oriented dialogues. We make two observations (1) minimizing a semantic objective improved diversity in responses in the smaller data set (Frames) but only as-good-as minimizing the NLL in the larger data set (MultiWoZ) (2) large language model embeddings can be more useful as a semantic loss objective than as initialization for token embeddings.

Predicting the next utterance in dialogue is contingent on encoding of users’ input text to generate appropriate and relevant response in data-driven approaches. Although the semantic and syntactic quality of the language generated is evaluated, more often than not, the encoded representation of input is not evaluated. As the representation of the encoder is essential for predicting the appropriate response, evaluation of encoder representation is a challenging yet important problem. In this work, we showcase evaluating the text generated through human or automatic metrics is not sufficient to appropriately evaluate soundness of the language understanding of dialogue models and, to that end, propose a set of probe tasks to evaluate encoder representation of different language encoders commonly used in dialogue models. From experiments, we observe that some of the probe tasks are easier and some are harder for even sophisticated model architectures to learn. And, through experiments we observe that RNN based architectures have lower performance on automatic metrics on text generation than transformer model but perform better than the transformer model on the probe tasks indicating that RNNs might preserve task information better than the Transformers.

Neural generative models have been become increasingly popular when building conversational agents. They offer flexibility, can be easily adapted to new domains, and require minimal domain engineering. A common criticism of these systems is that they seldom understand or use the available dialog history effectively. In this paper, we take an empirical approach to understanding how these models use the available dialog history by studying the sensitivity of the models to artificially introduced unnatural changes or perturbations to their context at test time. We experiment with 10 different types of perturbations on 4 multi-turn dialog datasets and find that commonly used neural dialog architectures like recurrent and transformer-based seq2seq models are rarely sensitive to most perturbations such as missing or reordering utterances, shuffling words, etc. Also, by open-sourcing our code, we believe that it will serve as a useful diagnostic tool for evaluating dialog systems in the future.

A lot of work has been done in the field of image compression via machine learning, but not much attention has been given to the compression of natural language. Compressing text into lossless representations while making features easily retrievable is not a trivial task, yet has huge benefits. Most methods designed to produce feature rich sentence embeddings focus solely on performing well on downstream tasks and are unable to properly reconstruct the original sequence from the learned embedding. In this work, we propose a near lossless method for encoding long sequences of texts as well as all of their sub-sequences into feature rich representations. We test our method on sentiment analysis and show good performance across all sub-sentence and sentence embeddings.

Neural Module Networks, originally proposed for the task of visual question answering, are a class of neural network architectures that involve human-specified neural modules, each designed for a specific form of reasoning. In current formulations of such networks only the parameters of the neural modules and/or the order of their execution is learned. In this work, we further expand this approach and also learn the underlying internal structure of modules in terms of the ordering and combination of simple and elementary arithmetic operators. We utilize a minimum amount of prior knowledge from the human-specified neural modules in the form of different input types and arithmetic operators used in these modules. Our results show that one is indeed able to simultaneously learn both internal module structure and module sequencing without extra supervisory signals for module execution sequencing. With this approach, we report performance comparable to models using hand-designed modules. In addition, we do a analysis of sensitivity of the learned modules w.r.t. the arithmetic operations and infer the analytical expressions of the learned modules.

Designing of general-purpose learning algorithms is a long-standing goal of artificial intelligence. A general purpose AI agent should be able to have a memory that it can store and retrieve information from. Despite the success of deep learning in particular with the introduction of LSTMs and GRUs to this area, there are still a set of complex tasks that can be challenging for conventional neural networks. Those tasks often require a neural network to be equipped with an explicit, external memory in which a larger, potentially unbounded, set of facts need to be stored. They include but are not limited to, reasoning, planning, episodic question-answering and learning compact algorithms. Recently two promising approaches based on neural networks to this type of tasks have been proposed: Memory Networks and Neural Turing Machines.In this tutorial, we will give an overview of this new paradigm of “neural networks with memory”. We will present a unified architecture for Memory Augmented Neural Networks (MANN) and discuss the ways in which one can address the external memory and hence read/write from it. Then we will introduce Neural Turing Machines and Memory Networks as specific instantiations of this general architecture. In the second half of the tutorial, we will focus on recent advances in MANN which focus on the following questions: How can we read/write from an extremely large memory in a scalable way? How can we design efficient non-linear addressing schemes? How can we do efficient reasoning using large scale memory and an episodic memory? The answer to any one of these questions introduces a variant of MANN. We will conclude the tutorial with several open challenges in MANN and its applications to NLP.We will introduce several applications of MANN in NLP throughout the tutorial. Few examples include language modeling, question answering, visual question answering, and dialogue systems.For updated information and material, please refer to our tutorial website: https://sites.google.com/view/mann-emnlp2017/.

Interlingua based Machine Translation (MT) aims to encode multiple languages into a common linguistic representation and then decode sentences in multiple target languages from this representation. In this work we explore this idea in the context of neural encoder decoder architectures, albeit on a smaller scale and without MT as the end goal. Specifically, we consider the case of three languages or modalities X, Z and Y wherein we are interested in generating sequences in Y starting from information available in X. However, there is no parallel training data available between X and Y but, training data is available between X & Z and Z & Y (as is often the case in many real world applications). Z thus acts as a pivot/bridge. An obvious solution, which is perhaps less elegant but works very well in practice is to train a two stage model which first converts from X to Z and then from Z to Y. Instead we explore an interlingua inspired solution which jointly learns to do the following (i) encode X and Z to a common representation and (ii) decode Y from this common representation. We evaluate our model on two tasks: (i) bridge transliteration and (ii) bridge captioning. We report promising results in both these applications and believe that this is a right step towards truly interlingua inspired encoder decoder architectures.