Transformer language models (LMs) have been shown to represent concepts as directions in the latent space of hidden activations. However, for any given human-interpretable concept, how can we find its direction in the latent space? We present a technique called linear relational concepts (LRC) for finding concept directions corresponding to human-interpretable concepts at a given hidden layer in a transformer LM by first modeling the relation between subject and object as a linear relational embedding (LRE). While the LRE work was mainly presented as an exercise in understanding model representations, we find that inverting the LRE while using earlier object layers results in a powerful technique to find concept directions that both work well as a classifier and causally influence model outputs.

In the rapidly evolving landscape of human-computer interaction, the integration of vision capabilities into conversational agents stands as a crucial advancement. This paper presents an initial implementation of a dialogue manager that leverages the latest progress in Large Language Models (e.g., GPT-4, IDEFICS) to enhance the traditional text-based prompts with real-time visual input. LLMs are used to interpret both textual prompts and visual stimuli, creating a more contextually aware conversational agent. The system's prompt engineering, incorporating dialogue with summarisation of the images, ensures a balance between context preservation and computational efficiency. Six interactions with a Furhat robot powered by this system are reported, illustrating and discussing the results obtained. By implementing this vision-enabled dialogue system, the paper envisions a future where conversational agents seamlessly blend textual and visual modalities, enabling richer, more context-aware dialogues.

One way that the current state of the art measures the reasoning ability of transformer-based models is by evaluating accuracy in downstream tasks like logical question answering or proof generation over synthetic contexts expressed in natural language. However, most of the contexts used are in practice very simple; in most cases, they are generated from short first-order logic sentences with only a few logical operators and quantifiers. In this work, we seek to answer the question how well a transformer-based model will perform reasoning over expressive contexts. For this purpose, we construct a synthetic natural language question-answering dataset, generated by description logic knowledge bases. For the generation of the knowledge bases, we use the expressive language $\mathcal{ALCQ}$. The resulting dataset contains 384K examples, and increases in two dimensions: i) reasoning depth, and ii) length of sentences. We show that the performance of our DeBERTa-based model, DELTA$_M$, is marginally affected when the reasoning depth is increased and it is not affected at all when the length of the sentences is increasing. We also evaluate the generalization ability of the model on reasoning depths unseen at training, both increasing and decreasing, revealing interesting insights into the model's adaptive generalization abilities.

Exploring the application of powerful large language models (LLMs) on the fundamental named entity recognition (NER) task has drawn much attention recently. This work aims to investigate the possibilities of pushing the boundary of zero-shot NER with LLM via a training-free self-improving strategy. We propose a self-improving framework, which utilize an unlabeled corpus to stimulate the self-learning ability of LLMs on NER. First, we use LLM to make predictions on the unlabeled corpus and obtain the self-annotated data. Second, we explore various strategies to select reliable samples from the self-annotated dataset as demonstrations, considering the similarity, diversity and reliability of demonstrations. Finally, we conduct inference for the test query via in-context learning with the selected self-annotated demonstrations. Through comprehensive experimental analysis, our study yielded the following findings: (1) The self-improving framework further pushes the boundary of zero-shot NER with LLMs, and achieves an obvious performance improvement; (2) Iterative self-improving or naively increasing the size of unlabeled corpus does not guarantee improvements; (3) There might still be space for improvement via more advanced strategy for reliable entity selection.

We address the zero-shot transfer learning setting for the knowledge base question answering (KBQA) problem, where a large volume of labeled training data is available for the source domain, but no such labeled examples are available for the target domain. Transfer learning for KBQA makes use of large volumes of unlabeled data in the target in addition to the labeled data in the source. More recently, few-shot in-context learning using Black-box Large Language Models (BLLMs) has been adapted for KBQA without considering any source domain data. In this work, we show how to meaningfully combine these two paradigms for KBQA so that their benefits add up. Specifically, we preserve the two stage retrieve-then-generate pipeline of supervised KBQA and introduce interaction between in-context learning using BLLMs and transfer learning from the source for both stages. In addition, we propose execution-guided self-refinement using BLLMs, decoupled from the transfer setting. With the help of experiments using benchmark datasets GrailQA as the source and WebQSP as the target, we show that the proposed combination brings significant improvements to both stages and also outperforms by a large margin state-of-the-art supervised KBQA models trained on the source. We also show that in the in-domain setting, the proposed BLLM augmentation significantly outperforms state-of-the-art supervised models, when the volume of labeled data is limited, and also outperforms these marginally even when using the entire large training dataset.

Realizing the recent advances in Natural Language Processing (NLP) to the legal sector poses challenging problems such as extremely long sequence lengths, specialized vocabulary that is usually only understood by legal professionals, and high amounts of data imbalance. The recent surge of Large Language Models (LLMs) has begun to provide new opportunities to apply NLP in the legal domain due to their ability to handle lengthy, complex sequences. Moreover, the emergence of domain-specific LLMs has displayed extremely promising results on various tasks. In this study, we aim to quantify how general LLMs perform in comparison to legal-domain models (be it an LLM or otherwise). Specifically, we compare the zero-shot performance of three general-purpose LLMs (ChatGPT-20b, LLaMA-2-70b, and Falcon-180b) on the LEDGAR subset of the LexGLUE benchmark for contract provision classification. Although the LLMs were not explicitly trained on legal data, we observe that they are still able to classify the theme correctly in most cases. However, we find that their mic-F1/mac-F1 performance is up to 19.2/26.8\% lesser than smaller models fine-tuned on the legal domain, thus underscoring the need for more powerful legal-domain LLMs.

Large language models (LLMs) have shown incredible performance in completing various real-world tasks. The current knowledge learning paradigm of LLMs is mainly based on learning from examples, in which LLMs learn the internal rule implicitly from a certain number of supervised examples. However, the learning paradigm may not well learn those complicated rules, especially when the training examples are limited. We are inspired that humans can learn the new tasks or knowledge in another way by learning from rules. That is, humans can grasp the new tasks or knowledge quickly and generalize well given only a detailed rule and a few optional examples. Therefore, in this paper, we aim to explore the feasibility of this new learning paradigm, which encodes the rule-based knowledge into LLMs. We propose rule distillation, which first uses the strong in-context abilities of LLMs to extract the knowledge from the textual rules and then explicitly encode the knowledge into LLMs' parameters by learning from the above in-context signals produced inside the model. Our experiments show that making LLMs learn from rules by our method is much more efficient than example-based learning in both the sample size and generalization ability.

To maintain user trust, large language models (LLMs) should signal low confidence on examples where they are incorrect, instead of misleading the user. The standard approach of estimating confidence is to use the softmax probabilities of these models, but as of November 2023, state-of-the-art LLMs such as GPT-4 and Claude-v1.3 We first study eliciting confidence linguistically -- asking an LLM for its confidence in its answer -- which performs reasonably (80.5% AUC on GPT-4 averaged across 12 question-answering datasets -- 7% above a random baseline) but leaves room for improvement. We then explore using a surrogate confidence model -- using a model where we do have probabilities to evaluate the original model's confidence in a given question. Surprisingly, even though these probabilities come from a different and often weaker model, this method leads to higher AUC than linguistic confidences on 9 out of 12 datasets. Our best method composing linguistic confidences and surrogate model probabilities gives state-of-the-art confidence estimates on all 12 datasets (84.6% average AUC on GPT-4). As large language models (LLMs) are increasingly deployed, it is important that they signal low confidence on examples where they are likely to make mistakes. This paper's goal is to produce good confidence estimates for state-of-the-art LLMs, which do not provide model probabilities or representations (such as GPT-4 and Claude-v1.3). We first examine a natural idea of eliciting linguistic confidence scores (Tian et al., 2023; Lin et al., 2022; Xiong et al., 2023) -- prompting the LLM to assess its confidence in its answer (Figure 1, GPT-4 Linguistic). We find that linguistic confidences work reasonably well for state-of-the-art models, and much better than a random guessing baseline, but still leave room for improvement (Section 3). Averaged across the datasets, GPT-4 achieves a selective classification AUC of 80.5%, which is 7% above a random guessing baseline. Our results hold across 12 standard datasets (8 MMLU datasets, TruthfulQA, CommonsenseQA, OpenbookQA, and MedQA), 5 models (GPT-4, Claude-v1.3,

Pretraining multilingual language models from scratch requires considerable computational resources and substantial training data. Therefore, a more efficient method is to adapt existing pretrained language models (PLMs) to new languages via vocabulary extension and continued pretraining. However, this method usually randomly initializes the embeddings of new subwords and introduces substantially more embedding parameters to the language model, thus weakening the efficiency. To address these issues, we propose a novel framework: \textbf{O}ne \textbf{F}or \textbf{A}ll (\textbf{\textsc{Ofa}}), which wisely initializes the embeddings of unseen subwords from target languages and thus can adapt a PLM to multiple languages efficiently and effectively. \textsc{Ofa} takes advantage of external well-aligned multilingual word embeddings and injects the alignment knowledge into the new embeddings. In addition, \textsc{Ofa} applies matrix factorization and replaces the cumbersome embeddings with two lower-dimensional matrices, which significantly reduces the number of parameters while not sacrificing the performance. Through extensive experiments, we show models initialized by \textsc{Ofa} are efficient and outperform several baselines. \textsc{Ofa} not only accelerates the convergence of continued pretraining, which is friendly to a limited computation budget, but also improves the zero-shot crosslingual transfer on a wide range of downstream tasks. We make our code and models publicly available.

Although image captioning has a vast array of applications, it has not reached its full potential in languages other than English. Arabic, for instance, although the native language of more than 400 million people, remains largely underrepresented in this area. This is due to the lack of labeled data and powerful Arabic generative models. We alleviate this issue by presenting a novel vision-language model dedicated to Arabic, dubbed \textit{Violet}. Our model is based on a vision encoder and a Gemini text decoder that maintains generation fluency while allowing fusion between the vision and language components. To train our model, we introduce a new method for automatically acquiring data from available English datasets. We also manually prepare a new dataset for evaluation. \textit{Violet} performs sizeably better than our baselines on all of our evaluation datasets. For example, it reaches a CIDEr score of $61.2$ on our manually annotated dataset and achieves an improvement of $13$ points on Flickr8k.