Pre-trained language models (PLMs) like BERT have made significant progress in various downstream NLP tasks. However, by asking models to do cloze-style tests, recent work finds that PLMs are short in acquiring knowledge from unstructured text. To understand the internal behaviour of PLMs in retrieving knowledge, we first define knowledge-baring (K-B) tokens and knowledge-free (K-F) tokens for unstructured text and ask professional annotators to label some samples manually. Then, we find that PLMs are more likely to give wrong predictions on K-B tokens and attend less attention to those tokens inside the self-attention module. Based on these observations, we develop two solutions to help the model learn more knowledge from unstructured text in a fully self-supervised manner. Experiments on knowledge-intensive tasks show the effectiveness of the proposed methods. To our best knowledge, we are the first to explore fully self-supervised learning of knowledge in continual pre-training.

Learning the distribution of data on Riemannian manifolds is crucial for modeling data from non-Euclidean space, which is required by many applications from diverse scientific fields. Yet, existing generative models on manifolds suffer from expensive divergence computation or rely on approximations of heat kernel. These limitations restrict their applicability to simple geometries and hinder scalability to high dimensions. In this work, we introduce the Riemannian Diffusion Mixture, a principled framework for building a generative process on manifolds as a mixture of endpoint-conditioned diffusion processes instead of relying on the denoising approach of previous diffusion models, for which the generative process is characterized by its drift guiding toward the most probable endpoint with respect to the geometry of the manifold. We further propose a simple yet efficient training objective for learning the mixture process, that is readily applicable to general manifolds. Our method outperforms previous generative models on various manifolds while scaling to high dimensions and requires a dramatically reduced number of in-training simulation steps for general manifolds. Deep generative models have shown great success in learning the distributions of the data represented in Euclidean space, e.g., images and text. While the focus of the previous works has been biased toward data in the Euclidean space, modeling the distribution of the data that naturally resides in the Riemannian manifold with specific geometry has been underexplored, while they are required for wide application: For example, the earth and climate science data (Karpatne et al., 2018; Mathieu & Nickel, 2020) lives in the sphere, whereas the protein structures (Jumper et al., 2021; Watson et al., 2022) and the robotic movements (Simeonov et al., 2022) are best represented by the group SE(3). Moreover, 3D computer graphics shapes (Hoppe et al., 1992) can be identified as a general closed manifold. However, previous generative methods are ill-suited for modeling these data as they do not take into consideration the specific geometry describing the data space and may assign a non-zero probability to regions outside the desired space.

SGD and AdamW are the two most used optimizers for fine-tuning large neural networks in computer vision. When the two methods perform the same, SGD is preferable because it uses less memory (12 bytes/parameter with momentum and 8 bytes/parameter without) than AdamW (16 bytes/parameter). However, on a suite of downstream tasks, especially those with distribution shifts, we find that fine-tuning with AdamW performs substantially better than SGD on modern Vision Transformer and ConvNeXt models. We find that large gaps in performance between SGD and AdamW occur when the fine-tuning gradients in the first "embedding" layer are much larger than in the rest of the model. Our analysis suggests an easy fix that works consistently across datasets and models: freezing the embedding layer (less than 1% of the parameters) leads to SGD with or without momentum performing slightly better than AdamW while using less memory (e.g., on ViT-L, SGD uses 33% less GPU memory). Our insights result in state-of-the-art accuracies on five popular distribution shift benchmarks: WILDS-FMoW, WILDS-Camelyon, BREEDS-Living-17, Waterbirds, and DomainNet.

Question answering over RDF data like knowledge graphs has been greatly advanced, with a number of good systems providing crisp answers for natural language questions or telegraphic queries. Some of these systems incorporate textual sources as additional evidence for the answering process, but cannot compute answers that are present in text alone. Conversely, the IR and NLP communities have addressed QA over text, but such systems barely utilize semantic data and knowledge. This paper presents a method for complex questions that can seamlessly operate over a mixture of RDF datasets and text corpora, or individual sources, in a unified framework. Our method, called UNIQORN, builds a context graph on-the-fly, by retrieving question-relevant evidences from the RDF data and/or a text corpus, using fine-tuned BERT models. The resulting graph typically contains all question-relevant evidences but also a lot of noise. UNIQORN copes with this input by a graph algorithm for Group Steiner Trees, that identifies the best answer candidates in the context graph. Experimental results on several benchmarks of complex questions with multiple entities and relations, show that UNIQORN significantly outperforms state-of-the-art methods for heterogeneous QA -- in a full training mode, as well as in zero-shot settings. The graph-based methodology provides user-interpretable evidence for the complete answering process.

In long context scenarios, large language models (LLMs) face three main challenges: higher computational/financial cost, longer latency, and inferior performance. Some studies reveal that the performance of LLMs depends on both the density and the position of the key information (question relevant) in the input prompt. Inspired by these findings, we propose LongLLMLingua for prompt compression towards improving LLMs' perception of the key information to simultaneously address the three challenges. We conduct evaluation on a wide range of long context scenarios including single-/multi-document QA, few-shot learning, summarization, synthetic tasks, and code completion. The experimental results show that LongLLMLingua compressed prompt can derive higher performance with much less cost. The latency of the end-to-end system is also reduced. For example, on NaturalQuestions benchmark, LongLLMLingua gains a performance boost of up to 17.1% over the original prompt with ~4x fewer tokens as input to GPT-3.5-Turbo. It can derive cost savings of \$28.5 and \$27.4 per 1,000 samples from the LongBench and ZeroScrolls benchmark, respectively. Additionally, when compressing prompts of ~10k tokens at a compression rate of 2x-10x, LongLLMLingua can speed up the end-to-end latency by 1.4x-3.8x. Our code is available at https://aka.ms/LLMLingua.

LLMs have shown great capabilities in various tasks but also exhibited memorization of training data, thus raising tremendous privacy and copyright concerns. While prior work has studied memorization during pre-training, the exploration of memorization during fine-tuning is rather limited. Compared with pre-training, fine-tuning typically involves sensitive data and diverse objectives, thus may bring unique memorization behaviors and distinct privacy risks. In this work, we conduct the first comprehensive analysis to explore LMs' memorization during fine-tuning across tasks. Our studies with open-sourced and our own fine-tuned LMs across various tasks indicate that fine-tuned memorization presents a strong disparity among tasks. We provide an understanding of this task disparity via sparse coding theory and unveil a strong correlation between memorization and attention score distribution. By investigating its memorization behavior, multi-task fine-tuning paves a potential strategy to mitigate fine-tuned memorization.

Maximum-a-posteriori (MAP) decoding is the most widely used decoding strategy for neural machine translation (NMT) models. The underlying assumption is that model probability correlates well with human judgment, with better translations being more likely. However, research has shown that this assumption does not always hold, and decoding strategies which directly optimize a utility function, like Minimum Bayes Risk (MBR) or Quality-Aware decoding can significantly improve translation quality over standard MAP decoding. The main disadvantage of these methods is that they require an additional model to predict the utility, and additional steps during decoding, which makes the entire process computationally demanding. In this paper, we propose to make the NMT models themselves quality-aware by training them to estimate the quality of their own output. During decoding, we can use the model's own quality estimates to guide the generation process and produce the highest-quality translations possible. We demonstrate that the model can self-evaluate its own output during translation, eliminating the need for a separate quality estimation model. Moreover, we show that using this quality signal as a prompt during MAP decoding can significantly improve translation quality. When using the internal quality estimate to prune the hypothesis space during MBR decoding, we can not only further improve translation quality, but also reduce inference speed by two orders of magnitude.

Counterfactually-Augmented Data (CAD) -- minimal editing of sentences to flip the corresponding labels -- has the potential to improve the Out-Of-Distribution (OOD) generalization capability of language models, as CAD induces language models to exploit domain-independent causal features and exclude spurious correlations. However, the empirical results of CAD's OOD generalization are not as efficient as anticipated. In this study, we attribute the inefficiency to the myopia phenomenon caused by CAD: language models only focus on causal features that are edited in the augmentation operation and exclude other non-edited causal features. Therefore, the potential of CAD is not fully exploited. To address this issue, we analyze the myopia phenomenon in feature space from the perspective of Fisher's Linear Discriminant, then we introduce two additional constraints based on CAD's structural properties (dataset-level and sentence-level) to help language models extract more complete causal features in CAD, thereby mitigating the myopia phenomenon and improving OOD generalization capability. We evaluate our method on two tasks: Sentiment Analysis and Natural Language Inference, and the experimental results demonstrate that our method could unlock the potential of CAD and improve the OOD generalization performance of language models by 1.0% to 5.9%.

Constant evolution and the emergence of new cyberattacks require the development of advanced techniques for defense. This paper aims to measure the impact of a supervised filter (classifier) in network anomaly detection. We perform our experiments by employing a hybrid anomaly detection approach in network flow data. For this purpose, we extended a state-of-the-art autoencoder-based anomaly detection method by prepending a binary classifier acting as a prefilter for the anomaly detector. The method was evaluated on the publicly available real-world dataset UGR'16. Our empirical results indicate that the hybrid approach does offer a higher detection rate of known attacks than a standalone anomaly detector while still retaining the ability to detect zero-day attacks. Employing a supervised binary prefilter has increased the AUC metric by over 11%, detecting 30% more attacks while keeping the number of false positives approximately the same.

Automatic speech recognition (ASR) systems are known to be sensitive to the sociolinguistic variability of speech data, in which gender plays a crucial role. This can result in disparities in recognition accuracy between male and female speakers, primarily due to the under-representation of the latter group in the training data. While in the context of hybrid ASR models several solutions have been proposed, the gender bias issue has not been explicitly addressed in end-to-end neural architectures. To fill this gap, we propose a data augmentation technique that manipulates the fundamental frequency (f0) and formants. This technique reduces the data unbalance among genders by simulating voices of the under-represented female speakers and increases the variability within each gender group. Experiments on spontaneous English speech show that our technique yields a relative WER improvement up to 9.87% for utterances by female speakers, with larger gains for the least-represented f0 ranges.