We introduce Born Again neural Rankers (BAR) in the Learning to Rank (LTR) setting, where student rankers, trained in the Knowledge Distillation (KD) framework, are parameterized identically to their teachers. Unlike the existing ranking distillation work which pursues a good trade-off between performance and efficiency, BAR adapts the idea of Born Again Networks (BAN) to ranking problems and significantly improves ranking performance of students over the teacher rankers without increasing model capacity. The key differences between BAR and common distillation techniques for classification are: (1) an appropriate teacher score transformation function, and (2) a novel listwise distillation framework. Both techniques are specifically designed for ranking problems and are rarely studied in the knowledge distillation literature. Using the state-of-the-art neural ranking structure, BAR is able to push the limits of neural rankers above a recent rigorous benchmark study and significantly outperforms traditionally strong gradient boosted decision tree based models on 7 out of 9 key metrics, the first time in the literature. In addition to the strong empirical results, we give theoretical explanations on why listwise distillation is effective for neural rankers.
The content on the web is in a constant state of flux. New entities, issues, and ideas continuously emerge, while the semantics of the existing conversation topics gradually shift. In recent years, pre-trained language models like BERT greatly improved the state-of-the-art for a large spectrum of content understanding tasks. Therefore, in this paper, we aim to study how these language models can be adapted to better handle continuously evolving web content. In our study, we first analyze the evolution of 2013 - 2019 Twitter data, and unequivocally confirm that a BERT model trained on past tweets would heavily deteriorate when directly applied to data from later years. Then, we investigate two possible sources of the deterioration: the semantic shift of existing tokens and the sub-optimal or failed understanding of new tokens. To this end, we both explore two different vocabulary composition methods, as well as propose three sampling methods which help in efficient incremental training for BERT-like models. Compared to a new model trained from scratch offline, our incremental training (a) reduces the training costs, (b) achieves better performance on evolving content, and (c) is suitable for online deployment. The superiority of our methods is validated using two downstream tasks. We demonstrate significant improvements when incrementally evolving the model from a particular base year, on the task of Country Hashtag Prediction, as well as on the OffensEval 2019 task.
When experiencing an information need, users want to engage with an expert, but often turn to an information retrieval system, such as a search engine, instead. Classical information retrieval systems do not answer information needs directly, but instead provide references to (hopefully authoritative) answers. Successful question answering systems offer a limited corpus created on-demand by human experts, which is neither timely nor scalable. Large pre-trained language models, by contrast, are capable of directly generating prose that may be responsive to an information need, but at present they are dilettantes rather than experts - they do not have a true understanding of the world, they are prone to hallucinating, and crucially they are incapable of justifying their utterances by referring to supporting documents in the corpus they were trained over. This paper examines how ideas from classical information retrieval and large pre-trained language models can be synthesized and evolved into systems that truly deliver on the promise of expert advice.
When trying to apply the recent advance of Natural Language Understanding (NLU) technologies to real-world applications, privacy preservation imposes a crucial challenge, which, unfortunately, has not been well resolved. To address this issue, we study how to improve the effectiveness of NLU models under a Local Privacy setting, using BERT, a widely-used pretrained Language Model (LM), as an example. We systematically study the strengths and weaknesses of imposing dx-privacy, a relaxed variant of Local Differential Privacy, at different stages of language modeling: input text, token embeddings, and sequence representations. We then focus on the former two with privacy-constrained fine-tuning experiments to reveal the utility of BERT under local privacy constraints. More importantly, to the best of our knowledge, we are the first to propose privacy-adaptive LM pretraining methods and demonstrate that they can significantly improve model performance on privatized text input. We also interpret the level of privacy preservation and provide our guidance on privacy parameter selections.
The milestone improvements brought about by deep representation learning and pre-training techniques have led to large performance gains across downstream NLP, IR and Vision tasks. Multimodal modeling techniques aim to leverage large high-quality visio-linguistic datasets for learning complementary information (across image and text modalities). In this paper, we introduce the Wikipedia-based Image Text (WIT) Dataset (https://github.com/google-research-datasets/wit) to better facilitate multimodal, multilingual learning. WIT is composed of a curated set of 37.6 million entity rich image-text examples with 11.5 million unique images across 108 Wikipedia languages. Its size enables WIT to be used as a pretraining dataset for multimodal models, as we show when applied to downstream tasks such as image-text retrieval. WIT has four main and unique advantages. First, WIT is the largest multimodal dataset by the number of image-text examples by 3x (at the time of writing). Second, WIT is massively multilingual (first of its kind) with coverage over 100+ languages (each of which has at least 12K examples) and provides cross-lingual texts for many images. Third, WIT represents a more diverse set of concepts and real world entities relative to what previous datasets cover. Lastly, WIT provides a very challenging real-world test set, as we empirically illustrate using an image-text retrieval task as an example.
Bottom-up algorithms such as the classic hierarchical agglomerative clustering, are highly effective for hierarchical as well as flat clustering. However, the large number of rounds and their sequential nature limit the scalability of agglomerative clustering. In this paper, we present an alternative round-based bottom-up hierarchical clustering, the Sub-Cluster Component Algorithm (SCC), that scales gracefully to massive datasets. Our method builds many sub-clusters in parallel in a given round and requires many fewer rounds -- usually an order of magnitude smaller than classic agglomerative clustering. Our theoretical analysis shows that, under a modest separability assumption, SCC will contain the optimal flat clustering. SCC also provides a 2-approx solution to the DP-means objective, thereby introducing a novel application of hierarchical clustering methods. Empirically, SCC finds better hierarchies and flat clusterings even when the data does not satisfy the separability assumption. We demonstrate the scalability of our method by applying it to a dataset of 30 billion points and showing that SCC produces higher quality clusterings than the state-of-the-art.
Pre-trained models like BERT (Devlin et al., 2018) have dominated NLP / IR applications such as single sentence classification, text pair classification, and question answering. However, deploying these models in real systems is highly non-trivial due to their exorbitant computational costs. A common remedy to this is knowledge distillation (Hinton et al., 2015), leading to faster inference. However -- as we show here -- existing works are not optimized for dealing with pairs (or tuples) of texts. Consequently, they are either not scalable or demonstrate subpar performance. In this work, we propose DiPair -- a novel framework for distilling fast and accurate models on text pair tasks. Coupled with an end-to-end training strategy, DiPair is both highly scalable and offers improved quality-speed tradeoffs. Empirical studies conducted on both academic and real-world e-commerce benchmarks demonstrate the efficacy of the proposed approach with speedups of over 350x and minimal quality drop relative to the cross-attention teacher BERT model.
Consider a sequential active learning problem where, at each round, an agent selects a batch of unlabeled data points, queries their labels and updates a binary classifier. While there exists a rich body of work on active learning in this general form, in this paper, we focus on problems with two distinguishing characteristics: severe class imbalance (skew) and small amounts of initial training data. Both of these problems occur with surprising frequency in many web applications. For instance, detecting offensive or sensitive content in online communities (pornography, violence, and hate-speech) is receiving enormous attention from industry as well as research communities. Such problems have both the characteristics we describe -- a vast majority of content is not offensive, so the number of positive examples for such content is orders of magnitude smaller than the negative examples. Furthermore, there is usually only a small amount of initial training data available when building machine-learned models to solve such problems. To address both these issues, we propose a hybrid active learning algorithm (HAL) that balances exploiting the knowledge available through the currently labeled training examples with exploring the large amount of unlabeled data available. Through simulation results, we show that HAL makes significantly better choices for what points to label when compared to strong baselines like margin-sampling. Classifiers trained on the examples selected for labeling by HAL easily out-perform the baselines on target metrics (like area under the precision-recall curve) given the same budget for labeling examples. We believe HAL offers a simple, intuitive, and computationally tractable way to structure active learning for a wide range of machine learning applications.
Many information retrieval and natural language processing problems can be formalized as a semantic matching task. However, the existing work in this area has been focused in large part on the matching between short texts like finding answer spans, sentences and passages given a query or a natural language question. Semantic matching between long-form texts like documents, which can be applied to applications such as document clustering, news recommendation and related article recommendation, is relatively less explored and needs more research effort. In recent years, self-attention based models like Transformers and BERT have achieved state-of-the-art performance in several natural language understanding tasks. These kinds of models, however, are still restricted to short text sequences like sentences due to the quadratic computational time and space complexity of self-attention with respect to the input sequence length. In this paper, we address these issues by proposing the Siamese Multi-depth Transformer-based Hierarchical (SMITH) Encoder for document representation learning and matching, which contains several novel design choices to adapt self-attention models for long text inputs. For model pre-training, we propose the masked sentence block language modeling task in addition to the original masked word language modeling task used in BERT, to capture sentence block relations within a document. The experimental results on several benchmark data sets for long-form document matching show that our proposed SMITH model outperforms the previous state-of-the-art Siamese matching models including hierarchical attention, multi-depth attention-based hierarchical recurrent neural network, and BERT for long-form document matching, and increases the maximum input text length from 512 to 2048 when compared with BERT-based baseline methods.