This paper proposes a transformer over transformer framework, called Transformer$^2$, to perform neural text segmentation. It consists of two components: bottom-level sentence encoders using pre-trained transformers, and an upper-level transformer-based segmentation model based on the sentence embeddings. The bottom-level component transfers the pre-trained knowledge learned from large external corpora under both single and pair-wise supervised NLP tasks to model the sentence embeddings for the documents. Given the sentence embeddings, the upper-level transformer is trained to recover the segmentation boundaries as well as the topic labels of each sentence. Equipped with a multi-task loss and the pre-trained knowledge, Transformer$^2$ can better capture the semantic coherence within the same segments. Our experiments show that (1) Transformer$^2$ manages to surpass state-of-the-art text segmentation models in terms of a commonly-used semantic coherence measure; (2) in most cases, both single and pair-wise pre-trained knowledge contribute to the model performance; (3) bottom-level sentence encoders pre-trained on specific languages yield better performance than those pre-trained on specific domains.
This paper presents an unsupervised extractive approach to summarize scientific long documents based on the Information Bottleneck principle. Inspired by previous work which uses the Information Bottleneck principle for sentence compression, we extend it to document level summarization with two separate steps. In the first step, we use signal(s) as queries to retrieve the key content from the source document. Then, a pre-trained language model conducts further sentence search and edit to return the final extracted summaries. Importantly, our work can be flexibly extended to a multi-view framework by different signals. Automatic evaluation on three scientific document datasets verifies the effectiveness of the proposed framework. The further human evaluation suggests that the extracted summaries cover more content aspects than previous systems.
Continual learning (CL) refers to a machine learning paradigm that using only a small account of training samples and previously learned knowledge to enhance learning performance. CL models learn tasks from various domains in a sequential manner. The major difficulty in CL is catastrophic forgetting of previously learned tasks, caused by shifts in data distributions. The existing CL models often employ a replay-based approach to diminish catastrophic forgetting. Most CL models stochastically select previously seen samples to retain learned knowledge. However, occupied memory size keeps enlarging along with accumulating learned tasks. Hereby, we propose a memory-efficient CL method. We devise a dynamic prototypes-guided memory replay module, incorporating it into an online meta-learning model. We conduct extensive experiments on text classification and additionally investigate the effect of training set orders on CL model performance. The experimental results testify the superiority of our method in alleviating catastrophic forgetting and enabling efficient knowledge transfer.
Graph neural networks (GNNs) are important tools for transductive learning tasks, such as node classification in graphs, due to their expressive power in capturing complex interdependency between nodes. To enable graph neural network learning, existing works typically assume that labeled nodes, from two or multiple classes, are provided, so that a discriminative classifier can be learned from the labeled data. In reality, this assumption might be too restrictive for applications, as users may only provide labels of interest in a single class for a small number of nodes. In addition, most GNN models only aggregate information from short distances (e.g., 1-hop neighbors) in each round, and fail to capture long distance relationship in graphs. In this paper, we propose a novel graph neural network framework, long-short distance aggregation networks (LSDAN), to overcome these limitations. By generating multiple graphs at different distance levels, based on the adjacency matrix, we develop a long-short distance attention model to model these graphs. The direct neighbors are captured via a short-distance attention mechanism, and neighbors with long distance are captured by a long distance attention mechanism. Two novel risk estimators are further employed to aggregate long-short-distance networks, for PU learning and the loss is back-propagated for model learning. Experimental results on real-world datasets demonstrate the effectiveness of our algorithm.
Extreme multi-label classification (XML) is becoming increasingly relevant in the era of big data. Yet, there is no method for effectively generating stratified partitions of XML datasets. Instead, researchers typically rely on provided test-train splits that, 1) aren't always representative of the entire dataset, and 2) are missing many of the labels. This can lead to poor generalization ability and unreliable performance estimates, as has been established in the binary and multi-class settings. As such, this paper presents a new and simple algorithm that can efficiently generate stratified partitions of XML datasets with millions of unique labels. We also examine the label distributions of prevailing benchmark splits, and investigate the issues that arise from using unrepresentative subsets of data for model development. The results highlight the difficulty of stratifying XML data, and demonstrate the importance of using stratified partitions for training and evaluation.
Topic modelling has been a successful technique for text analysis for almost twenty years. When topic modelling met deep neural networks, there emerged a new and increasingly popular research area, neural topic models, with over a hundred models developed and a wide range of applications in neural language understanding such as text generation, summarisation and language models. There is a need to summarise research developments and discuss open problems and future directions. In this paper, we provide a focused yet comprehensive overview of neural topic models for interested researchers in the AI community, so as to facilitate them to navigate and innovate in this fast-growing research area. To the best of our knowledge, ours is the first review focusing on this specific topic.
Knowledge distillation (KD), as an efficient and effective model compression technique, has been receiving considerable attention in deep learning. The key to its success is to transfer knowledge from a large teacher network to a small student one. However, most of the existing knowledge distillation methods consider only one type of knowledge learned from either instance features or instance relations via a specific distillation strategy in teacher-student learning. There are few works that explore the idea of transferring different types of knowledge with different distillation strategies in a unified framework. Moreover, the frequently used offline distillation suffers from a limited learning capacity due to the fixed teacher-student architecture. In this paper we propose a collaborative teacher-student learning via multiple knowledge transfer (CTSL-MKT) that prompts both self-learning and collaborative learning. It allows multiple students learn knowledge from both individual instances and instance relations in a collaborative way. While learning from themselves with self-distillation, they can also guide each other via online distillation. The experiments and ablation studies on four image datasets demonstrate that the proposed CTSL-MKT significantly outperforms the state-of-the-art KD methods.
Few/Zero-shot learning is a big challenge of many classifications tasks, where a classifier is required to recognise instances of classes that have very few or even no training samples. It becomes more difficult in multi-label classification, where each instance is labelled with more than one class. In this paper, we present a simple multi-graph aggregation model that fuses knowledge from multiple label graphs encoding different semantic label relationships in order to study how the aggregated knowledge can benefit multi-label zero/few-shot document classification. The model utilises three kinds of semantic information, i.e., the pre-trained word embeddings, label description, and pre-defined label relations. Experimental results derived on two large clinical datasets (i.e., MIMIC-II and MIMIC-III) and the EU legislation dataset show that methods equipped with the multi-graph knowledge aggregation achieve significant performance improvement across almost all the measures on few/zero-shot labels.
Obtaining training data for multi-document summarization (MDS) is time consuming and resource-intensive, so recent neural models can only be trained for limited domains. In this paper, we propose SummPip: an unsupervised method for multi-document summarization, in which we convert the original documents to a sentence graph, taking both linguistic and deep representation into account, then apply spectral clustering to obtain multiple clusters of sentences, and finally compress each cluster to generate the final summary. Experiments on Multi-News and DUC-2004 datasets show that our method is competitive to previous unsupervised methods and is even comparable to the neural supervised approaches. In addition, human evaluation shows our system produces consistent and complete summaries compared to human written ones.