By forcing at most N out of M consecutive weights to be non-zero, the recent N:M network sparsity has received increasing attention for its two attractive advantages: 1) Promising performance at a high sparsity. 2) Significant speedups on NVIDIA A100 GPUs. Recent studies require an expensive pre-training phase or a heavy dense-gradient computation. In this paper, we show that the N:M learning can be naturally characterized as a combinatorial problem which searches for the best combination candidate within a finite collection. Motivated by this characteristic, we solve N:M sparsity in an efficient divide-and-conquer manner. First, we divide the weight vector into $C_{\text{M}}^{\text{N}}$ combination subsets of a fixed size N. Then, we conquer the combinatorial problem by assigning each combination a learnable score that is jointly optimized with its associate weights. We prove that the introduced scoring mechanism can well model the relative importance between combination subsets. And by gradually removing low-scored subsets, N:M fine-grained sparsity can be efficiently optimized during the normal training phase. Comprehensive experiments demonstrate that our learning best combination (LBC) performs consistently better than off-the-shelf N:M sparsity methods across various networks. Our code is released at \url{https://github.com/zyxxmu/LBC}.
Much of the progress in contemporary NLP has come from learning representations, such as masked language model (MLM) contextual embeddings, that turn challenging problems into simple classification tasks. But how do we quantify and explain this effect? We adapt general tools from computational learning theory to fit the specific characteristics of text datasets and present a method to evaluate the compatibility between representations and tasks. Even though many tasks can be easily solved with simple bag-of-words (BOW) representations, BOW does poorly on hard natural language inference tasks. For one such task we find that BOW cannot distinguish between real and randomized labelings, while pre-trained MLM representations show 72x greater distinction between real and random labelings than BOW. This method provides a calibrated, quantitative measure of the difficulty of a classification-based NLP task, enabling comparisons between representations without requiring empirical evaluations that may be sensitive to initializations and hyperparameters. The method provides a fresh perspective on the patterns in a dataset and the alignment of those patterns with specific labels.
Large-scale single-stream pre-training has shown dramatic performance in image-text retrieval. Regrettably, it faces low inference efficiency due to heavy attention layers. Recently, two-stream methods like CLIP and ALIGN with high inference efficiency have also shown promising performance, however, they only consider instance-level alignment between the two streams (thus there is still room for improvement). To overcome these limitations, we propose a novel COllaborative Two-Stream vision-language pretraining model termed COTS for image-text retrieval by enhancing cross-modal interaction. In addition to instance level alignment via momentum contrastive learning, we leverage two extra levels of cross-modal interactions in our COTS: (1) Token-level interaction - a masked visionlanguage modeling (MVLM) learning objective is devised without using a cross-stream network module, where variational autoencoder is imposed on the visual encoder to generate visual tokens for each image. (2) Task-level interaction - a KL-alignment learning objective is devised between text-to-image and image-to-text retrieval tasks, where the probability distribution per task is computed with the negative queues in momentum contrastive learning. Under a fair comparison setting, our COTS achieves the highest performance among all two-stream methods and comparable performance (but with 10,800X faster in inference) w.r.t. the latest single-stream methods. Importantly, our COTS is also applicable to text-to-video retrieval, yielding new state-ofthe-art on the widely-used MSR-VTT dataset.
The design of complex engineering systems is an often long and articulated process that highly relies on engineers' expertise and professional judgment. As such, the typical pitfalls of activities involving the human factor often manifest themselves in terms of lack of completeness or exhaustiveness of the analysis, inconsistencies across design choices or documentation, as well as an implicit degree of subjectivity. An approach is proposed to assist systems engineers in the automatic generation of systems diagrams from unstructured natural language text. Natural Language Processing (NLP) techniques are used to extract entities and their relationships from textual resources (e.g., specifications, manuals, technical reports, maintenance reports) available within an organisation, and convert them into Systems Modelling Language (SysML) diagrams, with particular focus on structure and requirement diagrams. The intention is to provide the users with a more standardised, comprehensive and automated starting point onto which subsequently refine and adapt the diagrams according to their needs. The proposed approach is flexible and open-domain. It consists of six steps which leverage open-access tools, and it leads to an automatic generation of SysML diagrams without intermediate modelling requirement, but through the specification of a set of parameters by the user. The applicability and benefits of the proposed approach are shown through six case studies having different textual sources as inputs, and benchmarked against manually defined diagram elements.
End-to-end speech synthesis models directly convert the input characters into an audio representation (e.g., spectrograms). Despite their impressive performance, such models have difficulty disambiguating the pronunciations of identically spelled words. To mitigate this issue, a separate Grapheme-to-Phoneme (G2P) model can be employed to convert the characters into phonemes before synthesizing the audio. This paper proposes SoundChoice, a novel G2P architecture that processes entire sentences rather than operating at the word level. The proposed architecture takes advantage of a weighted homograph loss (that improves disambiguation), exploits curriculum learning (that gradually switches from word-level to sentence-level G2P), and integrates word embeddings from BERT (for further performance improvement). Moreover, the model inherits the best practices in speech recognition, including multi-task learning with Connectionist Temporal Classification (CTC) and beam search with an embedded language model. As a result, SoundChoice achieves a Phoneme Error Rate (PER) of 2.65% on whole-sentence transcription using data from LibriSpeech and Wikipedia. Index Terms grapheme-to-phoneme, speech synthesis, text-tospeech, phonetics, pronunciation, disambiguation.
Image memes and specifically their widely-known variation image macros, is a special new media type that combines text with images and is used in social media to playfully or subtly express humour, irony, sarcasm and even hate. It is important to accurately retrieve image memes from social media to better capture the cultural and social aspects of online phenomena and detect potential issues (hate-speech, disinformation). Essentially, the background image of an image macro is a regular image easily recognized as such by humans but cumbersome for the machine to do so due to feature map similarity with the complete image macro. Hence, accumulating suitable feature maps in such cases can lead to deep understanding of the notion of image memes. To this end, we propose a methodology that utilizes the visual part of image memes as instances of the regular image class and the initial image memes as instances of the image meme class to force the model to concentrate on the critical parts that characterize an image meme. Additionally, we employ a trainable attention mechanism on top of a standard ViT architecture to enhance the model's ability to focus on these critical parts and make the predictions interpretable. Several training and test scenarios involving web-scraped regular images of controlled text presence are considered in terms of model robustness and accuracy. The findings indicate that light visual part utilization combined with sufficient text presence during training provides the best and most robust model, surpassing state of the art.
Named entity recognition (NER) task aims at identifying entities from a piece of text that belong to predefined semantic types such as person, location, organization, etc. The state-of-the-art solutions for flat entities NER commonly suffer from capturing the fine-grained semantic information in underlying texts. The existing span-based approaches overcome this limitation, but the computation time is still a concern. In this work, we propose a novel span-based NER framework, namely Global Pointer (GP), that leverages the relative positions through a multiplicative attention mechanism. The ultimate goal is to enable a global view that considers the beginning and the end positions to predict the entity. To this end, we design two modules to identify the head and the tail of a given entity to enable the inconsistency between the training and inference processes. Moreover, we introduce a novel classification loss function to address the imbalance label problem. In terms of parameters, we introduce a simple but effective approximate method to reduce the training parameters. We extensively evaluate GP on various benchmark datasets. Our extensive experiments demonstrate that GP can outperform the existing solution. Moreover, the experimental results show the efficacy of the introduced loss function compared to softmax and entropy alternatives.
Hierarchical Text Classification (HTC), which aims to predict text labels organized in hierarchical space, is a significant task lacking in investigation in natural language processing. Existing methods usually encode the entire hierarchical structure and fail to construct a robust label-dependent model, making it hard to make accurate predictions on sparse lower-level labels and achieving low Macro-F1. In this paper, we propose a novel PAMM-HiA-T5 model for HTC: a hierarchy-aware T5 model with path-adaptive mask mechanism that not only builds the knowledge of upper-level labels into low-level ones but also introduces path dependency information in label prediction. Specifically, we generate a multi-level sequential label structure to exploit hierarchical dependency across different levels with Breadth-First Search (BFS) and T5 model. To further improve label dependency prediction within each path, we then propose an original path-adaptive mask mechanism (PAMM) to identify the label's path information, eliminating sources of noises from other paths. Comprehensive experiments on three benchmark datasets show that our novel PAMM-HiA-T5 model greatly outperforms all state-of-the-art HTC approaches especially in Macro-F1. The ablation studies show that the improvements mainly come from our innovative approach instead of T5.
We propose a novel text editing task, referred to as \textit{fact-based text editing}, in which the goal is to revise a given document to better describe the facts in a knowledge base (e.g., several triples). The task is important in practice because reflecting the truth is a common requirement in text editing. First, we propose a method for automatically generating a dataset for research on fact-based text editing, where each instance consists of a draft text, a revised text, and several facts represented in triples. We apply the method into two public table-to-text datasets, obtaining two new datasets consisting of 233k and 37k instances, respectively. Next, we propose a new neural network architecture for fact-based text editing, called \textsc{FactEditor}, which edits a draft text by referring to given facts using a buffer, a stream, and a memory. A straightforward approach to address the problem would be to employ an encoder-decoder model. Our experimental results on the two datasets show that \textsc{FactEditor} outperforms the encoder-decoder approach in terms of fidelity and fluency. The results also show that \textsc{FactEditor} conducts inference faster than the encoder-decoder approach.
We summarize the evaluation of the first Automated Legal Question Answering Competition (ALQAC 2021). The competition this year contains three tasks, which aims at processing the statute law document, which are Legal Text Information Retrieval (Task 1), Legal Text Entailment Prediction (Task 2), and Legal Text Question Answering (Task 3). The final goal of these tasks is to build a system that can automatically determine whether a particular statement is lawful. There is no limit to the approaches of the participating teams. This year, there are 5 teams participating in Task 1, 6 teams participating in Task 2, and 5 teams participating in Task 3. There are in total 36 runs submitted to the organizer. In this paper, we summarize each team's approaches, official results, and some discussion about the competition. Only results of the teams who successfully submit their approach description paper are reported in this paper.