Are Prompt-based Models Clueless?

Finetuning large pre-trained language models with a task-specific head has advanced the state-of-the-art on many natural language understanding benchmarks. However, models with a task-specific head require a lot of training data, making them susceptible to learning and exploiting dataset-specific superficial cues that do not generalize to other datasets. Prompting has reduced the data requirement by reusing the language model head and formatting the task input to match the pre-training objective. Therefore, it is expected that few-shot prompt-based models do not exploit superficial cues. This paper presents an empirical examination of whether few-shot prompt-based models also exploit superficial cues. Analyzing few-shot prompt-based models on MNLI, SNLI, HANS, and COPA has revealed that prompt-based models also exploit superficial cues. While the models perform well on instances with superficial cues, they often underperform or only marginally outperform random accuracy on instances without superficial cues.

Two Training Strategies for Improving Relation Extraction over Universal Graph

This paper explores how the Distantly Supervised Relation Extraction (DS-RE) can benefit from the use of a Universal Graph (UG), the combination of a Knowledge Graph (KG) and a large-scale text collection. A straightforward extension of a current state-of-the-art neural model for DS-RE with a UG may lead to degradation in performance. We first report that this degradation is associated with the difficulty in learning a UG and then propose two training strategies: (1) Path Type Adaptive Pretraining, which sequentially trains the model with different types of UG paths so as to prevent the reliance on a single type of UG path; and (2) Complexity Ranking Guided Attention mechanism, which restricts the attention span according to the complexity of a UG path so as to force the model to extract features not only from simple UG paths but also from complex ones. Experimental results on both biomedical and NYT10 datasets prove the robustness of our methods and achieve a new state-of-the-art result on the NYT10 dataset. The code and datasets used in this paper are available at https://github.com/baodaiqin/UGDSRE.

NeurIPS 2020 EfficientQA Competition: Systems, Analyses and Lessons Learned

We review the EfficientQA competition from NeurIPS 2020. The competition focused on open-domain question answering (QA), where systems take natural language questions as input and return natural language answers. The aim of the competition was to build systems that can predict correct answers while also satisfying strict on-disk memory budgets. These memory budgets were designed to encourage contestants to explore the trade-off between storing large, redundant, retrieval corpora or the parameters of large learned models. In this report, we describe the motivation and organization of the competition, review the best submissions, and analyze system predictions to inform a discussion of evaluation for open-domain QA.

An Empirical Study of Contextual Data Augmentation for Japanese Zero Anaphora Resolution

One critical issue of zero anaphora resolution (ZAR) is the scarcity of labeled data. This study explores how effectively this problem can be alleviated by data augmentation. We adopt a state-of-the-art data augmentation method, called the contextual data augmentation (CDA), that generates labeled training instances using a pretrained language model. The CDA has been reported to work well for several other natural language processing tasks, including text classification and machine translation. This study addresses two underexplored issues on CDA, that is, how to reduce the computational cost of data augmentation and how to ensure the quality of the generated data. We also propose two methods to adapt CDA to ZAR: [MASK]-based augmentation and linguistically-controlled masking. Consequently, the experimental results on Japanese ZAR show that our methods contribute to both the accuracy gain and the computation cost reduction. Our closer analysis reveals that the proposed method can improve the quality of the augmented training data when compared to the conventional CDA.

Modeling Event Salience in Narratives via Barthes' Cardinal Functions

Events in a narrative differ in salience: some are more important to the story than others. Estimating event salience is useful for tasks such as story generation, and as a tool for text analysis in narratology and folkloristics. To compute event salience without any annotations, we adopt Barthes' definition of event salience and propose several unsupervised methods that require only a pre-trained language model. Evaluating the proposed methods on folktales with event salience annotation, we show that the proposed methods outperform baseline methods and find fine-tuning a language model on narrative texts is a key factor in improving the proposed methods.

Word Rotator's Distance: Decomposing Vectors Gives Better Representations

One key principle for assessing semantic similarity between texts is to measure the degree of semantic overlap of them by considering word-by-word alignment. However, alignment-based approaches} are inferior to the generic sentence vectors in terms of performance. We hypothesize that the reason for the inferiority of alignment-based methods is due to the fact that they do not distinguish word importance and word meaning. To solve this, we propose to separate word importance and word meaning by decomposing word vectors into their norm and direction, then compute the alignment-based similarity with the help of earth mover's distance. We call the method word rotator's distance (WRD) because direction vectors are aligned by rotation on the unit hypersphere. In addition, to incorporate the advance of cutting edge additive sentence encoders, we propose to re-decompose such sentence vectors into word vectors and use them as inputs to WRD. Empirically, the proposed method outperforms current methods considering the word-by-word alignment including word mover's distance with a big difference; moreover, our method outperforms state-of-the-art additive sentence encoders on the most competitive dataset, STS-benchmark.

Data Augmentation using Random Image Cropping and Patching for Deep CNNs

Deep convolutional neural networks (CNNs) have achieved remarkable results in image processing tasks. However, their high expression ability risks overfitting. Consequently, data augmentation techniques have been proposed to prevent overfitting while enriching datasets. Recent CNN architectures with more parameters are rendering traditional data augmentation techniques insufficient. In this study, we propose a new data augmentation technique called random image cropping and patching (RICAP) which randomly crops four images and patches them to create a new training image. Moreover, RICAP mixes the class labels of the four images, resulting in an advantage similar to label smoothing. We evaluated RICAP with current state-of-the-art CNNs (e.g., the shake-shake regularization model) by comparison with competitive data augmentation techniques such as cutout and mixup. RICAP achieves a new state-of-the-art test error of $2.19\%$ on CIFAR-10. We also confirmed that deep CNNs with RICAP achieve better results on classification tasks using CIFAR-100 and ImageNet and an image-caption retrieval task using Microsoft COCO.

Interpretable and Compositional Relation Learning by Joint Training with an Autoencoder

Embedding models for entities and relations are extremely useful for recovering missing facts in a knowledge base. Intuitively, a relation can be modeled by a matrix mapping entity vectors. However, relations reside on low dimension sub-manifolds in the parameter space of arbitrary matrices---for one reason, composition of two relations $\boldsymbol{M}_1,\boldsymbol{M}_2$ may match a third $\boldsymbol{M}_3$ (e.g. composition of relations currency_of_country and country_of_film usually matches currency_of_film_budget), which imposes compositional constraints to be satisfied by the parameters (i.e. $\boldsymbol{M}_1\cdot \boldsymbol{M}_2\approx \boldsymbol{M}_3$). In this paper we investigate a dimension reduction technique by training relations jointly with an autoencoder, which is expected to better capture compositional constraints. We achieve state-of-the-art on Knowledge Base Completion tasks with strongly improved Mean Rank, and show that joint training with an autoencoder leads to interpretable sparse codings of relations, helps discovering compositional constraints and benefits from compositional training. Our source code is released at github.com/tianran/glimvec.

A Novel Weight-Shared Multi-Stage Network Architecture of CNNs for Scale Invariance

Convolutional neural networks (CNNs) have demonstrated remarkable results in image classification tasks for benchmark and practical uses. The CNNs with deeper architectures have achieved higher performances recently thanks to their robustness to parallel shift of objects in images aw well as their numerous parameters and resulting high expression ability. However, the CNNs have a limited robustness to other geometric transformations such as scaling and rotation. This problem is considered to limit performance improvement of the deep CNNs but there is no established solution. This study focuses on scale transformation and proposes a novel network architecture called weight-shared multi-stage network (WSMS-Net), consisting of multiple stages of CNNs. The WSMS-Net is easily combined with existing deep CNNs, such as ResNet and DenseNet, and enables them to acquire a robustness to scaling of objects. The experimental results demonstrate that existing deep CNNs combined with the proposed WSMS-Net achieve higher accuracy for image classification tasks only with a little increase in the number of parameters.