Data augmentation is an effective approach to tackle over-fitting. Many previous works have proposed different data augmentations strategies for NLP, such as noise injection, word replacement, back-translation etc. Though effective, they missed one important characteristic of language--compositionality, meaning of a complex expression is built from its sub-parts. Motivated by this, we propose a compositional data augmentation approach for natural language understanding called TreeMix. Specifically, TreeMix leverages constituency parsing tree to decompose sentences into constituent sub-structures and the Mixup data augmentation technique to recombine them to generate new sentences. Compared with previous approaches, TreeMix introduces greater diversity to the samples generated and encourages models to learn compositionality of NLP data. Extensive experiments on text classification and SCAN demonstrate that TreeMix outperforms current state-of-the-art data augmentation methods.
Transformers achieve state-of-the-art performance for natural language processing tasks by pre-training on large-scale text corpora. They are extremely compute-intensive and have very high sample complexity. Memory replay is a mechanism that remembers and reuses past examples by saving to and replaying from a memory buffer. It has been successfully used in reinforcement learning and GANs due to better sample efficiency. In this paper, we propose \emph{Transformer with Memory Replay} (TMR), which integrates memory replay with transformer, making transformer more sample-efficient. Experiments on GLUE and SQuAD benchmark datasets show that Transformer with Memory Replay achieves at least $1\%$ point increase compared to the baseline transformer model when pretrained with the same number of examples. Further, by adopting a careful design that reduces the wall-clock time overhead of memory replay, we also empirically achieve a better runtime efficiency.
Previous work on text generation from graph-structured data relies on pretrained language models (PLMs) and utilizes graph linearization heuristics rather than explicitly considering the graph structure. Efficiently encoding the graph structure in PLMs is challenging because they were pretrained on natural language, and modeling structured data may lead to catastrophic forgetting of distributional knowledge. In this paper, we propose StructAdapt, an adapter method to encode graph structure into PLMs. Contrary to prior work, StructAdapt effectively models interactions among the nodes based on the graph connectivity, only training graph structure-aware adapter parameters. In this way, we avoid catastrophic forgetting while maintaining the topological structure of the graph. We empirically show the benefits of explicitly encoding graph structure into PLMs using adapters and achieve state-of-the-art results on two AMR-to-text datasets, training only 5.1% of the PLM parameters.
Scene text magnifier aims to magnify text in natural scene images without recognition. It could help the special groups, who have myopia or dyslexia to better understand the scene. In this paper, we design the scene text magnifier through interacted four CNN-based networks: character erasing, character extraction, character magnify, and image synthesis. The architecture of the networks are extended based on the hourglass encoder-decoders. It inputs the original scene text image and outputs the text magnified image while keeps the background unchange. Intermediately, we can get the side-output results of text erasing and text extraction. The four sub-networks are first trained independently and fine-tuned in end-to-end mode. The training samples for each stage are processed through a flow with original image and text annotation in ICDAR2013 and Flickr dataset as input, and corresponding text erased image, magnified text annotation, and text magnified scene image as output. To evaluate the performance of text magnifier, the Structural Similarity is used to measure the regional changes in each character region. The experimental results demonstrate our method can magnify scene text effectively without effecting the background.
User evaluations include a significant quantity of information across online platforms. This information source has been neglected by the majority of existing recommendation systems, despite its potential to ease the sparsity issue and enhance the quality of suggestions. This work presents a deep model for concurrently learning item attributes and user behaviour from review text. Deep Cooperative Neural Network (DeepCoNN) is the suggested model consisting of two parallel neural networks connected in their final layers. One of the networks focuses on learning user behaviour from reviews submitted by the user, while the other network learns item attributes from user reviews. On top, a shared layer is added to connect these two networks. Similar to factorization machine approaches, the shared layer allows latent factors acquired for people and things to interact with each other. On a number of datasets, DeepCoNN surpasses all baseline recommendation systems, according to experimental findings.
We study a new application for text generation -- idiomatic sentence generation -- which aims to transfer literal phrases in sentences into their idiomatic counterparts. Inspired by psycholinguistic theories of idiom use in one's native language, we propose a novel approach for this task, which retrieves the appropriate idiom for a given literal sentence, extracts the span of the sentence to be replaced by the idiom, and generates the idiomatic sentence by using a neural model to combine the retrieved idiom and the remainder of the sentence. Experiments on a novel dataset created for this task show that our model is able to effectively transfer literal sentences into idiomatic ones. Furthermore, automatic and human evaluations show that for this task, the proposed model outperforms a series of competitive baseline models for text generation.
Aligning signals from different modalities is an important step in vision-language representation learning as it affects the performance of later stages such as cross-modality fusion. Since image and text typically reside in different regions of the feature space, directly aligning them at instance level is challenging especially when features are still evolving during training. In this paper, we propose to align at a higher and more stable level using cluster representation. Specifically, we treat image and text as two "views" of the same entity, and encode them into a joint vision-language coding space spanned by a dictionary of cluster centers (codebook). We contrast positive and negative samples via their cluster assignments while simultaneously optimizing the cluster centers. To further smooth out the learning process, we adopt a teacher-student distillation paradigm, where the momentum teacher of one view guides the student learning of the other. We evaluated our approach on common vision language benchmarks and obtain new SoTA on zero-shot cross modality retrieval while being competitive on various other transfer tasks.
Despite the widespread application of recurrent neural networks (RNNs) across a variety of tasks, a unified understanding of how RNNs solve these tasks remains elusive. In particular, it is unclear what dynamical patterns arise in trained RNNs, and how those patterns depend on the training dataset or task. This work addresses these questions in the context of a specific natural language processing task: text classification. Using tools from dynamical systems analysis, we study recurrent networks trained on a battery of both natural and synthetic text classification tasks. We find the dynamics of these trained RNNs to be both interpretable and low-dimensional. Specifically, across architectures and datasets, RNNs accumulate evidence for each class as they process the text, using a low-dimensional attractor manifold as the underlying mechanism. Moreover, the dimensionality and geometry of the attractor manifold are determined by the structure of the training dataset; in particular, we describe how simple word-count statistics computed on the training dataset can be used to predict these properties. Our observations span multiple architectures and datasets, reflecting a common mechanism RNNs employ to perform text classification. To the degree that integration of evidence towards a decision is a common computational primitive, this work lays the foundation for using dynamical systems techniques to study the inner workings of RNNs.
Semantic information of a sentence is crucial for improving the expressiveness of a text-to-speech (TTS) system, but can not be well learned from the limited training TTS dataset just by virtue of the nowadays encoder structures. As large scale pre-trained text representation develops, bidirectional encoder representations from transformers (BERT) has been proven to embody text-context semantic information and applied to TTS as additional input. However BERT can not explicitly associate semantic tokens from point of dependency relations in a sentence. In this paper, to enhance expressiveness, we propose a semantic representation learning method based on graph neural network, considering dependency relations of a sentence. Dependency graph of input text is composed of edges from dependency tree structure considering both the forward and the reverse directions. Semantic representations are then extracted at word level by the relational gated graph network (RGGN) fed with features from BERT as nodes input. Upsampled semantic representations and character-level embeddings are concatenated to serve as the encoder input of Tacotron-2. Experimental results show that our proposed method outperforms the baseline using vanilla BERT features both in LJSpeech and Bilzzard Challenge 2013 datasets, and semantic representations learned from the reverse direction are more effective for enhancing expressiveness.
Open information extraction (OIE) methods extract plenty of OIE triples <noun phrase, relation phrase, noun phrase> from unstructured text, which compose large open knowledge bases (OKBs). Noun phrases and relation phrases in such OKBs are not canonicalized, which leads to scattered and redundant facts. It is found that two views of knowledge (i.e., a fact view based on the fact triple and a context view based on the fact triple's source context) provide complementary information that is vital to the task of OKB canonicalization, which clusters synonymous noun phrases and relation phrases into the same group and assigns them unique identifiers. However, these two views of knowledge have so far been leveraged in isolation by existing works. In this paper, we propose CMVC, a novel unsupervised framework that leverages these two views of knowledge jointly for canonicalizing OKBs without the need of manually annotated labels. To achieve this goal, we propose a multi-view CH K-Means clustering algorithm to mutually reinforce the clustering of view-specific embeddings learned from each view by considering their different clustering qualities. In order to further enhance the canonicalization performance, we propose a training data optimization strategy in terms of data quantity and data quality respectively in each particular view to refine the learned view-specific embeddings in an iterative manner. Additionally, we propose a Log-Jump algorithm to predict the optimal number of clusters in a data-driven way without requiring any labels. We demonstrate the superiority of our framework through extensive experiments on multiple real-world OKB data sets against state-of-the-art methods.