Zewen Chi, Li Dong, Furu Wei, Nan Yang, Saksham Singhal, Wenhui Wang, Xia Song, Xian-Ling Mao, Heyan Huang, Ming Zhou. Proceedings of the 2021 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies. 2021.

In this work we focus on transferring supervision signals of natural language generation (NLG) tasks between multiple languages. We propose to pretrain the encoder and decoder a sequence-to-sequence model under both monolingual cross-lingual settings. The pre-training objective encourages represent different languages in shared space, so that can conduct zero-shot transfer. After procedure, use data fine-tune pre-trained downstream NLG tasks. Then trained single be directly evaluated beyond...

Knowledge graph (KG) plays an increasingly important role in recommender systems. Recently, neural networks (GNNs) based model has gradually become the theme of knowledge-aware recommendation (KGR). However, there is a natural deficiency for GNN-based KGR models, that is, sparse supervised signal problem, which may make their actual performance drop to some extent. Inspired by recent success contrastive learning mining signals from data itself, this paper, we focus on exploring KG-aware and...

Joint entity and relation extraction is an essential task in natural language processing knowledge graph construction. Existing approaches usually decompose the joint into several basic modules or steps to make it easy conduct. However, such a paradigm ignores fact that three elements of triple are interdependent indivisible. Therefore, previous methods suffer from problems cascading errors redundant information. To address these issues, this paper, we propose novel model, named OneRel,...

Session-based recommendation (SBR) is a challenging task, which aims at recommending items based on anonymous behavior sequences. Almost all the existing solutions for SBR model user preference only current session without exploiting other sessions, may contain both relevant and irrelevant item-transitions to session. This paper proposes novel approach, called Global Context Enhanced Graph Neural Networks (GCE-GNN) exploit item transitions over sessions in more subtle manner better inferring...

Bundle recommendation aims to recommend the user a bundle of items as whole. Previous models capture user’s preferences on both and association items. Nevertheless, they usually neglect diversity intents adopting fail disentangle in representations. In real scenario recommendation, intent may be naturally distributed different bundles that (Global view). And contain multiple (Local Each view has its advantages for disentangling: 1) global view, more are involved present each intent, which...

Chinese Spell Checking (CSC) aims to detect and correct erroneous characters for usergenerated text in language.Most of the spelling errors are misused semantically, phonetically or graphically similar characters.Previous attempts notice this phenomenon try utilize similarity relationship task.However, these methods use either heuristics handcrafted confusion sets predict character.In paper, we propose a spell checker called REALISE, by directly leveraging multimodal information...

Zewen Chi, Shaohan Huang, Li Dong, Shuming Ma, Bo Zheng, Saksham Singhal, Payal Bajaj, Xia Song, Xian-Ling Mao, Heyan Furu Wei. Proceedings of the 60th Annual Meeting Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Incorporating Knowledge Graphs (KG) into recommeder system as side information has attracted considerable attention. Recently, the technical trend of Knowledge-aware Recommendation (KGR) is to develop end-to-end models based on graph neural networks (GNNs). However, extremely sparse user-item interactions significantly degrade performance GNN-based models, from following aspects: 1) interaction, itself, means inadequate supervision signals and limits supervised models; 2) combination (CF...

Aspect-based sentiment analysis (ABSA) is a fine-grained task that aims to align aspects and corresponding sentiments for aspect-specific polarity inference. It challenging because sentence may contain multiple or complicated (e.g., conditional, coordinating, adversative) relations. Recently, exploiting dependency syntax information with graph neural networks has been the most popular trend. Despite its success, methods heavily rely on tree pose challenges in accurately modeling alignment of...

Cross-modal hashing has been widely used in multimedia retrieval tasks due to its fast speed and low storage cost. Recently, many deep unsupervised cross-modal methods have proposed deal the unlabeled datasets. These usually construct an instance similarity matrix by fusing image text modality-specific matrices as guiding information train networks. However, most of them directly use cosine similarities between bag-of-words (BoW) vectors datapoints define matrix, which fails mine semantic...

Aspect Sentiment Triplet Extraction (ASTE) has become an emerging task in sentiment analysis research, aiming to extract triplets of the aspect term, its corresponding opinion and associated polarity from a given sentence. Recently, many neural networks based models with different tagging schemes have been proposed, but almost all them their limitations: heavily relying on 1) prior assumption that each word is only single role (e.g., or etc. ) 2) word-level interactions treating...

In this work, we present an information-theoretic framework that formulates cross-lingual language model pre-training as maximizing mutual information between multilingual-multi-granularity texts. The unified view helps us to better understand the existing methods for learning representations. More importantly, inspired by framework, propose a new task based on contrastive learning. Specifically, regard bilingual sentence pair two views of same meaning and encourage their encoded...

Most of the unsupervised hashing methods usually map images into semantic similarity-preserving hash codes by constructing local similarity structure as guiding information, i.e., treating each point similar to its k nearest neighbours. However, for an image, some neighbours may be dissimilar it, they are noisy datapoints which will damage retrieval performance. Thus, tackle this problem, in paper, we propose a novel deep method, called MLS3RDUH, can reduce further enhance Specifically,...

The task of table structure recognition aims to recognize the internal a table, which is key step make machines understand tables. Currently, there are lots studies on this for different file formats such as ASCII text and HTML. It also attracts attention structures in PDF files. However, it hard existing methods accurately complicated tables contain spanning cells occupy at least two columns or rows. To address issue, we propose novel graph neural network recognizing files, named GraphTSR....

Due to their high retrieval efficiency and low storage cost, cross-modal hashing methods have attracted considerable attention. Generally, compared with shallow methods, deep can achieve a more satisfactory performance by integrating feature learning hash codes optimizing into same framework. However, most existing either cannot learn unified code for the two correlated data-points of different modalities in database instance or guide feedback function procedure, enhance accuracy. To address...

Zewen Chi, Li Dong, Shuming Ma, Shaohan Huang, Saksham Singhal, Xian-Ling Mao, Heyan Xia Song, Furu Wei. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021.

Zewen Chi, Li Dong, Bo Zheng, Shaohan Huang, Xian-Ling Mao, Heyan Furu Wei. Proceedings of the 59th Annual Meeting Association for Computational Linguistics and 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Sequential recommendation (SR) aims to predict the subsequent behaviors of users by understanding their successive historical behaviors. Recently, some methods for SR are devoted alleviating data sparsity problem (i.e., limited supervised signals training), which take account contrastive learning incorporate self-supervised into SR. Despite achievements, it is far from enough learn informative user/item embeddings due inadequacy modeling complex collaborative information and co-action...

Recently, deep supervised hashing methods have shown state-of-the-art performance by integrating feature learning and hash codes into an end-to-end network to generate high-quality codes. However, it is still a challenge learn discriminative for preserving the label information of images efficiently. To overcome this difficulty, in paper, we propose novel Partial-Softmax Loss based Deep Hashing, called PSLDH, Specifically, PSLDH first trains category code each category, will preserve...