Linyuan Gong, Chenyan Xiong, Xiaodong Liu, Payal Bajaj, Yiqing Xie, Alvin Cheung, Jianfeng Gao, Xia Song. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

A big convergence of model architectures across language, vision, speech, and multimodal is emerging. However, under the same name ”Transformers”, the above areas use different implementations for better performance, e.g., Post-LayerNorm for BERT, and Pre-LayerNorm for GPT and vision Transformers. We call for the development of Foundation Transformer for true general-purpose modeling, which serves as a go-to architecture for various tasks and modalities with guaranteed training stability. In this work, we introduce a Transformer variant, named Magneto, to fulfill the goal. Specifically, we propose Sub-LayerNorm for good expressivity, and the initialization strategy theoretically derived from DeepNet for stable scaling up. Extensive experiments demonstrate its superior performance and better stability than the de facto Transformer variants designed for various applications, including language modeling (i.e., BERT, and GPT), machine translation, vision pretraining (i.e., BEiT), speech recognition, and multimodal pretraining (i.e., BEiT-3).

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

Sparse mixture of experts provides larger model capacity while requiring a constant computational overhead. It employs the routing mechanism to distribute input tokens to the best-matched experts according to their hidden representations. However, learning such a routing mechanism encourages token clustering around expert centroids, implying a trend toward representation collapse. In this work, we propose to estimate the routing scores between tokens and experts on a low-dimensional hypersphere. We conduct extensive experiments on cross-lingual language model pre-training and fine-tuning on downstream tasks. Experimental results across seven multilingual benchmarks show that our method achieves consistent gains. We also present a comprehensive analysis on the representation and routing behaviors of our models. Our method alleviates the representation collapse issue and achieves more consistent routing than the baseline mixture-of-experts methods.

We present a self-supervised learning framework, COCO-LM, that pretrains Language Models by COrrecting and COntrasting corrupted text sequences. Following ELECTRA-style pretraining, COCO-LM employs an auxiliary language model to corrupt text sequences, upon which it constructs two new tasks for pretraining the main model. The first token-level task, Corrective Language Modeling, is to detect and correct tokens replaced by the auxiliary model, in order to better capture token-level semantics. The second sequence-level task, Sequence Contrastive Learning, is to align text sequences originated from the same source input while ensuring uniformity in the representation space. Experiments on GLUE and SQuAD demonstrate that COCO-LM not only outperforms recent state-of-the-art pretrained models in accuracy, but also improves pretraining efficiency. It achieves the MNLI accuracy of ELECTRA with 50% of its pretraining GPU hours. With the same pretraining steps of standard base/large-sized models, COCO-LM outperforms the previous best models by 1+ GLUE average points.

Learning low-dimensional embeddings of knowledge graphs is a powerful approach used to predict unobserved or missing edges between entities. However, an open challenge in this area is developing techniques that can go beyond simple edge prediction and handle more complex logical queries, which might involve multiple unobserved edges, entities, and variables. For instance, given an incomplete biological knowledge graph, we might want to predict "em what drugs are likely to target proteins involved with both diseases X and Y?" -- a query that requires reasoning about all possible proteins that might interact with diseases X and Y. Here we introduce a framework to efficiently make predictions about conjunctive logical queries -- a flexible but tractable subset of first-order logic -- on incomplete knowledge graphs. In our approach, we embed graph nodes in a low-dimensional space and represent logical operators as learned geometric operations (e.g., translation, rotation) in this embedding space. By performing logical operations within a low-dimensional embedding space, our approach achieves a time complexity that is linear in the number of query variables, compared to the exponential complexity required by a naive enumeration-based approach. We demonstrate the utility of this framework in two application studies on real-world datasets with millions of relations: predicting logical relationships in a network of drug-gene-disease interactions and in a graph-based representation of social interactions derived from a popular web forum.

Geo-fencing is a location based service that allows sending of messages to users who enter/exit a specified geographical area, known as a geo-fence. Today, it has become one of the popular location based mobile marketing strategies. However, the process of designing geo-fences is presently manual, i.e. a retailer must specify the location and the radius of area around it to setup the geo-fences. Moreover, this process does not consider the user's preference towards the targeted product/service and thus, can compromise his/her experience of the app that sends these communications. We attempt to solve this problem by presenting a novel end-to-end system for automated design of affinity based smart geo-fences. Affinity towards a product/service refers to the user's interest in a product/service. Our unique formulation to estimate affinity, using historical app usage data, is sensitive to a user's location and thus, the affinity is termed as location sensitive product affinity (LSPA). The geo-fence logic tries to capture contiguous groups of locations where the affinity high. Experiments on real world e-commerce dataset reveals that geo-fences designed by our approach performs significantly better at accurately targeting the users who are interested in a product. We thus show that, using historical app usage data, geo-fences can be designed in an automated manner and can help enterprises target interested users with better accuracy as compared to the present industry practices.

Obtaining enough labeled data to robustly train complex discriminative models is a major bottleneck in the machine learning pipeline. A popular solution is combining multiple sources of weak supervision using generative models. The structure of these models affects the quality of the training labels, but is difficult to learn without any ground truth labels. We instead rely on weak supervision sources having some structure by virtue of being encoded programmatically. We present Coral, a paradigm that infers generative model structure by statically analyzing the code for these heuristics, thus significantly reducing the amount of data required to learn structure. We prove that Coral's sample complexity scales quasilinearly with the number of heuristics and number of relations identified, improving over the standard sample complexity, which is exponential in n for learning n-th degree relations. Empirically, Coral matches or outperforms traditional structure learning approaches by up to 3.81 F1 points. Using Coral to model dependencies instead of assuming independence results in better performance than a fully supervised model by 3.07 accuracy points when heuristics are used to label radiology data without ground truth labels.

Online TV has seen rapid growth in recent years, with most of the large media companies broadcasting their linear content online. Access to the online TV accounts is protected by an authentication, and like the traditional cable TV subscription, users in the same household share the online TV credentials. However, as the standard data collection techniques have capability to collect only account level information, online TV measurements fail to capture individual level viewing characteristics in shared accounts. Thus, individual profile identification and experience individualization are challenging and difficult for online TV platforms. In this paper, we propose a novel approach to decompose online TV account into distinct personas sharing the account through analyzing viewing characteristics. A recommendation algorithm is then proposed to individualize the experience for each persona. Finally, we demonstrate the usefulness of the proposed approach through experiments on a large online TV database.

YouTube draws large number of users who contribute actively by uploading videos or commenting on existing videos. However, being a crowd sourced and large content pushed onto it, there is limited control over the content. This makes malicious users push content (videos and comments) which is inappropriate (unsafe), particularly when such content is placed around cartoon videos which are typically watched by kids. In this paper, we focus on presence of unsafe content for children and users who promote it. For detection of child unsafe content and its promoters, we perform two approaches, one based on supervised classification which uses an extensive set of video-level, user-level and comment-level features and another based Convolutional Neural Network using video frames. Detection accuracy of 85.7% is achieved which can be leveraged to build a system to provide a safe YouTube experience for kids. Through detailed characterization studies, we are able to successfully conclude that unsafe content promoters are less popular and engage less as compared with other users. Finally, using a network of unsafe content promoters and other users based on their engagements (likes, subscription and playlist addition) and other factors, we find that unsafe content is present very close to safe content and unsafe content promoters form very close knit communities with other users, thereby further increasing the likelihood of a child getting getting exposed to unsafe content.