United States

This work demonstrates the development process of a machine learning architecture for inference that can scale to a large volume of requests. In our experiments, we used a BERT model that was fine-tuned for emotion analysis, returning a probability distribution of emotions given a paragraph. The model was deployed as a gRPC service on Kubernetes. Apache Spark was used to perform inference in batches by calling the service. We encountered some performance and concurrency challenges and created solutions to achieve faster running time. Starting with 3.3 successful inference requests per second, we were able to achieve as high as 300 successful requests per second with the same batch job resource allocation. As a result, we successfully stored emotion probabilities for 95 million paragraphs within 96 hours.


NAACL 2021

An Architecture for Accelerated Large-Scale Inference of Transformer-Based Language Models

2021 Annual Conference of the North American Chapter of the Association for Computational Linguistics.

**Whova App** 
Stay in touch with your fellow conference attendees via the [Whova App](https://whova.com/portal/webapp/nacon_202106/)

**Conference Structure**
https://2021.naacl.org/blog/conference-structure/

**Walkthrough video of how to NAACL 2021** 

Please take a moment to view this video explaining how to navigate the platform, attend sessions network with other attendees. 


<figure class="video_container">
  <iframe src="https://screencast-o-matic.com/watch/crhwbGVh3vx?v=6&ff=1&title=0&controls=1" width=640  height=350  frameborder="0" allowfullscreen="true"> </iframe>
</figure>

<br>
<br>

technical paper

This presentation will talk about how LightSeq helps ByteDance's products and why it is so fast.

LightSeq: A High Performance Inference Library for Transformers

Can unsupervised machine translation be improved by other languages? This paper proposed to improve unsupervised neural machine translation with cross-lingual supervision, which utilizes supervision signals from high resource language pairs to improve the translation of zero-source languages. Specifically, for training En-Ro system without parallel corpus, we can leverage the corpus from En-Fr and En-De to collectively train the translation from one language into many languages under one model.

Cross-lingual Supervision Improves Unsupervised Neural Machine Translation

Most of the recent natural language processing (NLP) studies are based on the pretrain-finetuning approach (PFA). However for small and medium-sized industries with insufficient hardware, there are many limitations in servicing latest PFA based NLP application software, due to slow speed and insufficient memory. Since these approaches generally require large amounts of data, it is much more difficult to service with PFA especially for low-resource languages. We propose a new tokenization method, ONE-Piece, to address this limitation. ONE-Piece combines morphologically-aware subword tokenization and vocabulary communicating method, which has not been carefully considered before. Our proposed method can also be utilized without modifying the model structure. We experiment by applying ONE-Piece to Korean, a morphologically-rich and low-resource language. We revealed that ONE-Piece with vanilla transformer model can achieve comparable performance to the current Korean-English machine translation state-of-the-art model.

Should we find another model?: Improving Neural Machine Translation Performance with ONE-Piece Tokenization Method without Model Modification

Automatic translation of dialogue texts is a much needed demand in many real life scenarios. However, current neural machine translation systems usually deliver unsatisfying translation results of dialogue texts. In this paper, we conduct a deep analysis of a dialogue corpus and summarize three major issues on dialogue translation. In response to these challenges, we propose a joint learning method to identify omission and typo in the process of translating and utilize context to translate dialogue utterances.

Autocorrect in the Process of Translation --- Multi-task Learning Improves Dialogue Machine Translation

Neural Machine Translation (NMT) has achieved significant breakthrough in performance but is known to suffer vulnerability to input perturbations. As real input noise is difficult to predict during training, robustness is a big issue for system deployment. In this paper, we improve the robustness of NMT models by reducing the effect of noisy words through a Context-Enhanced Reconstruction (CER) approach. CER trains the model to resist noise in two steps: (1) perturbation step that breaks the naturalness of input sequence with made-up words; (2) reconstruction step that defends the noise propagation by generating better and more robust contextual representation. Experimental results on Chinese-English (ZH-EN) and French-English (FR-EN) translation tasks demonstrate robustness improvement on both news and social media text. Further finetuning experiments on social media text show our approach can converge at a higher position and provide a better adaptation.

Addressing the Vulnerability of NMT in Input Perturbations

Transformer-based methods are appealing for multilingual text classification, but common research benchmarks like XNLI (Conneau et al., 2018) do not reflect the data availability and task variety of industry applications. We present an empirical comparison of transformer-based text classification models in a variety of practical monolingual and multilingual pretraining and fine-tuning settings. We evaluate these methods on two distinct tasks in five different languages. Departing from prior work, our results show that multilingual language models can outperform monolingual ones in some downstream tasks and target languages. We additionally show that practical modifications such as task- and domain-adaptive pretraining and data augmentation can improve classification performance without the need for additional labeled data.

Practical Transformer-based Multilingual Text Classification

E-commerce has grown substantially over the last several years, and chatbots for intelligent customer service are concurrently drawing attention. We presented AliMe Assist, a Chinese intelligent assistant designed for creating an innovative online shopping experience in E-commerce. Based on question answering (QA), AliMe Assist offers assistance service, customer service, and chatting service. According to the survey of user studies and the real online testing, emotional comfort of customers’ negative emotions, which make up more than 5% of whole number of customer visits on AliMe, is a key point for providing considerate service. In this paper, we propose a framework to obtain proper replies to customers’ emotional questions. The framework takes emotion classification model as a core, and the final reply selection is based on topic classification and text matching. Our experiments on real online systems show that the framework is very promising.

An Emotional Comfort Framework for Improving User Satisfaction in E-Commerce Customer Service Chatbots

Although deep neural networks have been widely employed and proven effective in sentiment analysis tasks, it remains challenging for model developers to assess their models for erroneous predictions that might exist prior to deployment. Once deployed, emergent errors can be hard to identify in prediction runtime and impossible to trace back to their sources. To address such gaps, in this paper we propose an error detection framework for sentiment analysis based on explainable features. We perform global-level feature validation with human-in-the-loop assessment, followed by an integration of global and locallevel feature contribution analysis. Experimental results show that, given limited human-inthe- loop intervention, our method is able to identify erroneous model predictions on unseen data with high precision.


When and Why a Model Fails? A Human-in-the-loop Error Detection Framework for Sentiment Analysis

In this presentation, we considered the problem of scaling automated suggested replies for Outlook email system to multiple languages. Faced with increased compute requirements and low resource languages for language expansion, we build a single universal model for improving the quality and reducing run-time costs of our production system. Restricted data movement across regional clusters prevents joint training across languages. Thus, we proposed a multi-task continual learning framework with auxiliary tasks and language adapters. The results show positive cross-lingual transfer across languages and significant gains in CTR and char-saved online metrics. 


Language Scaling for Universal Suggested Replies Model

Product search is an important problem for retailers to solve as it is a deciding factor to drive business revenues. However, unlike web search, accurate word embeddings for product queries in mid-sized shops are especially difficult to obtain, considering the challenges posed by query length, data sparsity and limited computational capacity. In this talk, we present Query2Prod2Vec, a model that grounds lexical representations for product search in product embeddings: in our model, meaning is a mapping between words and a latent space of products, where the mapping is learned directly from noisy behavioral data. Our experiments show that our model is sample- and cost-efficient, and more accurate than known embedding techniques from the recent NLP and IR literature. Finally, by proposing Query2Prod2Vec, we hope to raise awareness on the interesting challenges of doing NLP inference at “reasonable scale”, and to showcase how research can produce practical solutions that are beneficial for practitioners even outside a few retail giants.


An Architecture for Accelerated Large-Scale Inference of Transformer-Based Language Models

Downloads

Next from NAACL 2021

LightSeq: A High Performance Inference Library for Transformers

Similar lecture

Efficient Attentions for Long Document Summarization

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

An Architecture for Accelerated Large-Scale Inference of Transformer-Based Language Models

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from NAACL 2021

LightSeq: A High Performance Inference Library for Transformers

Similar lecture

Efficient Attentions for Long Document Summarization

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads