[DEMO] GluonNLP demos

[eric-haibin-lin]

Similar to https://demo.allennlp.org, it would be great to have online demo applications for various models available in gluonnlp.

List of demos from AllenNLP for reference:

Annotate a sentence

• Semantic Role Labeling
• Named Entity Recognition
• Constituency Parsing
• Dependency Parsing
• Open Information Extraction
• Sentiment Analysis

Annotate a passage

• Coreference Resolution

Answer a question

• Reading Comprehension

Semantic parsing

• WikiTableQuestions Semantic Parser
• Cornell NLVR Semantic Parser
• Text to SQL (ATIS)
• QuaRel Zero

Other

• Textual Entailment
• Event2Mind
• Language Modeling
• Masked Language Modeling

[eric-haibin-lin]

1. We can add dependency parsing, coreference resolution, named entity recognition, and POS tagging similar to https://emorynlp.github.io/ddr/viz/

Specifically, we want users to enter a paragraph (or selection one from a set of candidate paragraphs), and we show the result of coreference, NER, POS tagging inline with the paragraph. When the user clicks on any sentence in the paragraph, we show an additional dependency parsing visualization result for that sentence.

2. Abstractive Summarization

   • UniLM: https://github.com/microsoft/unilm , or
   • Get To The Point: Summarization with Pointer-Generator Networks: https://github.com/abisee/pointer-generator

3. Natural Language for Visual Reasoning

   • LXMERT: Learning Cross-Modality Encoder Representations from Transformers: https://github.com/airsplay/lxmert
   • VisualBERT: https://github.com/uclanlp/visualbert

[eric-haibin-lin]

2. Abstractive Summarization
   • UniLM: https://github.com/microsoft/unilm , or
   • Get To The Point: Summarization with Pointer-Generator Networks: https://github.com/abisee/pointer-generator

[eric-haibin-lin]

3. Natural Language for Visual Reasoning
   • LXMERT: Learning Cross-Modality Encoder Representations from Transformers: https://github.com/airsplay/lxmert
   • VisualBERT: https://github.com/uclanlp/visualbert

[eric-haibin-lin]

@StrayBird-ATSH we can probably pick 1 or 2

[eric-haibin-lin]

We can have the following stack:

• dedicated g4dn.xlarge instance for server hosting
• docker and conda for environment & dependency management
• MMS for serving logics (load pre-trained model, inference per request)
• colab & SageMaker script for users' customization

[leezu]

dedicated g4dn.xlarge instance for server hosting

how to handle traffic spikes? It may require auto-scaling and batching of user requests?

[eric-haibin-lin]

Maybe we can add auto scaling for the second phase of the project, considering team's capacity. For now, I assume the traffic (number_unique_visitors) is not a lot. Because the Web UI will not allow large amount of automated inference jobs, we will have at most number_unique_visitors concurrent connections. If users want to leverage our demo for running a large amount of data, they can use the sagemaker/colab script we prepared.