Applied Machine Learning for Text Analysis Community of Practice: 2021 in review

In an effort to improve comment analysis from the 2021 Census of Population, Statistics Canada's Data Science Division worked in collaboration with the Census Subject Matter Secretariat to create a proof-of-concept on the use of ML techniques to quickly and objectively classify census comments. In addition to classifying comments by subject matter area, the models also classified comments regarding technical issues and privacy concerns.

The IRCC has collected client feedback about their services since 2014. The IRCC Client Experience Branch planned to launch a ML project to analyze the feedback in more detail. The analysis would help the department gain insight into their services, better align their innovation projects and inform their decision-making process. The presenter discussed with the CoP members about the appropriate approaches to demonstrate the added value of ML to senior management.

The Data Analysis section within the CRA's Audit, Evaluation, and Risk Branch created a data museum in 2016. The data museum enabled:

1. the delivery of descriptive and diagnostic data analytics for multiple assurance and advisory engagements,
2. the use of ML for better pattern recognition, classification, and outlier detection, and
3. the development of data privacy and protection standard operating procedures.

This presentation provided an overview of the challenges and solutions related to data engineering and the domain that deals with the automation of data processing and analysis. The presenter discussed the taxonomy of data engineering tasks and the tools to address them, and also described the efforts to build the Data Engineering Toolkit and the Data Engineering Community of Practice. They also displayed the Shiny Apps for linking and deduplicating noisy data and the NLP analysis of the Open Canada database of Completed Access to Information Requests.

This presentation gave a technical overview of the methodology behind topic modelling, explained the basis of Latent Dirichlet Allocation, and introduced a temporal dimension into the topic modelling analysis. This was done in the context of event detection using data from the Canadian Coroner and Medical Examiner Database.

The Emergency Preparedness Branch sends text notifications on occurrences (i.e. traffic accidents, etc.) affecting Canada's air, marine and surface transportation infrastructure, on a 24/7 basis to email subscribers. The presenter's team engineered bilingual notifications to produce analytical datasets in order to mine intelligence, enable text analytics, and detect patterns for similar occurring events.

This proof-of-concept web application mined social media data in real-time and provided geo-spatial insights and text analytics using NLP. The aim of the application was to help analysts determine to what extent safety and security occurrences could be extracted from social media. In addition, a sentiment-scoring component was added to enable sentiment analysis based on topics or organizational handles. While never operationalized, the proof-of-concept permitted rapid analysis for emerging issues for a subpopulation of social media users including individuals, news providers and national police.

Quantum technologies are set to disrupt common tasks in ML, including text classification. The presenters provided a review of three approaches to quantum ML in the noisy intermediate scale quantum device age, with the goal of introducing these methods to existing researchers and data scientists in the field.

The presenter shared his brief journey through practical applications of NLP to problems in health sciences research. This included using rule-based NLP to extract drug-disease interactions from medical research abstracts, or using ML to analyze tweets about vaccinations to predict disease outbreaks. He also discussed his lessons learned and mistakes he made along the way.

PHAC (with Xtract AI in Vancouver, BC) was examining the automation of different stages of evidence synthesis to increase efficiencies. The presenter shared an overview of an initial version of a novel ML-based system that was powered by recent advances in NLP, such as BioBERT, with further optimizations completed using a new immunization-specific document database. The resulting optimized NLP model at the core of this system was able to identify and extract Population, Intervention, Control, Outcomes (PICO) and PICO-related fields from publications on immunization with 88% average accuracy, across five classes of text.

Building dashboards have proven to be useful in the field of data science. With the current advancements in this field, there are emerging open source tools that are powerful, highly customizable and free to use. The presenter shared the python tools suitable for building dashboards, showed examples of relevant work the Data Science Division has done and presented a brief overview of how to get started with the two most popular tools – Dash and Django.

Portable Document Format (PDF) documents are commonly used by companies for financial reporting. The absence of an effective way to extract data from unstructured PDF files into a tabular format caused significant challenges for financial analysts to efficiently analyze and process information in a timely manner. Spatial Layout based Information and Content Extraction (SLICE) is a unique computer vision algorithm that simultaneously uses textual, visual, and layout information to segment several data points into a tabular structure. This proposed solution significantly reduces the manual hours spent in identifying and capturing required information by automating the financial variable extraction process for close to 70 000 PDFs per year in near real-time. It also includes the development of a robust metadata management system that indexes close to 150 variables for each financial document as well as a web application that allows users to interact with the extracted data points. Check out the recent Data Science Network article on Document Intelligence: The art of PDF information extraction.

In 2019, ISED began implementing a Virtual Assistant Technology with a product called a chatbot. Virtual Assistant is a communication channel Canadians can currently use when visiting ISED's webpages and mobile application. The Canadian Intellectual Property Office, Corporations Canada, Strategic Communication and Marketing Sector, and Canada Business App are using a web and mobile chatbot to help reduce calls to call centres and offer enhanced service when providing information to Canadians who visit their websites or applications. The current chatbot is built on a Microsoft architecture leveraging Microsoft Azure and Microsoft's Language Understanding AI, called LUIS. A similar presentation was held at the DSN's Chatbot Workshop, as the presenters explained the background and purpose of their Virtual Assistant technology. See Chatting About Chatbots: A review of the Chatbot Workshop.