Today's construction industry is predominantly built on an outsourcing model that prioritises flexible capacity over efficiency and productivity. A principle contractor will often outsource 80% or more of the work to subcontractors and this creates many interfaces between organisations with different systems. Whilst the future may be integration, for now the result is a patchwork of manual processes, often utilising low skilled labour. There is much work being done to agree standards for information management and exchange throughout the lifecycle of a building, however many processes occur indirectly to the asset. Agreeing the scope for an electrician to carry out an electrical installation, organising delivery of a crane, or ensuring that a subcontractor has a plan to deliver their project safely, for example.
Many of these processes involve the exchange of unstructured information such as e-mails, letters and other documents which present challenges to automation due to the variety and ambiguity of written or spoken language. This form of communication will persist, however. There is a limit to how much can be structured through forms and drop-down boxes; it simply isn’t natural to communicate many types of information this way.
A trend that has been gathering pace in recent years is the use of a computing technique called Natural Language Processing (NLP) that enables a machine to interact with the kind of language that we use in day-to-day discourse. Witness the rise of consumer tech products such as the Amazon Echo or Google Home whose smart assistants use NLP to enable us to play music or set a reminder by voice command alone.
Intuety, a startup based in Wales, have developed a product that brings NLP to the construction industry, changing the way we carry out some of these manual, unstructured processes. The goal is to build a language model based on deep domain knowledge gathered from the construction industry and related sectors, creating some ready-made applications of the technology.
The problem
The starting point is to improve a process that is often done badly and inefficiently, wasting money, but more importantly, putting lives at risk; the review and approval of Risk Assessments and Method Statements (RAMS), a mandatory requirement under CDM regulations.
The current norm is for RAMS documents to be produced with office applications using templates unique to the subcontractor and sent to the site manager for review. The documents often miss important risks and control measures whilst overloading the reader with superfluous information that distracts from the key safety messages. If the approver requires amendments, then managing the workflow is inefficient and lacking control.
Furthermore, because the documents are disconnected islands of information, obtaining overall trends and statistics about the activities being carried out and their associated risks cannot easily be acquired.
The solution
Intuety automates much of the process. The application ingests and reads the documents parsing the text into activities, risks and control measures. It then compares the risks and control measures for each activity against a curated library and makes recommendations where variations are identified. The user has a final say and ultimately accepts/declines any recommendations, but the key is that they have the aggregated, digitised knowledge of many projects at their fingertips; augmented intelligence, if you will.
Amendments and approvals are automated and logged, and the structuring of risk information enables trends and patterns to be gather across a company's projects such as the most common risks identified or the most commonly overlooked risks.
The application will even be able to generate a new risk assessment autonomously, with the user merely inputting the activities to be carried out and reviewing the risks and control measures generated.
Not as easy as it sounds
Getting a computer to assess whether the user has missed any risks or control measures for a given activity is not a trivial technical challenge. Consider the following examples:
Coming into contact with plant
Coming into contact with asbestos
Hit by moving equipment
1 and 2 are identical but for a single word, yet the control measures would be very different. 1 and 3 don’t share a single common word and yet the meaning is essentially the same. Therefore, rather than simply matching words, you need to extract a deeper meaning behind the words.
There are many different ways of doing this but they broadly fall in to two categories:
provide the machine with rules e.g. “coming into contact” and “hit by” are synonymous, as are “plant” and “equipment” however “plant” and “asbestos” are not; or
label 1 and 3 as being analogous along with n other examples and build a machine learning model to cluster similar phrases.
Intuety are using a combination of rules-based and machine learning techniques to optimise the trade-off between the accuracy but lack of breadth of rules-based approaches, with the intuitive but less certain results (particularly with smaller datasets) that accompany machine learning.
The early signs are very encouraging, and we are now just weeks away from trialling the product.
Comments