This browser is not actively supported anymore. For the best passle experience, we strongly recommend you upgrade your browser.
| 6 minutes read

Cobots use Natural Language to Deliver Business Value

Twenty-five years ago, IBM’s Deep Blue artificial intelligence system defeated chess Grand Master Gary Kasparov. Deep Blue became the first computer to defeat a world champion under standard chess tournament rules, marking an historic milestone for AI.

Kasparov’s six-game duel with Deep Blue struck an adversarial tone of human vs. machine. However, with natural language processing (NLP) entering the business mainstream, a different approach is now evident: cobots and people working back and forth together. We see widespread adoption of natural language applications at clients. In most cases, cobots bring together machines’ ability to sift through large volumes of data with decisions made by people.

NLP cobots have a similarity with chess: using a few basic principles (or “moves”) that combine to achieve complex goals. At root, NLP can be viewed as consisting of four basic moves that identify:

  • Entities such as people or companies
  • Topics forming the subject of a text
  • Sentiment expressing affect
  • Word relationships measuring how closely words tend to be used together and their positions in a text

As an example, consider a bank contact center that receives a SMS text: “I need to make an offer on a great house ASAP”. The entity is the author, topic is an offer, sentiment indicates positivity, and word relationships can map “ASAP” to the concept of acting fast. The business action can be to fast-track underwriting a loan.

Natural language applications bring together the four base concepts to form higher level actionable insights such as summaries, timelines or finding conflicting factual claims. NLP can answer questions such as:

  • How have perceptions of our products changed over time?
  • What is happening at ports that may affect our supply chain?
  • Do any available news items, filings, or social media suggest any of our creditors may be violating loan covenants?
  • Which topics are regulators focusing on in Europe compared to the US?

AlixPartners has collaborated with clients across several industries using natural language applications. We use the 4D framework to guide design across:

  • Detect relevant text data
  • Decipher the business meaning
  • Decide relevant actions
  • Do actions or workflows

Examples of how clients have achieved business goals using NLP:

1. Global architecture and design firm: What are the top reputation risks in our industry and to our brand? 

The client had grown through engineering excellence and a reputation for delivering complex assignments on time and budget. However, the client found that government ministries and banks increasingly show sensitivity to how the industry, the client, and specific projects are viewed in media. Additionally, the client found stiff competition for recruiting engineering talent and wanted to understand how potential employees viewed the firm.

We applied the 4D framework:

  • Detect: ingesting hundreds of thousands of articles across news, social media, and web content
  • Decipher: identifying topics, authors, and publications most affecting sentiment
  • Decide: identifying the best course of action including when and how to respond
  • Do: orchestrating workflows to implement responses and ongoing proactive improvements to communication

AlixPartners applied our rapid AI development methodology and platform to jump start building the solution. The system follows the cobot approach where the machine brings distilled insights to people who then confirm analyses and responses.

2. Post-merger integration of packaging products companies: How do we rapidly create a single product master file?

NLP can enable newly combined companies to move faster. Optimizing product hierarchies and inventory management offers an example. A client grew through acquisitions providing packaging products such as boxes, foam, and bags. While each acquisition manufactured and distributed similar products, they also use used different names. After multiple acquisitions, the client had several hundred thousand SKUs with inconsistent descriptions. They needed to establish a consistent product hierarchy to enable inventory visibility, provide manufacturing and distribution footprint visibility, and optimize integration with suppliers.

AlixPartners combined semantic search, which finds matches based on words’ meaning instead of spelling, with a custom word relationship (embedding) model to identify related products. The client team extended FastText – a technology initially developed by Facebook that represents word meaning as vectors (numbers) – to allow machine learning algorithms to understand them and work within the client’s business domain of packaging-related terms.

The cobot approach was key, drawing on client input and implementing subsequent iterations, covering misspellings, global languages, and other needs in a series of custom-trained models. Business results include reducing aged/E&O inventory, improving order fulfilment timing and completeness, optimizing warehouse capacity and improving strategic sourcing.

3. Leading private equity investor: outside-in due diligence on a prospective investment

Employees and job candidates share data on social networks and employer review sites that provide a window onto company leadership, culture, and practices. Natural language techniques can answer questions such as:

  • What are the drivers of turnover?
  • How has sentiment changed with new senior leadership?
  • What are the estimated size and ratios of corporate functions compared to benchmarks?
  • Does a target company use modern technology?

These data points contribute to analytics to identify value creation theses and plan implementation.


What’s next for business application of natural language understanding?

  • Large language models that map word relationships increasingly reflect nuances of how people communicate. These models first became commercially useful in 2018 with Google’s BERT for generating marketing content, chat bots, and other applications. Models are often measured by the number of logical “cells” (layers, hidden size, and attention heads) that work together to model understanding. BERT uses 340 million cells. More recently, Microsoft and Nvidia teamed up on the Megatron-Turing model with 540 billion “cells” – and Microsoft is working on a model called 1T with one trillion parameters. These models contribute to new applications such as real-time translation and instantly generating highly personalized marketing messages.

  • Multi-modal models combine natural language and visual understanding-and potentially auditory as well. Multi-modal models recognize both what people say and how they say it. This approach incorporates the range of verbal and visual signals that people use communicating. For example, researchers are training multi-modal models to triage patients for emergency rooms.

  • Training models on specific industries and business functional areas presents further opportunity. To date, most new natural language research focused on general purpose language, which means the business applications they enable tend to center on consumers. For example, ULMFit, a popular word relationship model, uses Wikipedia as the default corpus, or body of text, to connect word relationships.

    Instead, language models can be trained on business topics. Consider a model based on several hundred thousand accounts payable invoices. This model could identify aberrations such as potential fraud, find outliers and link to causes, and contribute to working capital optimization models.

    One example of domain-specific models in action is accelerating how software developers write code. Word relationship models can predict how to complete a line of code or even a whole component that a developer has started. For example, GitHub is a source code repository and collaboration tool widely used among developers. OpenAI, a research company funded by a $1 billion investment from Microsoft, built models trained on code in GitHub. These models serve as a virtual companion making suggestions as an engineer writes code. However, like any AI model, the suggestions are only as good as the data used to train the model. This means bugs or other issues in the code used to train the model could potentially be found in the suggestions.

  • One area where AlixPartners sees a pattern of need is controls and content moderation for digital platforms where people interact: crypto exchanges, gaming, metaverse, and social media. Each share the need to confirm users’ identity and that activity on the account matches the identity used to open the account. These platforms also share the need to restrict harmful content and behavior, and to use confirmed positive examples of bad content or behavior to find other bad actors. AlixPartners uses customized NLP models to protect these businesses and fulfill compliance with new mandates, such as the Digital Services Act in the European Union.

In the not-so-distant future, AI models will understand the state of an entire business and converse naturally with people. For example, a CFO could ask “Hey, FinanceBot, if we reduced day sales outstanding by two days, what would that mean for working capital?” or a CMO could ask “MarketingBot, what’s the best way to improve conversion?”.  NLP models will be subsumed into broader multimodal models and linked together in ensembles that comprehend industries and functional areas like marketing or finance. These models will further the cobot partnership, pairing instant understanding of what has and could happen with human judgement and creativity.

For a deeper discussion, please contact Jeremy Lehman or Amelia Green. With additional thanks to Greg Adams, Jonas Berlin, Nico Bottini, Rob Cerff, Charlie de Montfort, and Chris Pocek in developing this article.

Tags

ai, cobot, nlp, natural language processing, digital
post featured image

Latest Insights