NLPs and the Hurdles Halting Their True Potential
With Generative AI (Gen AI) having thoroughly taken hold of enterprises’ imagination, Natural Language Processing (NLP) has emerged as a more commonly known noun for businesses across sectors.
From customer service chatbots to advanced data analysis, NLP is reshaping how companies interact with information and clients alike.
"NLP has become a lot more ubiquitous and has grown into something that drives user experiences, rather than 'only' enriching them," says Sebastian Gehrmann, Head of NLP in the Office of the CTO at Bloomberg.
Previously, the technology played a supporting role in organisations operations. But as its capabilities have grown, so have expectations of it. This change of position has put enterprise in position to reap its rewards, but also feel its wrath if left unchecked.
NLP’s evolution
NLP has come a long way since its inception in the late 1940s. What began as a niche field of AI has now evolved into a powerhouse technology with far-reaching implications for businesses across the globe.
"The last 2 years have seen a huge rise due to the Generative Transformer model progress (driven by huge amounts of computing, data and advancements in techniques to tune models) in the abilities of NLP," explains Brad Mallard, Chief Technology Officer at Version 1.
Modern NLP systems employ a variety of sophisticated techniques to understand and generate human language. These include large language models (LLMs), deep learning algorithms, and advanced neural networks.
These systems analyse vast amounts of textual data, considering context, sentiment, and linguistic nuances to provide human-like responses or extract valuable insights from unstructured information.
This progress has opened up new possibilities for businesses to engage with their data and customers in more natural and intuitive ways.
NLP’s performance
Text generation, question answering, text summarisation; text classification and named entity recognition are all some of the most notable NLP tasks that the system is currently capable of.
The applications are diverse and growing, ranging from automated report generation to sentiment analysis of customer feedback.
Despite these obvious utilities and indeed its current application, NLP faces significant hurdles in mission-critical enterprise applications.
The price of progress
"When NLP is the technology driving the entire user experience, failure is much more obvious and frustrating," Brad states.
This increased visibility of NLP in customer-facing applications raises the stakes for accuracy and reliability.
One of the primary challenges is the phenomenon known as "hallucinations" - where AI models generate plausible but factually incorrect information.
"These models produce answers that are based on internal algorithms that sometimes do not reach the output we expect,” explains Bern Elliot, Distinguished VP Analyst at Gartner. “This is called 'hallucinations' because the output is not based on clear evidence."
For instance, a hallucinating system could give incorrect information, like when a certain system advised the prompter to put ‘non-toxic’ glue on a pizza to help the cheese stick.
But these hallucinations equally pose a problem when being felt in internal operations too.
Imagine, for instance, an NLP system summarising a financial document incorrectly, or misinterpreting crucial legal documents.
The potential for reputational damage and financial loss is considerable. Yet, NLP’s ability to present this information is where the danger really lies.
“Since generated natural language sounds plausible and convincing even when wrong, there is overt danger in relying on it completely without verifying the output,” explains Sebastian.
This can be damaging enough in isolated incidents, but imagine these systems being used at the enterprise level and at the beginning of a project. One hallucination can go on to act as the foundation for whole campaigns or strategy.
Equally, the world has changed beyond recognition since NLP’s initial creation, when the Western world made up almost all of the globe’s computer and internet users.
Complexity of human language, including cultural nuances and contextual understanding, presents ongoing challenges for NLP systems.
"Contextual understanding is critical based on cultural norms, and we have to build models that are situationally and geographically aware if we are to ensure a more natural engagement between digital and human over time,” explains Brad.
This cultural and linguistic diversity adds another layer of complexity to NLP development and deployment.
Sebastian notes: "Excluding non-English speakers, or imbuing NLP systems with only the values and cultural context of a small development team, effectively excludes the majority of the global population."
This lack of understanding may preclude those of different cultures, and therefore different ways of operating, from getting the best out of these NLP systems.
The data dilemma: Quality over quantity
These issues mentioned above are focused on the application side, but really, their root stems from the same place: the data.
Accuracy and effectiveness of NLP systems heavily depend on the quality and diversity of the data they're trained on. But ensuring this data is representative, unbiased, and of high quality is a laborious and complex task.
For large tech companies with vast resources, this may not be as significant an issue. However, in the age of AI democratisation, where small startups are now providing NLP services, this process of data curation and model fine-tuning can be a significant challenge.
"There are many steps in the development lifecycle of NLP products,” Sebastian explains. “Training on raw unstructured textual data, tuning on human- or machine-created annotations, and deployment in live environments where the models may interact with knowledge graphs and metadata, extract information from structured databases, or be run on documents."
Because these systems are so data hungry, the quality of data is sometimes overlooked in order to meet the quantity these systems need.
Although pre-processing of such data can occur, Brad argues it would be better to be more mindful of the source: “The quality of the written or spoken words on Reddit or social media streams isn’t always the best quality and therefore a good use of data.”
Poor, incomplete or biased training data can lead to skewed results, perpetuating existing inaccuracies, flaws, biases, or creating new ones.
Halting hallucinations
As NLP technology continues to evolve, new approaches are emerging to address these challenges.
One promising solution is known as Retrieval-augmented generation (RAG). This technique aims to reduce hallucinations by grounding language models in verified information sources.
"RAG is a practical way to overcome the limitations of general LLMs by making enterprise data and information available for LLM processing,” explains Bern. “It is essentially a way to allow targeted information to be retrieved (often via a search) and then submitted as part of a prompt, to a LLM, enabling the LLM to provide focused answers."
Another area of innovation is the integration of NLP with other AI technologies.
Sebastian predicts that AI applications powered by NLP will help make us more effective at making decisions, with the convergence of technologies leading to “more sophisticated and capable AI systems” that can handle complex tasks across various domains.
Aside from managing models’ shortcomings with techniques such as RAG, new and improved models are what the industry is looking toward as a way to improve accuracy.
“One trend we are already starting to see is the capacity to handle more diverse input in models,” says Sebastian. “Another trend is the improved efficiency of models. We will see a continued trend to reduce this burden by training smaller models, for example by having a large teacher model that trains smaller knowledge-distilled versions of itself.”
Brad believes the change could be more fundamental: “We will likely see a change in the approach from transformer models to alternative approaches that are less about prediction of what comes next, to models that are more contextually grounded in world models or understanding of the world and its rules.”
NLP fully realised
NLP has already brought benefits to organisations and the individuals within them who benefit from their systems.
Yet, NLP is yet to reach its full, unfettered potential. “You can’t sustain development with compute alone, so improved techniques and maths will enable better efficiency – and efficiency, or doing more with less, will be the aim of the game for NLP.” explains Michael Conway, Executive Partner, AI & Business Transformation Services at IBM UK&I.
These efficiencies, alongside developments with new transformer models, could enable more complex tasks to be completed.
“Imagine A world where a query such as ‘Propose three strategies to invest in eco-friendly battery manufacturing’ creates associated research reports and offers to execute an investment strategy at the click of a button,” says Sebastian. “Or where ‘prepare my day’ leads to the curation of news articles to catch up on based on your portfolio holdings.”
As enterprises continue to integrate NLP into their operations, they must navigate the current balance of risks v rewards before them. The journey of NLP in enterprise applications is far from over.
“In future, NLP will become more efficient. Not only will we have more compute, but systems will be more trusted through organisations being more transparent and using trusted data for training models,” Michael concludes.
Success will depend on addressing current challenges head-on and embracing emerging solutions as the technology continues to evolve in order to capture the “true promise” of NLP.
To read the full story in the magazine click HERE
******
Make sure you check out the latest edition of AI Magazine and also sign up to our global conference series - Tech & AI LIVE 2024
******
AI Magazine is a BizClik brand
