NLP Cloud - Add Language Processing to Your Apps

The Rise of No-Code Tools and the Need for Specialized AI Expertise

While user-friendly ‘no code’ platforms are empowering businesses to leverage computing power without requiring large internal technical teams to manage software configurations, accessing the most advanced technologies – particularly in the realm of ‘deep tech’ AI – still necessitates specialized knowledge and potentially significant investment in expertise.

French startup NLPCloud.io is addressing this gap by providing MLOps/AIOps – essentially a ‘compute platform as a service’ – focusing specifically on natural language processing (NLP). The platform executes queries on its own infrastructure.

Advances in Natural Language Processing

Recent advancements in artificial intelligence have spurred considerable progress in NLP, a technology that enables businesses to efficiently manage communications. This is achieved through automation of tasks such as Named Entity Recognition, sentiment analysis, text classification, summarization, question answering, and Part-Of-Speech tagging. This automation frees up human employees to concentrate on more complex and nuanced responsibilities.

However, it’s important to note that the majority of NLP research has been concentrated on the English language, resulting in greater maturity and wider availability of the technology for English-based applications. Consequently, the benefits of these AI advancements are not yet universally accessible.

Pre-trained NLP models for English are readily available for immediate use. Furthermore, open-source frameworks provide assistance with model training. Nevertheless, businesses seeking to implement NLP still require the necessary DevOps resources and skills.

NLPCloud.io: Production-Ready NLP Without DevOps

NLPCloud.io caters to organizations that lack the internal capacity for NLP implementation, offering a “production-ready NLP API” that promises to eliminate the need for dedicated DevOps personnel.

The API leverages open-source models from Hugging Face and spaCy. Clients can select from pre-trained models curated by NLPCloud.io, or they can upload custom models developed by their own data science teams – a feature the company highlights as a differentiator from services like Google Natural Language or Amazon Comprehend and Monkey Learn.

Democratizing NLP Access

NLPCloud.io aims to make NLP more accessible by enabling developers and data scientists to deploy projects “quickly and affordably”. The company offers a tiered pricing structure based on requests per minute, starting at $39 per month and scaling up to $1,199 per month for enterprise clients requiring a dedicated GPU for a custom model. A free tier is also available for testing models at lower request volumes.

“My experience as a software engineer revealed that many AI projects falter during the production deployment phase,” explains Julien Salinas, the founder and CTO. “Numerous excellent open-source models are now available, delivering impressive results. Therefore, the primary challenge now lies in efficiently utilizing these models in a production environment. This requires expertise in AI, DevOps, and programming – a significant hurdle for many companies, which prompted the launch of NLPCloud.io.”

Growth and Future Plans

Launched in January 2021, the platform currently serves approximately 500 users, with 30 subscribing to paid plans. The startup, located in Grenoble, France, currently operates with a core team of three, supplemented by independent contractors. Salinas intends to expand the team to five by the end of the year.

“Our user base primarily consists of tech startups, but we are also attracting larger companies,” Salinas shares with TechCrunch. “The strongest demand comes from both software engineers and data scientists. Some teams possess data science skills but lack DevOps expertise (or prefer not to allocate time to it). Others seek to integrate NLP capabilities without the expense of hiring a dedicated data science team.”

“Our clientele is diverse, ranging from solo startup founders to established companies like BBVA, Mintel, and Senuto, spanning various sectors including banking, public relations, and market research,” he adds.

Real-World Applications

Customers are utilizing the platform for applications such as lead generation from unstructured text – achieved through named entity extraction – and prioritizing support tickets based on sentiment analysis.

Content marketers are leveraging summarization capabilities for headline generation, while text classification is being employed for economic intelligence and financial data extraction, according to Salinas.

Addressing the Challenges of AI Implementation

Salinas’s experience as a CTO and software engineer working on NLP projects at multiple tech companies highlighted the challenges of translating successful models into practical applications.

“Building acceptable NLP models with frameworks like spaCy and Hugging Face Transformers is now relatively straightforward. However, deploying these models into production presents significant difficulties. It demands programming skills to develop an API, robust DevOps skills to build a scalable infrastructure, and, of course, data science expertise.”

“I explored existing cloud solutions to streamline the process, but found none that fully met my needs. I believed a dedicated platform could save tech teams substantial time, potentially months of work, particularly those without strong DevOps capabilities.”

The Commoditization of NLP

“NLP has been a field of study for decades, but until recently, building acceptable models required large teams of data scientists. Over the past few years, we’ve witnessed remarkable improvements in both the accuracy and speed of NLP models. Many experts who have dedicated their careers to NLP now agree that it is becoming a ‘commodity’,” he explains. “Frameworks like spaCy simplify the process of leveraging NLP models without requiring advanced data science knowledge. Hugging Face’s open-source repository further accelerates this trend.”

“However, running these models in production remains challenging, and potentially even more so given the resource demands of these new models.”

Model Selection and Future Development

NLPCloud.io prioritizes model performance, selecting models that offer “the best compromise between accuracy and speed”. Salinas emphasizes the importance of considering context, as NLP applications vary widely, and therefore offers a range of models to suit different use cases.

“We initially focused on entity extraction models, but customer feedback led us to expand our offerings. We will continue to add models from spaCy and Hugging Face to support a wider range of applications and languages.”

The choice of spaCy and Hugging Face was based on their established reputations, the quality of their NLP libraries, and their commitment to production-ready frameworks. This combination allows NLPCloud.io to provide a selection of models that balance speed and accuracy.

“SpaCy, developed by Explosion.ai in Germany, is a widely used NLP library among companies deploying NLP in production. It’s known for its speed, accuracy, and opinionated framework, which simplifies use for non-data scientists, although it offers less customization. Hugging Face has disrupted the NLP landscape with its ‘transformers’ library, significantly improving model accuracy, albeit at the cost of increased resource consumption. Their open-source repository simplifies model selection.”

Addressing Bias in NLP

While AI is rapidly evolving, particularly in areas like NLP for English, potential pitfalls remain. Automated language processing and analysis can be susceptible to errors and biases embedded in the data used to train the models.

Salinas acknowledges the potential for “concerning bias issues”, such as racism and misogyny, in NLP. However, he expresses confidence in the models selected by NLPCloud.io.

“Bias often stems from the underlying data used to train the models, highlighting the need for careful data sourcing. The best solution is for the NLP community to actively report inappropriate behavior when using specific models, allowing for prompt correction.”

“Even if we are confident that our models are free of bias, we encourage users to report any concerns so we can take appropriate action.”