NLP Case Study

Serokell developed a solution that allows the customer to set precisely targeted ad campaigns automatically.

Our client, an international financial investment consulting and marketing group specializing in connecting hedge funds, asset managers, and institutional investors, developed an in-house advertising platform.

The client was looking for a solution that could automatically define the campaign's target audience based on requests written in simple language. The functionality should be easily accessible for setting up a large number of ad campaigns simultaneously.

Serokell automated audience segmentation for the platform using natural language processing (NLP) technologies. The tool can identify desired audience segments through keyword analysis and determine optimal targeting with a high degree of flexibility, all without manual input or human involvement.

The main goals were:

Automating ad campaign settings

Providing a flexible tool for selecting target groups

Increasing ad conversion

Reducing marketing expenses for each market segment

Increasing the number of targeted leads

Innovative solution

Setting up advertising campaigns on the client’s platform was a complicated process involving dozens of operations to define the target audience. This process became even more complex and time-consuming when creating ads for multiple types of campaigns, each with different settings and requirements.

Serokell provided an innovative solution to this problem: automatic AI-driven targeting, which identifies audience parameters from queries written in plain English.

We have taught the model to split a query into structured language fragments and recognize the main components. Then, each query from the client is translated into campaign settings. The user does not have to fill out any forms: all information is extracted from a simple query written in plain conversational English.

Development challenges

The main difficulty was the requirement for these campaigns to use different semantic structures and identify audiences and target subgroups based on written language patterns rather than a predefined set of forms.

Understanding context and nuance

Sarcasm, idioms, or industry-specific jargon are difficult to interpret by NLP algorithms, leading to inaccurate targeting or model overfitting.

Keyword ambiguity

Keywords can have multiple meanings depending on the context. AI models must accurately interpret the intent behind a user’s search query to match it with the most relevant idea.

Integration complexity

Integrating NLP analysis into existing PPC management tools and workflows is a complex task. Our team has deeply explored both technical and strategic aspects.

Overcoming difficulties

As there was no pre-built solution meeting the client's requirements, and the introduction of a large language model (LLM) exceeded the customer’s budget for this project, Serokell designed custom software with an ML core that operates  on top of the existing layers. To address these challenges, along with testing multiple models, continuous refinement of NLP models with human supervision was implemented.
At the beginning, we encountered a difficulty that slowed our progress: insufficient labeled raw data. Collecting enough data is a traditional challenge in ML modeling. In our case, thousands of test data records were required. To overcome this issue, we developed a dataset generator.
To ensure that there were no significant discrepancies between the customer’s datasets and the generated records, we compared the artificially created ones  with real data provided by the client. The quality of both was found to be identical.
With enough data available, we trained several models, compared the results, and selected the most effective one. After successfully verifying the NLP model with new data, the campaign automation process received confirmation of the required accuracy and relevance, leading to the implementation of this model in production.

Stages of the project

Developing the overall logic

Our project aimed to extract criteria for the target audience from a simple, human-written request, specifically to recognize logical operators such as or/and/not in the query. The first phase involved developing the logic to recognize all the expressions related to conjunctions, disjunctions, and identities and translate them into queries. The objective was to develop functionality for automatically defining market segments based on customer descriptions using AI models.

Language processing was necessary to recognize patterns and segment the audience into smaller groups based on their interests. By carefully dividing these groups, each with similar interests, the ad campaign can effectively target their needs and encourage participation in financial activities, whether investments or loans. A well-optimized campaign enables the company to achieve successful outreach with significantly reduced effort and time and minimized conversion costs per lead.

In particular, at this stage we mapped out
the general framework for the future solution:

Implement semantic analysis of written queries for advertising campaigns using NLP methods.

Determine the key components of the queries and the way they affect the segmentation settings.

Identify keywords related to target audiences and determine overlapping (conjunction) and non-overlapping (disjunction) sets.

Later, we progressed to data labeling, selecting NLP models, and teaching them.

Providing the solution

There was no ready-to-use framework on the market for our needs.
Therefore, to extract the necessary information from customer queries, we wrote a parser capable of recognizing linguistic structures, shingles, and keywords for analysis. This was a critical missing component needed to accomplish the task. Once a test dataset was ready, the parser could extract patterns and apply them to different ML models for analysis and segmentation. We employed several ML models:

  • Stanza and Trankit for sentence parsing;
  • MPNet, RoBERTa, XLM-R for segmentation.

After conducting a series of experiments, we determined that Sentence-Bert was the most suitable model to utilize with real data for the client's application.

Technical stack

Our accomplishments

The new parser efficiently structures raw data. The implemented ML model successfully classifies objects from the provided customer text queries and matches them with one of 400+ predetermined segments for ad campaign targeting.

Read more

Serokell's solution provides a tool that functions like a team of professional ad marketers, reducing the need to expand the staff to meet the company's increasing demand for advertising its financial products to a wide audience. It offers customization for PPC campaigns targeted at specific segments, whether they are investors, borrowers, or job seekers.

Serokell's key achievements in this project include:

  • Designing a system that enables the highest level of customer satisfaction without human interaction;
  • Ensuring a stable and smooth operation of the system with an accuracy of over 89%;
  • Automation in designing and implementing advertising campaigns led to a 35% decrease in campaign cost, while proper targeting increased lead flow.

By introducing this NLP-driven solution,
the customer achieved:

Increased efficiency and time savings

Automation streamlines the PPC campaign setup process, enabling PPC specialists to operate successfully with an increased number of users and attract twice as many leads with the same budget.

Enhanced targeting precision

Leveraging NLP for keyword analysis and audience targeting resulted in more relevant ad placements, improved click-through rates (CTR), and higher conversion rates.

Cost-efficiency

Automation helped optimize bidding strategies and budget allocation in real time, covering hundreds of low-competitive keywords and ensuring that ads are properly targeted. This led to lower cost per acquisition (CPA) and higher return on ad spend (ROAS).

The significant reduction in lead costs was achieved through a combination of intelligent automation, strategic optimization, and data-driven decision-making.

Results

The 35% decrease in campaign cost directly translates into a higher ROI for the advertising budget, enabling the allocation of resources to other strategic areas.

Moreover, through proper targeting, the lead flow increased twofold. This cost efficiency enhances the scalability and sustainability of marketing efforts, influencing the overall company ROI.

Let’s Have a Talk

We would be happy to discuss your project and propose solutions.

Contact us