For our client from Southern Germany we designed, extensively tested and now deployed to the cloud a production system which each day transforms ten-thousands texts within a few hours and reports them back. The client sent short texts are automatically transformed using ChatGPT from OpenAI. As part of this process ChatGPT answers are linguistically checked for undesired words whose presence triggers regeneration of such texts. To reach the required processing speed while staying below load limitations, we use a high degree of parallelization on different architectural levels.

As already reported here in a joint research phase with a leading northern German company in the field of technical documentation LangTec has been working on a AI-based authoring support system for text completion in the text domain of technical documentation.

Now we can report extremely convincing results for this proof of concept. Among other results we were able to improve prediction accuracy on the next word by 45 (top 10 predictions) to 62 (top 1) percentage points due to finetuning the base language model on client data. Additionally, this project allowed LangTec to use and extend its expertise in the field of neural language models, especially with respect to deeper interventions in the standard software architecture of the language model.

This year the weather did not mean well with us at first: It took us three attempts, but on that sunny day it worked out and we met at the Supper Club Hamburg for cool drinks and water sport activities. With an SUP, a canoe and a pedal boat we toured over the Alster and met Irish musicians and the waves of the Alster steamer on the way… Everyone stayed dry and so in the evening we moved towards Schanze to the restaurant JING JING. An excellent, super flavourful 4-course menu propelled us straight to Thailand! Now what exactly was part of the concept, and what not, remains open, but we can say for sure that we had a fantastic time 🙂

Technical documentation requires a specific sublanguage by convention and regulation, which poses a challenge when different technical writers must create new documents in a coherent manner. So far, our client employs rule-based systems that enforce consistent structure and style, which can only detect problems that trigger manually-created style checking rules.

In a joint research phase, LangTec will now develop the proof of concept for an authoring support system that offers text completion features in document creation without the need to specify any explicitly defined rules up-front.

The goal of this AI project is to support technical writers by reducing any repetitive manual work. The central task will be to suggest the most fitting continuation of sentences when creating new technical documents. LangTec will select one of the the available very large language models, perform a domain-specific finetuning based on a large number of existing documents, and formally evaluate the resulting model’s prediction accuracy.

In order to increase the search engine relevance of text extracts, we are investigating the paraphrasing competence of the language model ChatGPT for one of our customers. The goal is to automatically reduce full texts (text summarization) and to generate text snippets using paraphrasing that are evaluated by the search engine as “unique” as possible. By optimizing the textual uniqueness, the snippets should be ranked as high as possible in the hit list in the web search and thus lead to a better conversion.
For the implementation of this project, LangTec specifically designed and operationalized a measure to quantify textual uniqueness. In this project, LangTec thus also contributes its computational linguistics expertise and many years of experience in the automated generation of texts.

For a provider in the field of recruiting, we successfully implemented the generation of sample job advertisements using the large-scale language model ChatGPT. The texts generated with our solution serve recruiting companies in the creation of their job advertisements as a starting point for further individual adaptations.
For this purpose, LangTec developed a web application as well as suitable instructions, so-called “prompts,” for the language model of OpenAI. The prompts also included textually relevant information from a domain-specific database, which should be considered by ChatGPT in the output.
After having implemented extensive text generation solutions based on template-based text generation (NLG) over many years in the past with our in-house developed solution TextWriter, we are now happy to use generative language models in our text generation projects with this project.

To celebrate LangTec’s 12th anniversary, we met for brunch today at Pynk Coffee, a cozy café close to the office. While we tasted delicious macarons, croissants and cakes, we reminisced about past projects and also shared the excitement about upcoming projects. Accompanied by the one or the other cup of coffee or tea, we really enjoyed ourselves..

Our client is aiming to prequalify suitable investment targets. In a yet another project involving large transformer-based language models, LangTec has developed a solution to identify all relevant company types based on company website information. The key challenge in this task is to deal with huge amounts of website content whose length exceeds the typical sequence length limitation posed by transformer-based language models. LangTec’s solution was optimised for recall, i.e., it was designd to capture all potentially interesting companies in the training and test set.
In addition to designing, training and optimising the perfect-recall classifier, LangTec successfully trained a hybrid language model that uses features from a another, non-neural-network statistical model along with features from the transformer-based model to arrive at a joint classification decision. This model architecture permits to combine transformer-based models with other machine-learning models in a hybrid architecture.

In a new project, LangTec gets to shine with its expertise in the areas of Data Science, Machine Learning and Big Data – and this time, actually quite directly. We are pleased to have been awarded a new contract to create comprehensive training materials for an advanced training program for adults. Specifically, we are designing documents for the modules Data Science, Machine Learning and Big Data.

We would be very happy to make our expertise in these areas available to anyone interested in gaining a foothold in this exciting field and learning advanced techniques.

We celebrated St. Nicholas Day with plenty of chocolate!

We have not managed the whole kilogram, but fortunately the Christmas season is not yet over…. 🙂