AI for E-Mails
Open sourced GPT Neo writes individualized emails
GPT Neo model can certainly dispatch rule-based patterns directly, but it is by no means a plug-and-play application that is suitable for all possible use cases without training. We explain why creativity is still needed to create well written e-mails.
We all might have seen autocomplete suggestions for writing emails as GMail offers. This solution uses machine learning to create suggestions as you type and already offers little improvement for the tiresome writing and responding to emails every day. But whenever someone wants to generate not only parts but entire emails for sales or marketing campaigns that are sent out to update our followers in a very personalized email, other solutions are needed to save time.
There are different ways to create those automatically generated and personalized emails. For example, the usage of rule-based email marketing software or even a simple Excel sheet can help with assembling text sections. The latter was also used by our sales team for sending out contact emails after attending conferences or forums. However, this makeshift solution takes a long time to set up when aiming to address the recipient correctly, include a reference to an attended event, or add a personal note to the email also with regards to the correct grammatical usage.
That is why we were looking for an alternative that we could implement quickly and no longer needed a rule-based solution for. The publication of Open-AI researchers on the GPT-3 language model in June 2020 gave us the appearance of being able to generate our emails. The GPT-3 in simple terms describes a language processing model with 175 billion parameters that supplements given individual words, sentences, or entire paragraphs with further text that is considered likely in the respective context. [1] Since access to this model is restricted, we used the open-source GPT-3-like model GPT Neo [2]. It was actually trained on more data than the original GPT-3 (~825GB compared to ~570GB) but works with fewer parameters (2.7 billion compared to 175 billion). Therefore, it is assumed that the GPT Neo project will only deliver an AI with fewer capabilities. So, we wanted to know how simple the task needs to be since it seems like the open-source version is not a simple plug-and-play solution for everything. Can the largest open-source alternative to GPT-3 write our emails? We would like to share the results with you now.
GPT Neo was able to automatically generate personalized emails with individual fields like surname, name, title, company, and reason of contact
The training data was created by using emails in PDF format stored in the archive:
The model was immediately able to create emails by only providing four initial examples including the tokens surname, name, title, company, and reason of contact. Here is an excerpt of the anonymized test data of emails in the prompt:
surname: Stromberg
name: Bernd
title: Herr
reason: our discussion yesterday at the Helios Congress
companies: Capitol Versicherung
email:
Good afternoon Mr. Stromberg,
does AI also add value in your daily practice? This is a question we answered with yes
in our discussion yesterday. It was great to see you again at the Helios Congress. As
discussed we would appreciate this as a good starting point for real digitalization with AI
at Capitol Versicherung. In addition to our software Konfuzio, we are also happy to bring
our AI knowledge to your individual project so that your company can also write further success stories.
Are you interested to discuss things further? Feel free to use the following link to book an appointment with me:
https://calendly.com/konfuzio/meeting
surname: Scott
name: Michael
title: Herr
reason: our discussion yesterday at the IHL Messe
companies: Paper Company
email:
After just providing the necessary fields described above, the model was creating this email:
Good afternoon Mr. Scott,
does AI also add value in your daily practice? This is a question we answered with yes
in our discussion yesterday. It was great to see you again at the IHL Messe. As
discussed we would appreciate this as a good starting point for real digitalization with AI
at Paper Company. In addition to our software Konfuzio, we are also happy to bring
our AI knowledge to your individual project so that your company can also write further success stories.
Are you interested to discuss things further? Feel free to use the following link to book an appointment with me:
https://calendly.com/konfuzio/meeting
The model was able to create a new email by exchanging the name, the company, and the event. Also, the model was right in distinguishing between Mr. and Mrs. as two examples of both types of greetings were provided in the prompt. There was no change of words or additions made in the rest of the text. So we see that it does work to quickly replace our rule-based system for writing customized emails.
GPT Neo even works better with German compared to the German GPT-2 model
We usually write most of our emails in German and therefore we have also tested the German version of the model which is the GPT-2. This didn’t work at all and did not generate any output. However, even the GPT Neo works better with German inputs than the German GPT-2 model. This could be explained by the fact that this model is older and has been trained with considerably fewer data. According to the team who trained the German GPT-2 model, it “(…) is meant to be an entry point for fine-tuning on other texts, and it is definitely not as good or “dangerous” as the English GPT-3 model.” [3] The GPT Neo is trained on less German data, but on more data overall, which could explain the general functionality for German input.
Nevertheless, the GPT Neo only gets completely correct results with English inputs. So to generate our German emails automatically, we still have to take the workaround of translating German into English, and after the email generation from English back into German.
What is the benefit of automatically generating emails with GPT Neo?
As we could see, the rule-based system previously used to generate emails can be immediately replaced by using GPT Neo without any adjustments and without having to train the model on our specific application beforehand. But as said, this only covers the rule-based emails so far. If more complex use cases e.g., more individualization in the text or grammatical changes have to be added, it requires further training on the individual data set which can be done with the help of the AI specialists of Konfuzio.
Conclusion
In summary, we can say that the GPT Neo model can certainly dispatch rule-based patterns directly, but it is by no means a plug-and-play application that is suitable for all possible use cases without training. So if we now let the model write our emails, the very big advantage is that no big dataset is needed, we quickly have our results, and we don’t have to deal with the maintenance of the rule-based system. But of course, there is still room for improvement and need for efficient use of other languages as well. But keep in mind to revise any automated generated email manually before sending. Over the years we have collected means to increase the response rate drastically from 3.5 % to over 15. Automatically creating prebuild drafts is only one step in this journey. Get in touch with us, if you have an email project you would like to automate.
Sources:
[2] EleutherAI (2021). GPT Neo, retrieved from GitHub.
[3] Dbmdz (2020). German GPT-2 model, retrieved from Hugging Face.