Email is one of the highest volume customer engagement channels for many organizations. Customers email businesses to ask billing questions, buy new products, register complaints, and understand details about their account. Often these emails are sent to a centralized email address:
info, support, or something similar, and need to be routed to the correct team for action to be taken, consuming anywhere from 2-5 minutes per email for an agent to manually review. This can create a bottleneck in resolving critical issues, lowering customer satisfaction, and increasing costs.
AI can address this bottleneck by analyzing emails to understand their context and intent, then automatically forward them to the correct department for resolution. However, large language models can determine this with greater accuracy than smaller models, such as BERT. One way in which a large language model like GPT can better solve this by utilizing a larger sequence length. Sequence length refers to the ability to analyze more information from the document at once, to generate better context and understanding.
To understand how this works in the real world, try the GPT Challenge below:
When I discovered my credit score dropping in XX/XX/2020 I requested a free report from XXXX. I discovered the reporting by Mercedes-Benz Financial Services USA LLC (MBFS) has issues. On or about XX/XX/2020, I filed a dispute with XXXX and contacted MBFS disputing its negative payment rating. During that communication with MBFS, representative XXXX XXXX admitted MBFS erroneously reported negative history concerning my account to credit reporting agencies and volunteered that MBFS was aware that it had made the error with my account and the accounts of other customers.
We are conditioned to accept that standing up for oneself is a time consuming and ultimately futile act in which drains already limited time and energy. We are stopped in our tracks by fine print in agreement that we had no choice but to sign. Sign this or don’t go to school to better your life is not a choice. There is no shopping around for the best contract for those of us with few financial resources.
Capital One sent me a letter stating that the contact information used to open the account has been linked to me by public records. That is verification not validation, and that only means that this alleged account could have been opened by anyone. Capital One also sent over bank statements and an alleged charge that they assume was made by me, and that is also not validation. Anyone could have made those charges. Pursuant to 1692g no court nor debt collector can validate any debt, they can only assume that the alleged debt is valid. A belief or an assumption is not validation. Only I, the consumer who is a natural person, and the original creditor can validate this alleged debt. I do not validate this alleged debt, this alleged debt does not belong to me.
Longer sequence length allows the models to analyze more relevant information at a single time. By processing more information, such as the entire document at the same time, it can understand greater context and details about the document. This can result in greater accuracy when classifying the category of the document, improved understanding of the sentiment, and more relevant search results.