---
title: "Reduced Transcription Automation"
slug: "reduced-transcription-automation"
updated: 2024-10-30T05:16:59Z
published: 2024-10-30T05:16:59Z
canonical: "help.hyperscience.ai/reduced-transcription-automation"
---

> ## Documentation Index
> Fetch the complete documentation index at: https://help.hyperscience.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Reduced Transcription Automation

## Issue

Transcription automation rates are lower than expected.

## Affected versions

- All versions

## Affected system configurations

- All system configurations

## Causes

- **Fine-tuning is trained automatically every day and is sensitive to fluctuations in training-record volume** – Due to the fact that fine-tuning is trained automatically every day, you can expect to see periodic decreases in automation followed by subsequent increases. Before further investigation, check whether the decrease in automation is large or prolonged based on the automation rates for at least the past two months.
- **Not enough QA tasks completed for a new layout** – If you do not complete enough QA tasks for a new layout, you can expect to see short-term decreases in automation. As you complete QA tasks on the new layout, the automation rate will likely rebound. If the automation rates do not rebound even after completing sufficient number of QA tasks on the new layout, the causes of poor performance or low confidence should be investigated.
- **Systematic keyer errors for particular fields during QA tasks result in decreases in automation** – Examples of systematic errors during QA tasks include:
  - A field is blank because the value was located outside the field, but a user annotated the value regardless.
  - Incorrect concatenation of transcription or removal of existing punctuation.
  - A field that Hyperscience does not support.

The above-mentioned systematic issues with QA annotation reduce automation and should be investigated.

## Solution

To debug automation issues, you need to aggregate the QA records that generate high-confidence errors. All investigations should target high-confidence errors around the date automation started decreasing.

Note that issues could start appearing a few days before the automation rates start decreasing.

In v32 and later, to find problematic fields, you can use the fine-tuning debug tool at */admin/forms_qa/finetuningmodelmeta/*. You can open this tool by clicking the **Debug** button for a given fine-tuning model.

All QA records with errors are sorted by confidence in the **Samples Where Machine is Wrong** section. In an additional table called **Grouped by template and field uuid**, you can find the same QA records with errors grouped by *template_uuid* and *data_type_uuid*. If you identify a field that causes more errors than others, this field could be considered problematic. Once you’ve identified all problematic fields, you can additionally filter the records by appending *?template_uuid=&data_type_uuid=* to the URL link where:

- needs to be replaced by the problematic *template_uuid*, and
- needs to be replaced by the problematic *data_type_uuid*.

By default, crops and machine transcriptions are disabled, but you can enable them by appending *output_type=stats* to the URL link. Enabling the crops and machine transcriptions allows you to further investigate the causes of the errors.

### Next steps

Once you’ve identified the causes of the annotation errors, you should address these issues with your keyers. You should ensure that the keyers annotate exactly what is present in the field. If you find fields that Hyperscience cannot handle properly, a layout redesign might be necessary.

We recommend against manually deleting QA records in bulk. These records expire with time, and the automation rates should rebound if you’ve taken steps to correct the problematic annotation patterns.

For further information and questions, contact your Hyperscience representative.