概要

主な手順

[探索] ページにはさまざまなトレーニング モードがあり、このフェーズでは主にそのうちの 3 つに焦点を当てます。

• Shuffle - Shows a random selection of messages for users to annotate. Make sure you complete a significant chunk of training in Shuffle, in order to create a training set of examples that is representative of the wider dataset.
• Teach - Used for for unreviewed messages. As soon as the platform makes some reasonable predictions for a label, you can improve the ability of predicting the label for more varied examples by reviewing messages in the default Teach mode, which is for unreviewed messages. This will show you messages where the platform is unsure whether the selected label applies or not.
• Low Confidence - Shows you messages that are not well covered by informative label predictions. These messages will have either no predictions or very low confidence predictions for labels that the platform understands to be informative.

This section will also cover training using Search in Explore, similar to training using Search in Discover.

Teach, for reviewed messages, is another training mode in Explore. For more details, check Refining models and using Validation.

レイアウト

The layout from the previously shown image is explained in the following table:

1	Adjust the date range or period of messages shown.
2	Add various other filters based on the metadata of the messages, e.g. score or sender.
3	Add a general field filter.
4	Toggle from all messages to either reviewed or unreviewed messages, also adjusts pinned vs predicted label counts.
5	Add a label filter.
6	Search for specific labels within your taxonomy.
7	Add additional labels.
8	Expand message's metadata.
9	Refresh the current query.
10	Switch between different training modes such as recent, shuffle, teach and low confidence, and select label to sort by.
11	Search the dataset for messages containing specific words or phrases.
12	Download all of the messages on this page or export the dataset with applied filters as a CSV file.

How much training to do for each label

各ラベルを正確に予測するために必要な例の数は、ラベルの概念の幅や特異性によって大きく異なります。

ラベルが一般的に、非常に具体的で簡単に識別できる単語、語句、意図に関連付けられている場合があります。この場合、プラットフォームは比較的少数のトレーニング例で一貫性を持ってそのラベルを予測できます。また、ラベルが幅広いトピックをキャプチャし、そのトピックに関連する言語のさまざまなバリエーションがある場合もあります。この場合、ラベルを適用すべきインスタンスを一貫性を持って識別できるようにするには、はるかに多くのトレーニング例が必要になる可能性があります。

The platform can often start making predictions for a label with as little as five examples, though in order to accurately estimate the performance of a label, that is, how well the platform is able to predict it, each label requires at least 25 examples.

When annotating in Explore, the little red dials next to each label indicate whether more examples are needed to accurately estimate the performance of the label. The dial starts to disappear as you provide more training examples and will disappear completely once you reach 25.

This does not mean that with 25 examples the platform will be able to accurately predict every label, but it will at least be able to validate how well it can predict each label and alert you if additional training is required.

During the Explore phase, make sure that you have provided at least 25 examples for all of the labels that you are interested in, using a combination of the steps mentioned previously, that is, mostly Shuffle, and Teach and Unreviewed.

During the Refine phase it may become clear that more training is required for certain labels to improve their performance. For more details, check Refining models and using Validation.

ラベルのパフォーマンスに関する警告

In Explore, once you reach 25 pinned examples for a label, you may notice one of the following label performance indicators in place of the training dial:

• Grey is an indicator that the platform is calculating the performance of that label. This means that it will update to either disappear, or an amber or red circle once calculated.
• Amber is an indicator that the label has slightly less than satisfactory performance and could be improved.
• Red is an indicator that the label is performing poorly and needs additional training or corrective actions to improve it.
• If there is no circle, the label is performing at a satisfactory level, though it may still need improving depending on the use case and desired accuracy levels.
• To understand more about label performance and how to improve it, check Understanding and improving model performance.

予測されたラベル数とピン留めされたラベル数

If you select the tick icon, as shown in the following images, at the top of the label filter bar to filter to reviewed messages, you will be shown the number of reviewed messages that have that label applied.

If you select the computer icon to filter to unreviewed messages, you will be shown the total number of predictions for that label, which includes the number of reviewed examples too.

[探索] では、確認済みまたは未確認のメッセージのどちらも選択していない場合は、既定で、ラベルにピン留めされたメッセージの合計数が表示されます。[レポート] では、既定で、予測された合計数が表示されます。

Note: The predicted number is an aggregation of all the probabilities that the platform calculates for this label. For example, two messages with a confidence level of 50% would be counted as one predicted label.

Tips for using Explore

• The model can make predictions with only a few annotated messages, though for it to make reliable predictions, you should annotate at a minimum of 25 messages per label. Some will require more than this, it will depend on the complexity of the data, the label and the consistency with which the labels have been applied
• In Explore, you should also try and find messages where the model has predicted a label incorrectly. You should remove incorrect labels and apply correct ones. This process helps to prevent the model from making a similar incorrect prediction in future

Important: During this phase you will be applying a lot of labels, so make sure to adhere to the key annotating best practices of adding all labels that apply. You can do this by applying the labels consistently, and annotating what you can view in front of you.