You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
> For single-label classification models, the count of false negatives and false positives are always equal. Custom single-label classification models always predict one class for each document. If the prediction is not correct, FP count of the predicted class increases by one and FN of the actual class increases by one, overall count of FP and FN for the model will always be equal. This is not the case for multi-label classification, because failing to predict one of the classes of a document is counted as a false negative.
92
+
> For single-label classification models, the number of false negatives and false positives are always equal. Custom single-label classification models always predict one class for each document. If the prediction is not correct, FP count of the predicted class increases by one and FN of the actual class increases by one, overall number of FP and FN for the model will always be equal. This is not the case for multi-label classification, because failing to predict one of the classes of a document is counted as a false negative.
93
93
## Interpreting class-level evaluation metrics
94
94
95
95
So what does it actually mean to have a high precision or a high recall for a certain class?
Copy file name to clipboardExpand all lines: articles/ai-services/language-service/question-answering/concepts/best-practices.md
+4-4Lines changed: 4 additions & 4 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -5,7 +5,7 @@ ms.service: azure-ai-language
5
5
author: jboback
6
6
ms.author: jboback
7
7
ms.topic: conceptual
8
-
ms.date: 11/21/2024
8
+
ms.date: 06/06/2025
9
9
ms.custom: language-service-question-answering
10
10
---
11
11
@@ -55,7 +55,7 @@ Users can add as many alternate questions as they want, but only first 5 will be
55
55
56
56
Semantic understanding in custom question answering should be able to take care of similar alternate questions.
57
57
58
-
The return on investment will start diminishing once you exceed 10 questions. Even if you’re adding more than 10 alternate questions, try to make the initial 10 questions as semantically dissimilar as possible so that all kinds of intents for the answer are captured by these 10 questions. For the project at the beginning of this section, in question answer pair #1, adding alternate questions such as “How can I buy a car”, “I wanna buy a car” aren’t required. Whereas adding alternate questions such as “How to purchase a car”, “What are the options of buying a vehicle” can be useful.
58
+
The return on investment will start diminishing once you exceed 10 questions. Even if you’re adding more than 10 alternate questions, try to make the initial 10 questions as semantically dissimilar as possible so that all kinds of intents for the answer are captured by these 10 questions. For the project at the beginning of this section, in question answer pair #1, adding alternate questions such as “How can I buy a car,” “I wanna buy a car” aren’t required. Whereas adding alternate questions such as “How to purchase a car,” “What are the options of buying a vehicle” can be useful.
59
59
60
60
### When to add synonyms to a project?
61
61
@@ -67,11 +67,11 @@ For better relevance, you need to provide a list of acronyms that the end user i
67
67
*`ID` – Identification
68
68
*`ETA` – Estimated time of Arrival
69
69
70
-
Other than acronyms, if you think your words are similar in context of a particular domain and generic language models won’t consider them similar, it’s better to add them as synonyms. For instance, if an auto company producing a car model X receives queries such as “my car’s audio isn’t working” and the project has questions on “fixing audio for car X”, then we need to add ‘X’ and ‘car’ as synonyms.
70
+
Other than acronyms, if you think your words are similar in context of a particular domain and generic language models won’t consider them similar, it’s better to add them as synonyms. For instance, if an auto company producing a car model X receives queries such as “my car’s audio isn’t working” and the project has questions on “fixing audio for car X,” then we need to add ‘X’ and ‘car’ as synonyms.
71
71
72
72
The transformer-based model already takes care of most of the common synonym cases, for example: `Purchase – Buy`, `Sell - Auction`, `Price – Value`. For another example, consider the following question answer pair: Q: “What is the price of Microsoft Stock?” A: “$200”.
73
73
74
-
If we receive user queries like “Microsoft stock value”,” Microsoft share value”, “Microsoft stock worth”, “Microsoft share worth”, “stock value”, etc., you should be able to get the correct answer even though these queries have words like "share", "value", and "worth", which aren’t originally present in the project.
74
+
If we receive user queries like “Microsoft stock value,”” Microsoft share value,” “Microsoft stock worth,” “Microsoft share worth,” “stock value,” etc., you should be able to get the correct answer even though these queries have words like "share", "value", and "worth", which aren’t originally present in the project.
Copy file name to clipboardExpand all lines: articles/ai-services/language-service/question-answering/how-to/network-isolation.md
+7-7Lines changed: 7 additions & 7 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -5,13 +5,13 @@ ms.service: azure-ai-language
5
5
ms.topic: how-to
6
6
author: jboback
7
7
ms.author: jboback
8
-
ms.date: 11/21/2024
8
+
ms.date: 06/06/2025
9
9
ms.custom: language-service-question-answering
10
10
---
11
11
12
12
# Network isolation and private endpoints
13
13
14
-
The steps below describe how to restrict public access to custom question answering resources as well as how to enable Azure Private Link. Protect an AI Foundry resource from public access by [configuring the virtual network](../../../cognitive-services-virtual-networks.md?tabs=portal).
14
+
The following steps describe how to restrict public access to custom question answering resources and how to enable Azure Private Link. Protect an AI Foundry resource from public access by [configuring the virtual network](../../../cognitive-services-virtual-networks.md?tabs=portal).
15
15
16
16
## Private Endpoints
17
17
@@ -21,14 +21,14 @@ Private endpoints are provided by [Azure Private Link](/azure/private-link/priva
21
21
22
22
## Steps to enable private endpoint
23
23
24
-
1. Assign *Contributer* role to language resource (Depending on the context this may appear as a Text Analytics resource) in the Azure Search Service instance. This operation requires *Owner* access to the subscription. Go to Identity tab in the service resource to get the identity.
24
+
1. Assign the *contributer* role to the resource in the Azure Search Service instance. This operation requires *Owner* access to the subscription. Go to Identity tab in the service resource to get the identity.
5. Go to *Networking* tab of language resource and under the *Allow access from*, select the *Selected Networks and private endpoints* option and select *save*.
41
+
5. Go to the **Networking** tab of the resource. Under the **Allow access from** section, select the **Selected Networks and private endpoints** option and select **save**.
@@ -56,7 +56,7 @@ This will establish a private endpoint connection between language resource and
56
56
57
57
Follow the steps below to restrict public access to custom question answering language resources. Protect an AI Foundry resource from public access by [configuring the virtual network](../../../cognitive-services-virtual-networks.md?tabs=portal).
58
58
59
-
After restricting access to an AI Foundry resource based on VNet, To browse projects on Language Studio from your on-premises network or your local browser.
59
+
After you restrict access to an AI Foundry resource based on virtual network, to browse projects on Language Studio from your on-premises network or your local browser.
60
60
- Grant access to [on-premises network](../../../cognitive-services-virtual-networks.md?tabs=portal#configure-access-from-on-premises-networks).
61
61
- Grant access to your [local browser/machine](../../../cognitive-services-virtual-networks.md?tabs=portal#managing-ip-network-rules).
62
62
- Add the **public IP address of the machine under the Firewall** section of the **Networking** tab. By default `portal.azure.com` shows the current browsing machine's public IP (select this entry) and then select **Save**.
0 commit comments