REST Resource: projects.conversationModels.evaluations

Resource: ConversationModelEvaluation

Represents evaluation result of a conversation model.

JSON representation 
{
  "name": string,
  "displayName": string,
  "evaluationConfig": {
    object (EvaluationConfig)
  },
  "createTime": string,
  "rawHumanEvalTemplateCsv": string,

  // Union field metrics can be only one of the following:
  "smartReplyMetrics": {
    object (SmartReplyMetrics)
  }
  // End of list of possible types for union field metrics.
}
Fields
name

string

The resource name of the evaluation. Format: projects/<Project ID>/conversationModels/<Conversation Model ID>/evaluations/<Evaluation ID>
displayName

string

Optional. The display name of the model evaluation. At most 64 bytes long.
evaluationConfig

object (EvaluationConfig)

Optional. The configuration of the evaluation task.
createTime

string (Timestamp format)

Output only. Creation time of this model.

A timestamp in RFC3339 UTC "Zulu" format, with nanosecond resolution and up to nine fractional digits. Examples: "2014-10-02T15:01:23Z" and "2014-10-02T15:01:23.045123456Z".
rawHumanEvalTemplateCsv

string

Output only. Human eval template in csv format. It tooks real-world conversations provided through input dataset, generates example suggestions for customer to verify quality of the model. For Smart Reply, the generated csv file contains columns of Context, (Suggestions,Q1,Q2)*3, Actual reply. Context contains at most 10 latest messages in the conversation prior to the current suggestion. Q1: "Would you send it as the next message of agent?" Evaluated based on whether the suggest is appropriate to be sent by agent in current context. Q2: "Does the suggestion move the conversation closer to resolution?" Evaluated based on whether the suggestion provide solutions, or answers customer's question or collect information from customer to resolve the customer's issue. Actual reply column contains the actual agent reply sent in the context.
Union field metrics. Metrics details. metrics can be only one of the following:
smartReplyMetrics

object (SmartReplyMetrics)

Output only. Only available when model is for smart reply.

EvaluationConfig

The configuration for model evaluation.

JSON representation 
{
  "datasets": [
    {
      object (InputDataset)
    }
  ],

  // Union field model_specific_config can be only one of the following:
  "smartReplyConfig": {
    object (SmartReplyConfig)
  },
  "smartComposeConfig": {
    object (SmartComposeConfig)
  }
  // End of list of possible types for union field model_specific_config.
}
Fields
datasets[]

object (InputDataset)

Required. Datasets used for evaluation.
Union field model_specific_config. Specific configurations for different models in order to do evaluation. model_specific_config can be only one of the following:
smartReplyConfig

object (SmartReplyConfig)

Configuration for smart reply model evalution.
smartComposeConfig

object (SmartComposeConfig)

Configuration for smart compose model evalution.

SmartReplyConfig

Smart reply specific configuration for evaluation job.

JSON representation 
{
  "allowlistDocument": string,
  "maxResultCount": integer
}
Fields
allowlistDocument

string

The allowlist document resource name. Format: projects/<Project ID>/knowledgeBases/<Knowledge Base ID>/documents/<Document ID>. Only used for smart reply model.
maxResultCount

integer

Required. The model to be evaluated can return multiple results with confidence score on each query. These results will be sorted by the descending order of the scores and we only keep the first maxResultCount results as the final results to evaluate.

SmartComposeConfig

Smart compose specific configuration for evaluation job.

JSON representation 
{
  "allowlistDocument": string,
  "maxResultCount": integer
}
Fields
allowlistDocument

string

The allowlist document resource name. Format: projects/<Project ID>/knowledgeBases/<Knowledge Base ID>/documents/<Document ID>. Only used for smart compose model.
maxResultCount

integer

Required. The model to be evaluated can return multiple results with confidence score on each query. These results will be sorted by the descending order of the scores and we only keep the first maxResultCount results as the final results to evaluate.

SmartReplyMetrics

The evaluation metrics for smart reply model.

JSON representation 
{
  "allowlistCoverage": number,
  "topNMetrics": [
    {
      object (TopNMetrics)
    }
  ],
  "conversationCount": string
}
Fields
allowlistCoverage

number

Percentage of target participant messages in the evaluation dataset for which similar messages have appeared at least once in the allowlist. Should be [0, 1].
topNMetrics[]

object (TopNMetrics)

Metrics of top n smart replies, sorted by [TopNMetric.n][].
conversationCount

string (int64 format)

Total number of conversations used to generate this metric.

TopNMetrics

Evaluation metrics when retrieving n smart replies with the model.

JSON representation 
{
  "n": integer,
  "recall": number
}
Fields
n

integer

Number of retrieved smart replies. For example, when n is 3, this evaluation contains metrics for when Dialogflow retrieves 3 smart replies with the model.
recall

number

Defined as number of queries whose top n smart replies have at least one similar (token match similarity above the defined threshold) reply as the real reply divided by number of queries with at least one smart reply. Value ranges from 0.0 to 1.0 inclusive.

Methods

get

 Gets an evaluation of conversation model.

list

 Lists evaluations of a conversation model.