Sentiment task-specific AI functions¶

Sentiment quality¶

ENTITY_SENTIMENT delivers industry-leading aspect based and overall sentiment accuracy. The following table provides information on how popular models perform on the Aspect Based Sentiment (ABSA-mix) and Overall Sentiment benchmarks, including the models offered by Snowflake as well as other popular models.

ENTITY_SENTIMENT¶

Cortex AI ENTITY_SENTIMENT captures a spectrum of customer opinions beyond overall positive, negative, and neutral buckets. For aspect-based sentiment analysis, specify the content (such as a customer comment or a review) and the entities (aspects or categories) for which you want to analyze sentiment. ENTITY_SENTIMENT returns sentiment scores for each entity as well as an overall sentiment score.

The following example uses ENTITY_SENTIMENT to get the sentiment classification of a product review.

SELECT SNOWFLAKE.CORTEX.ENTITY_SENTIMENT('A tourist\'s delight, in low urban light,
  Recommended gem, a pizza night sight.  Swift arrival, a pleasure so right,
  Yet, pockets felt lighter, a slight pricey bite. 💰🍕🚀',
  ['Cost', 'Quality', 'Waiting Time']);

Response:

{
  "categories": [
    {
      "name": "overall",
      "sentiment": "mixed"
    },
    {
      "name": "Cost",
      "sentiment": "negative"
    },
    {
      "name": "Quality",
      "sentiment": "positive"
    },
    {
      "name": "Waiting Time",
      "sentiment": "positive"
    }
  ]
}

The following example “passes through” entities unrelated to the content of the product review, Professionalism and Brand, and classifies them as “unknown”.

SELECT SNOWFLAKE.CORTEX.ENTITY_SENTIMENT('A tourist\'s delight, in low urban light,
  Recommended gem, a pizza night sight. Swift arrival, a pleasure so right,
  Yet, pockets felt lighter, a slight pricey bite. 💰🍕🚀',
  ['Cost', 'Professionalism' ,'Brand']);

Response:

{
  "categories": [
    {
      "name": "overall",
      "sentiment": "mixed"
    },
    {
      "name": "Brand",
      "sentiment": "unknown"
    },
    {
      "name": "Cost",
      "sentiment": "negative"
    },
    {
      "name": "Professionalism",
      "sentiment": "unknown"
    }
  ]
}

If you want only the overall sentiment, specify the content without analysis entities.

SELECT SNOWFLAKE.CORTEX.ENTITY_SENTIMENT('I went to the store, bought the leggings and exact same as shorts...
  they are expensive but i heard such great things. After wearing them twice i noticed a string popping out already.
  And aince i believed that they were this amazing luxury brand i didnt keep the receipt 😭 ');

Response:

{
  "categories": [
    {
      "name": "overall",
      "sentiment": "mixed"
    }
  ]
}

SENTIMENT¶

The Cortex AI SENTIMENT function returns overal sentiment as a score between -1.0 to 1.0t. The closer to 1, the the more likely that the text has positive sentiment; the the closer to -1 the more likely the text has negative sentiment. The chart below provides guidance on how to interpret the sentiment scores:

The following example uses SENTIMENT to get the sentiment classification of a food service review, which we can infer as positive given the probability score of 0.54.

SELECT SNOWFLAKE.CORTEX.SENTIMENT('A tourist\'s delight, in low urban light,
  Recommended gem, a pizza night sight. Swift arrival, a pleasure so right,
  Yet, pockets felt lighter, a slight pricey bite. 💰🍕🚀');

Response:

0.5424458

Model restrictions¶

All models available in Snowflake Cortex AI have limitations on the total number of input and output tokens, which is referred to as the model’s context window.

The Context Window for ENTITY_SENTIMENT is set such that the model can sustain a high level of accuracy. For more context, ENTITY_SENTIMENT was trained and optimized for text inputs of 2,048 tokens (roughly 1,600 words), and you can pass a maximum of 10 entity categories, each no greater than 30 characters.

Regional availability¶

ENTITY_SENTIMENT and SENTIMENT are available in the following regions: