Sentiment classification

Last updated: Oct 09, 2024

The Watson Natural Language Processing Sentiment classification models classify the sentiment of the input text.

Supported languages

Sentiment classification is available for the following languages. For a list of the language codes and the corresponding language, see Language codes.

ar, cs, da, de, en, es, fi, fr, he, hi, it, ja, ko, nb, nl, nn, pl, pt, ro, ru, sk, sv, tr, zh-cn

Sentiment

The sentiment of text can be positive, negative or neutral.

The sentiment model computes the sentiment for each sentence in the input document. The aggregated sentiment for the entire document is also calculated using the sentiment transformer workflow in Runtime 23.1. If you are using the sentiment models in Runtime 22.x the overall document sentiment can be computed by the helper method called predict_document_sentiment.

The classifications returned contain a probability. The sentiment score varies from -1 to 1. A score greater than 0 denotes a positive sentiment, a score less than 0 a negative sentiment, and a score of 0 a neutral sentiment.

Sentence sentiment workflows in runtime 23.1

Workflow names

sentiment-aggregated_transformer-workflow_multilingual_slate.153m.distilled
sentiment-aggregated_transformer-workflow_multilingual_slate.153m.distilled-cpu

The sentiment-aggregated_transformer-workflow_multilingual_slate.153m.distilled workflow can be used on both CPUs and GPUs.

The sentiment-aggregated_transformer-workflow_multilingual_slate.153m.distilled-cpu workflow is optimized for CPU-based runtimes.

Code sample using the sentiment-aggregated_transformer-workflow_multilingual_slate.153m.distilled workflow

# Load the Sentiment workflow
sentiment_model = watson_nlp.load('sentiment-aggregated_transformer-workflow_multilingual_slate.153m.distilled-cpu')

# Run the sentiment model on the result of the syntax results
sentiment_result = sentiment_model.run('The rooms are nice. But the beds are not very comfortable.')

# Print the sentence sentiment results
print(sentiment_result)

Output of the code sample

{
  "document_sentiment": {
    "score": -0.339735,
    "label": "SENT_NEGATIVE",
    "mixed": true,
    "sentiment_mentions": [
      {
        "span": {
          "begin": 0,
          "end": 19,
          "text": "The rooms are nice."
        },
        "sentimentprob": {
          "positive": 0.9720447063446045,
          "neutral": 0.011838269419968128,
          "negative": 0.016117043793201447
        }
      },
      {
        "span": {
          "begin": 20,
          "end": 58,
          "text": "But the beds are not very comfortable."
        },
        "sentimentprob": {
          "positive": 0.0011594508541747928,
          "neutral": 0.006315878126770258,
          "negative": 0.9925248026847839
        }
      }
    ]
  },
  "targeted_sentiments": {
    "targeted_sentiments": {},
    "producer_id": {
      "name": "Aggregated Sentiment Workflow",
      "version": "0.0.1"
    }
  },
  "producer_id": {
    "name": "Aggregated Sentiment Workflow",
    "version": "0.0.1"
  }
}

Sentence sentiment blocks in 22.x runtimes

Block name

sentiment_sentence-bert_multi_stock

Dependencies on other blocks

The following block must run before you can run the Sentence sentiment block:

syntax_izumo_<language>_stock

Code sample using the sentiment_sentence-bert_multi_stock block

import watson_nlp
from watson_nlp.toolkit.sentiment_analysis_utils import predict_document_sentiment
# Load Syntax and a Sentiment model for English
syntax_model = watson_nlp.load('syntax_izumo_en_stock')
sentiment_model = watson_nlp.load('sentiment_sentence-bert_multi_stock')

# Run the syntax model on the input text
syntax_result = syntax_model.run('The rooms are nice. But the beds are not very comfortable.')

# Run the sentiment model on the result of the syntax results
sentiment_result = sentiment_model.run_batch(syntax_result.get_sentence_texts(), syntax_result.sentences)

# Print the sentence sentiment results
print(sentiment_result)

# Get the aggregated document sentiment
document_sentiment = predict_document_sentiment(sentiment_result, sentiment_model.class_idxs)
print(document_sentiment)

Output of the code sample:

[{
  "score": 0.9540348989256836,
  "label": "SENT_POSITIVE",
  "sentiment_mention": {
    "span": {
      "begin": 0,
      "end": 19,
      "text": "The rooms are nice."
    },
    "sentimentprob": {
      "positive": 0.919123649597168,
      "neutral": 0.05862388014793396,
      "negative": 0.022252488881349564
    }
  },
  "producer_id": {
    "name": "Sentence Sentiment Bert Processing",
    "version": "0.1.0"
  }
}, {
  "score": -0.9772116371114815,
  "label": "SENT_NEGATIVE",
  "sentiment_mention": {
    "span": {
      "begin": 20,
      "end": 58,
      "text": "But the beds are not very comfortable."
    },
    "sentimentprob": {
      "positive": 0.015949789434671402,
      "neutral": 0.025898978114128113,
      "negative": 0.9581512808799744
    }
  },
  "producer_id": {
    "name": "Sentence Sentiment Bert Processing",
    "version": "0.1.0"
  }
}]
{
  "score": -0.335185,
  "label": "SENT_NEGATIVE",
  "mixed": true,
  "sentiment_mentions": [
    {
      "span": {
        "begin": 0,
        "end": 19,
        "text": "The rooms are nice."
      },
      "sentimentprob": {
        "positive": 0.919123649597168,
        "neutral": 0.05862388014793396,
        "negative": 0.022252488881349564
      }
    },
    {
      "span": {
        "begin": 20,
        "end": 58,
        "text": "But the beds are not very comfortable."
      },
      "sentimentprob": {
        "positive": 0.015949789434671402,
        "neutral": 0.025898978114128113,
        "negative": 0.9581512808799744
      }
    }
  ]
}

Targets sentiment extraction

Targets sentiment extraction extracts sentiments expressed in text and identifies the targets of those sentiments.

It can handle multiple targets with different sentiment in one sentence as opposed to the sentiment block described above.

For example, given the input sentence The served food was delicious, yet the service was slow., the Targets sentiment block identifies that there is a positive sentiment expressed in the target "food", and a negative sentiment expressed in "service".

The model has been fine-tuned on English data only. Although you can use the model on the other languages listed under Supported languages, the results might vary.

Targets sentiment workflows in Runtime 23.1

Workflow names

targets-sentiment_transformer-workflow_multilingual_slate.153m.distilled
targets-sentiment_transformer-workflow_multilingual_slate.153m.distilled-cpu

The targets-sentiment_transformer-workflow_multilingual_slate.153m.distilled workflow can be used on both CPUs and GPUs.

The targets-sentiment_transformer-workflow_multilingual_slate.153m.distilled-cpu workflow is optimized for CPU-based runtimes.

Code sample for the targets-sentiment_transformer-workflow_multilingual_slate.153m.distilled workflow

import watson_nlp
# Load Targets Sentiment model for English
targets_sentiment_model = watson_nlp.load('targets-sentiment_transformer-workflow_multilingual_slate.153m.distilled')
# Run the targets sentiment model on the input text
targets_sentiments = targets_sentiment_model.run('The rooms are nice, but the bed was not very comfortable.')
# Print the targets with the associated sentiment
print(targets_sentiments)

Output of the code sample:

{
  "targeted_sentiments": {
    "rooms": {
      "score": 0.990798830986023,
      "label": "SENT_POSITIVE",
      "mixed": false,
      "sentiment_mentions": [
        {
          "span": {
            "begin": 4,
            "end": 9,
            "text": "rooms"
          },
          "sentimentprob": {
            "positive": 0.990798830986023,
            "neutral": 0.0,
            "negative": 0.00920116901397705
          }
        }
      ]
    },
    "bed": {
      "score": -0.9920912981033325,
      "label": "SENT_NEGATIVE",
      "mixed": false,
      "sentiment_mentions": [
        {
          "span": {
            "begin": 28,
            "end": 31,
            "text": "bed"
          },
          "sentimentprob": {
            "positive": 0.00790870189666748,
            "neutral": 0.0,
            "negative": 0.9920912981033325
          }
        }
      ]
    }
  },
  "producer_id": {
    "name": "Transformer-based Targets Sentiment Extraction Workflow",
    "version": "0.0.1"
  }
}

Targets sentiment blocks in 22.x runtimes

Block name targets-sentiment_sequence-bert_multi_stock

Dependencies on other blocks

The following block must run before you can run the Targets sentiment extraction block:

syntax_izumo_<language>_stock

Code sample using the sentiment-targeted_bert_multi_stock block

import watson_nlp

# Load Syntax and the Targets Sentiment model for English
syntax_model = watson_nlp.load('syntax_izumo_en_stock')
targets_sentiment_model = watson_nlp.load('targets-sentiment_sequence-bert_multi_stock')

# Run the syntax model on the input text
syntax_result = syntax_model.run('The rooms are nice, but the bed was not very comfortable.')

# Run the targets sentiment model on the syntax results
targets_sentiments = targets_sentiment_model.run(syntax_result)

# Print the targets with the associated sentiment
print(targets_sentiments)

Output of the code sample:

{
  "targeted_sentiments": {
    "rooms": {
      "score": 0.9989274144172668,
      "label": "SENT_POSITIVE",
      "mixed": false,
      "sentiment_mentions": [
        {
          "span": {
            "begin": 4,
            "end": 9,
            "text": "rooms"
          },
          "sentimentprob": {
            "positive": 0.9989274144172668,
            "neutral": 0.0,
            "negative": 0.0010725855827331543
          }
        }
      ]
    },
    "bed": {
      "score": -0.9977545142173767,
      "label": "SENT_NEGATIVE",
      "mixed": false,
      "sentiment_mentions": [
        {
          "span": {
            "begin": 28,
            "end": 31,
            "text": "bed"
          },
          "sentimentprob": {
            "positive": 0.002245485782623291,
            "neutral": 0.0,
            "negative": 0.9977545142173767
          }
        }
      ]
    }
  },
  "producer_id": {
    "name": "BERT TSA",
    "version": "0.0.1"
  }
}

Parent topic: Watson Natural Language Processing task catalog