Aspect based Sentiment Analysis for restaurant reviews

Description

Automatically detect positive, negative and neutral aspects about restaurants from user reviews. Instead of labelling the entire review as negative or positive, this model helps identify which exact phrases relate to sentiment identified in the review.

Predicted Entities

NEG, POS

Live Demo Open in Colab Download Copy S3 URI

How to use

Use as part of an nlp pipeline with the following stages: DocumentAssembler, SentenceDetector, Tokenizer, WordEmbeddingsModel, NerDLModel. Add the NerConverter to the end of the pipeline to convert entity tokens into full entity chunks.

...
word_embeddings = WordEmbeddingsModel.pretrained("glove_6B_300", "xx")\
.setInputCols(["document", "token"])\
.setOutputCol("embeddings")
ner_model = NerDLModel.pretrained("ner_aspect_based_sentiment")\
.setInputCols(["document", "token", "embeddings"])\
.setOutputCol("ner")
...
nlp_pipeline = Pipeline(stages=[document_assembler, sentence_detector, tokenizer, word_embeddings, ner_model, ner_converter])
model = nlp_pipeline.fit(spark.createDataFrame([[""]]).toDF("text"))
results = model.transform(spark.createDataFrame([["Came for lunch my sister. We loved our Thai-style main which amazing with lots of flavours very impressive for vegetarian. But the service was below average and the chips were too terrible to finish."]]).toDF("text"))
...
val word_embeddings = WordEmbeddingsModel.pretrained("glove_6B_300", "xx")
.setInputCols(Array("document", "token"))
.setOutputCol("embeddings")
val ner_model = NerDLModel.pretrained("ner_aspect_based_sentiment")
.setInputCols(Array("document", "token", "embeddings"))
.setOutputCol("ner")
...
val pipeline = new Pipeline().setStages(Array(document_assembler, sentence_detector, tokenizer, word_embeddings, ner_model, ner_converter))
val data = Seq("Came for lunch my sister. We loved our Thai-style main which amazing with lots of flavours very impressive for vegetarian. But the service was below average and the chips were too terrible to finish.").toDF("text")
val result = pipeline.fit(data).transform(data)
import nlu
text = ["""Came for lunch my sister. We loved our Thai-style main which amazing with lots of flavours very impressive for vegetarian. But the service was below average and the chips were too terrible to finish."""]

ner_df = nlu.load('en.ner.aspect_sentiment').predict(text, output_level='token')
list(zip(ner_df["entities"].values[0], ner_df["entities_confidence"].values[0])

Results

+----------------------------------------------------------------------------------------------------+-------------------+-----------+
| sentence                                                                                           | aspect            | sentiment |
+----------------------------------------------------------------------------------------------------+-------------------+-----------+
| We loved our Thai-style main which amazing with lots of flavours very impressive for vegetarian.   | Thai-style main   | positive  |
| We loved our Thai-style main which amazing with lots of flavours very impressive for vegetarian.   | lots of flavours  | positive  |
| But the service was below average and the chips were too terrible to finish.                       | service           | negative  |
| But the service was below average and the chips were too terrible to finish.                       | chips             | negative  |
+----------------------------------------------------------------------------------------------------+-------------------+-----------+

Model Information

Model Name: ner_aspect_based_sentiment
Type: ner
Compatibility: Spark NLP 2.6.2+
License: Open Source
Edition: Official
Input Labels: [token, embeddings]
Output Labels: [absa]
Language: en
Dependencies: glove_6B_300