Emotion analysis

English only, needs to translate first

We use the IsaacZhy/roberta-large-goemotions model for emotion identification as it supports 28 different emotion labels.

tokenizer = transformers.RobertaTokenizerFast.from_pretrained("IsaacZhy/roberta-large-goemotions")
model = transformers.RobertaForSequenceClassification.from_pretrained("IsaacZhy/roberta-large-goemotions")

id2label = {
    "0": "admiration",
    "1": "amusement",
    "2": "anger",
    "3": "annoyance",
    "4": "approval",
    "5": "caring",
    "6": "confusion",
    "7": "curiosity",
    "8": "desire",
    "9": "disappointment",
    "10": "disapproval",
    "11": "disgust",
    "12": "embarrassment",
    "13": "excitement",
    "14": "fear",
    "15": "gratitude",
    "16": "grief",
    "17": "joy",
    "18": "love",
    "19": "nervousness",
    "20": "optimism",
    "21": "pride",
    "22": "realization",
    "23": "relief",
    "24": "remorse",
    "25": "sadness",
    "26": "surprise",
    "27": "neutral"
  }
def labelProbabilities(probabilities):
    return {id2label[str(p)]: float(probabilities[0][p])  for p in range(len(probabilities[0]))}

def evaluate(sentence):
    tokens = tokenizer.encode(sentence, return_tensors="pt")
    with torch.no_grad():
        result = model(tokens)
        probabilities = torch.nn.functional.softmax(result.logits, dim=-1)
        return labelProbabilities(probabilities)

def most_likely(sentence):
    probabilities = evaluate(sentence)
    return max(probabilities, key=probabilities.get)

It works well on tests but only works for English data. This is the case for all of the emotion classifiers I could find. Luckily, google automatically translate play store reviews to the language you need.

Some examples:

>>> most_likely("During covid, when all restaurants where closed OpenTable expired all my points. For I am deleting this App! Bad customer relationship management!")
'disappointment'

>>> most_likely("Try 3 crédit cards for japanese restaurant at Tokyo... Never opentable want our credit card... Several restaurant at Tokyo use opentable, how is possible ? Gaijin can't use money ?? Bad night with family")
'curiosity'

>>> most_likely("Convenient, great experience")
'admiration'

>>> most_likely("Had the best time of my life")
'joy'

>>> most_likely("Obviously not the best way to pick a restaurant")
'disapproval'

While it can struggle on long comments, it can do pretty well on short form messages, with an understanding of the meaning of words and of negations.