empatico

Emotion detection microservice based on natural language inference

Table of Contents

Demo

Launch the model (dev mode, auto-reloading):

$ python3 -m venv .venv
$ source ./venv/bin/activate
(.venv) $ pip install -r requirements.txt
(.venv) $ pip install flit
(.venv) $ flit install --symlink
(.venv) $ uvicorn empatico:app --reload
INFO:     Will watch for changes in these directories: ['/home/caleb/Documents/repos/empatico']
INFO:     Uvicorn running on http://127.0.0.1:8000 (Press CTRL+C to quit)
INFO:     Started reloader process [303534] using watchgod

There are several classifications already set up. You just need to provide the text:

$ curl --request POST \
    --header "Content-Type: application/json" \
    --data '{"text": "I love when you speak to me rudely"}' \
    http://127.0.0.1:8000/emotions | jq
[
  {
    "label": "mixed",
    "score": 92.90633797645569
  },
  {
    "label": "anger",
    "score": 95.19177079200745
  },
  {
    "label": "sarcastic",
    "score": 97.97245264053345
  }
]

Wow, a label for sarcasm! Here's another one:

$ curl --request POST \
    --header "Content-Type: application/json" \
    --data '{"text": "Best insurance company in europe what a joke"}' \
    http://127.0.0.1:8000/emotions | jq
[
  {
    "label": "negative",
    "score": 99.07029867172241
  },
  {
    "label": "mixed",
    "score": 92.89084672927856
  },
  {
    "label": "anger",
    "score": 99.08776879310608
  },
  {
    "label": "sadness",
    "score": 92.85369515419006
  },
  {
    "label": "bitter",
    "score": 99.18588995933533
  },
  {
    "label": "sarcastic",
    "score": 99.7484564781189
  },
  {
    "label": "fear",
    "score": 89.00700807571411
  },
  {
    "label": "disgust",
    "score": 99.62377548217773
  },
  {
    "label": "surprise",
    "score": 99.7963547706604
  }
]

Again sarcasm is detected (99.7% probability), with a few other comorbid emotions. Another test of the sarcasm detector:

$ curl --request POST \
    --header "Content-Type: application/json" \
    --data '{"text": "yes go ahead and sue me, I''m sure that will work out fine for you"}' \
    http://127.0.0.1:8000/emotions | jq
[
  {
    "label": "sarcastic",
    "score": 96.60570025444031
  },
  {
    "label": "helpful",
    "score": 87.13959455490112
  },
  {
    "label": "trust",
    "score": 89.57158923149109
  }
]

It isn't always negative:

$ curl --request POST \
    --header "Content-Type: application/json" \
    --data '{"text": "Staff were wonderful and made the trip that much more pleasant. Thank you!"}' \
    http://127.0.0.1:8000/emotions | jq
[
  {
    "label": "positive",
    "score": 94.82141733169556
  },
  {
    "label": "satisfied",
    "score": 95.2349305152893
  },
  {
    "label": "helpful",
    "score": 95.06783485412598
  },
  {
    "label": "joy",
    "score": 95.37110924720764
  }
]

By default, a rich array of emotional labels is provided:

$ curl --request POST \
    --header "Content-Type: application/json" \
    --data '{"text": "The only way you could have done any worse is lose my package completely."}' \
    http://127.0.0.1:8000/emotions | jq
[
  {
    "label": "negative",
    "score": 98.56123924255371
  },
  {
    "label": "anger",
    "score": 97.950679063797
  },
  {
    "label": "sadness",
    "score": 89.63329792022705
  },
  {
    "label": "disappointment",
    "score": 87.91854977607727
  },
  {
    "label": "bitter",
    "score": 90.94756841659546
  },
  {
    "label": "fear",
    "score": 90.21917581558228
  },
  {
    "label": "disgust",
    "score": 90.52256941795349
  },
  {
    "label": "surprise",
    "score": 83.91632437705994
  }
]

Default labels (and reporting threshold)

You can also control the cutoff threshold for reporting. Here we set it to zero, which disables the filter. This shows all the default hypotheses (which are expressed as labels):

$ curl --request POST \
    --header "Content-Type: application/json" \
    --data '{"text": "The kids were so looking forward to the trip but the rain washed away all our plans.", \
        "report_threshold": 0.0}' \
    http://127.0.0.1:8000/emotions | jq
[
  {
    "label": "positive",
    "score": 0.7067840080708265
  },
  {
    "label": "negative",
    "score": 98.16489219665527
  },
  {
    "label": "mixed",
    "score": 98.80892634391785
  },
  {
    "label": "satisfied",
    "score": 0.7183659821748734
  },
  {
    "label": "neutral1",
    "score": 0.5342578981071711
  },
  {
    "label": "neutral2",
    "score": 0.034320083796046674
  },
  {
    "label": "neutral3",
    "score": 4.02584969997406
  },
  {
    "label": "factual",
    "score": 6.705068796873093
  },
  {
    "label": "anger",
    "score": 33.71554911136627
  },
  {
    "label": "sadness",
    "score": 98.14655780792236
  },
  {
    "label": "disappointment",
    "score": 99.36606287956238
  },
  {
    "label": "bitter",
    "score": 61.47879958152771
  },
  {
    "label": "sarcastic",
    "score": 31.848391890525818
  },
  {
    "label": "helpful",
    "score": 71.62957191467285
  },
  {
    "label": "fear",
    "score": 17.39620268344879
  },
  {
    "label": "disgust",
    "score": 4.311040416359901
  },
  {
    "label": "surprise",
    "score": 90.2463436126709
  },
  {
    "label": "hope",
    "score": 0.21583051420748234
  },
  {
    "label": "trust",
    "score": 37.46950924396515
  },
  {
    "label": "joy",
    "score": 0.47363536432385445
  }
]

It is really interesting that while the "overall" positive score is 0.7% and the "overall" negative score is 98%, we do still see a "mixed" score of 98.8%. This can be interpreted as "while there were both positive and negative sentiments expressed, the negative outweighs the positive".

Customizable hypotheses

You can also provide your own hypotheses, which means you can generalise this to many difference kinds of classifications:

$ curl --request POST --header "Content-Type: application/json" \
    --data '{"text": "The democrats are ruining this country", \
    "report_threshold": 0.0, \
    "hypotheses": {"politics": "this text is about politics", \
    "sport": "this text is about sport"}}' http://127.0.0.1:8000/emotions | jq
[
  {
    "label": "politics",
    "score": 97.42230772972107
  },
  {
    "label": "sport",
    "score": 0.16288807382807136
  }
]
~
$ curl --request POST --header "Content-Type: application/json" \
    --data '{"text": "The tour de france was exhilarating to watch", \
    "report_threshold": 0.0, \
    "hypotheses": {"politics": "this text is about politics", \
    "sport": "this text is about sport"}}' http://127.0.0.1:8000/emotions | jq
[
  {
    "label": "politics",
    "score": 0.4232536070048809
  },
  {
    "label": "sport",
    "score": 97.9870855808258
  }
]

Background

The underlying technique for using natural language inference for classification was described by Joe Davison here:

https://joeddav.github.io/blog/2020/05/29/ZSL.html

The underlying idea is that, instead of using a model specifically trained for binary, ternary, or n-ary classification, we might instead try to use a model designed for inference and test whether a particular hypothesis is supported by a piece of text. This dramatically expands the scope of possible labels to apply to a given piece of text, and you don't need the model to have been specifally trained for those labels either.

The model being used in the code is facebook/bart-large-mnli.

Running the server in dev

$ uvicorn empatico:app --reload