feat: midas cls sent tokenize only if needed (#101)

* feat: midas cls sent tokenize only if needed

* feat: take into account tokenized uttrs by bot

* fix: codestyle

annotators/midas_classification/Dockerfile

@@ -16,6 +16,7 @@ COPY . /src/
 WORKDIR /src
 
 RUN python -m deeppavlov install $CONFIG
+RUN python -m spacy download en_core_web_sm
 
 RUN sed -i "s|$SED_ARG|g" "$CONFIG"
 

annotators/midas_classification/requirements.txt

@@ -4,4 +4,4 @@ sentry-sdk==0.14.2
 requests==2.23.0
 gunicorn==19.9.0
 numpy==1.17.2
-nltk==3.2.5
+spacy==3.0.6

annotators/midas_classification/server.py

@@ -4,9 +4,9 @@
 
 import numpy as np
 import sentry_sdk
+import spacy
 from deeppavlov import build_model
 from flask import Flask, request, jsonify
-from nltk.tokenize import sent_tokenize
 
 sentry_sdk.init(os.getenv("SENTRY_DSN"))
 
@@ -18,6 +18,7 @@
 EMPTY_SIGN = ": EMPTY >"
 
 try:
+    spacy_nlp = spacy.load("en_core_web_sm")
     model = build_model("midas_conv_bert.json", download=True)
     m = model(["hi"])
 except Exception as e:
@@ -76,6 +77,14 @@ def recombine_responses(responses, dialog_ids, n_dialogs):
     return final_responses
 
 
+def spacy_sent_tokenize(sentence):
+    doc = spacy_nlp(sentence)
+    segments = []
+    for sent in doc.sents:
+        segments.append(sent.text)
+    return segments
+
+
 @app.route("/model", methods=["POST"])
 def respond():
     st_time = time.time()
@@ -84,17 +93,22 @@ def respond():
     inputs = []
     for i, dialog in enumerate(dialogs):
         if len(dialog["bot_utterances"]):
-            prev_bot_uttr_text = dialog["bot_utterances"][-1].get("text", "").lower()
-            tokenized_sentences = sent_tokenize(prev_bot_uttr_text)
+            tokenized_sentences = (
+                dialog["bot_utterances"][-1].get("annotations", {}).get("sentseg", {}).get("segments", [])
+            )
+            if len(tokenized_sentences) == 0:
+                prev_bot_uttr_text = dialog["bot_utterances"][-1].get("text", "").lower()
+                tokenized_sentences = spacy_sent_tokenize(prev_bot_uttr_text)
             context = tokenized_sentences[-1].lower() if len(tokenized_sentences) > 0 else ""
         else:
             context = ""
         if len(dialog["human_utterances"]):
-            curr_human_uttr_text = dialog["human_utterances"][-1].get("text", "").lower()
+            sentences = dialog["human_utterances"][-1].get("annotations", {}).get("sentseg", {}).get("segments", [])
+            if len(sentences) == 0:
+                sentences = spacy_sent_tokenize(dialog["human_utterances"][-1].get("text", "").lower())
         else:
-            curr_human_uttr_text = ""
+            sentences = []
 
-        sentences = sent_tokenize(curr_human_uttr_text)
         for sent in sentences:
             input_ = f"{context} {EMPTY_SIGN} {sent}"
             inputs.append(input_)
@@ -113,12 +127,19 @@ def batch_respond():
     st_time = time.time()
     bot_utterances = request.json["sentences"]
     human_utterances = request.json["last_human_utterances"]
-    bot_utterances_sentences = [sent_tokenize(utterance) for utterance in bot_utterances]
+    bot_utterances_sentences = [
+        spacy_sent_tokenize(utterance) if isinstance(utterance, str) else utterance for utterance in bot_utterances
+    ]
     dialog_ids = []
     inputs = []
     for i, bot_utterance_sents in enumerate(bot_utterances_sentences):
         if human_utterances[i]:
-            context = sent_tokenize(human_utterances[i])[-1].lower()
+            if isinstance(human_utterances[i], str):
+                context = spacy_sent_tokenize(human_utterances[i])[-1].lower()
+            elif isinstance(human_utterances[i], list):
+                context = human_utterances[i][-1].lower()
+            else:
+                context = ""
         else:
             context = ""
         for utterance in bot_utterance_sents: