Finishing up NLU module

.gitignore

@@ -0,0 +1,4 @@
+__pycache__
+slot-model*
+frame-model*
+nlu_cache

README.md

@@ -34,4 +34,13 @@ Agent powinien wykazywać elastyczność, adaptując się do potrzeb klienta, np
 | inform                      | poinformowanie użytkownika, o przyjętej wartości slotu  |
 | offer                       | rekomendacja (restauracji)                              |
 | request                     | pytanie użytkownika o wartość slotu                     |
-| select                      | prośba o dokonanie wyboru spośród przedstawionych opcji |
+| select                      | prośba o dokonanie wyboru spośród przedstawionych opcji |
+
+# Obsługa projektu
+
+- Python 3.10.12
+- Instalacja dependencies `pip3 install -r requirements.txt`
+- Centralna część systemu - uruchamiamy `python3 src/main.py`
+- NLU:
+    - uczenie modeli od zera `python3 nlu_train.py`
+    - Ewaluacja `python3 evaluate.py`

evaluate.py

@@ -2,11 +2,10 @@ import re
 import os
 import pandas as pd
 import numpy as np
-from nlu_utils import predict_multiple
 from flair.models import SequenceTagger
 from conllu import parse_incr
 from flair.data import Corpus
-from nlu_utils import conllu2flair, nolabel2o
+from src.utils.nlu_utils import conllu2flair, nolabel2o, predict_multiple
 
 # Frame model evaluation
 frame_model = SequenceTagger.load('frame-model-prod/best-model.pt')

nlu_train.py

@@ -6,7 +6,7 @@ from flair.embeddings import CharacterEmbeddings
 from flair.embeddings import FlairEmbeddings
 from flair.models import SequenceTagger
 from flair.trainers import ModelTrainer
-from nlu_utils import conllu2flair, nolabel2o
+from src.utils.nlu_utils import conllu2flair, nolabel2o
 
 import torch
 if torch.cuda.is_available():
@@ -38,4 +38,4 @@ def train_model(label_type, field_parsers = {}):
 
 if __name__ == '__main__':
     train_model("frame")
-    train_model('slot', field_parsers={'slot': nolabel2o})
+    train_model('slot', field_parsers={'slot': nolabel2o})

requirements.txt

@@ -0,0 +1,5 @@
+flair==0.13.1
+conllu==4.5.3
+pandas==1.5.3
+numpy==1.26.4
+torch==2.3.0

src/main.py

@@ -1,26 +1,38 @@
-from model.frame import Frame
 from service.dialog_state_monitor import DialogStateMonitor
 from service.dialog_policy import DialogPolicy
 from service.natural_languag_understanding import NaturalLanguageUnderstanding
 from service.natural_language_generation import NaturalLanguageGeneration
 
-print("Natural language understanding, example:")
-naturalLanguageUnderstanding = NaturalLanguageUnderstanding()
-print(naturalLanguageUnderstanding.convert_text_to_frame("Cześć, jak masz na imię?"))
+# initialize classes
+nlu = NaturalLanguageUnderstanding() # NLU
+monitor = DialogStateMonitor() # DSM
+dialog_policy = DialogPolicy() # DP
+language_generation = NaturalLanguageGeneration() # NLG
+
+# Main loop
+user_input = input("Możesz zacząć pisać.\n")
+while True:
+    # NLU
+    frame = nlu.process_input(user_input)
+    print(frame)
+
+    # DSM
+    monitor.append(frame)
+
+    # DP
+    print(dialog_policy.next_dialogue_act(monitor.get_all()).act)
+
+    # NLG
+    response = language_generation.respond_to_name_query("Jak masz na imię?")
+    print(response)
+
+    if frame.act == "bye":
+        break
+    
+    user_input = input(">\n")
+
+
+
 
-# Example
-print("Dialog state monitor, examples:")
-monitor = DialogStateMonitor()
-monitor.append(Frame('system', 'hello', []))
-monitor.append(Frame('user', 'some_text', []))
-print(monitor.get_all()[0].act)
-print(monitor.get_last().act)
 
-print("Dialog policy, next dialogue act:")
-dialog_policy = DialogPolicy(monitor.get_all())
-print(dialog_policy.next_dialogue_act().act)
 
-print("Natural Language Generation example:")
-agent = NaturalLanguageGeneration()
-response = agent.respond_to_name_query("Jak masz na imię?")
-print(response)

src/model/frame.py

@@ -1,7 +1,14 @@
-from model.slot import Slot
+from .slot import Slot
 
 class Frame:
     def __init__(self, source: str, act: str, slots: list[Slot]):
         self.source = source
         self.slots = slots
         self.act = act
+
+    def __str__(self):
+        msg = f"Act: {self.act}, Slots: ["
+        for slot in self.slots:
+            msg += f"({slot}), "
+        msg += "]"
+        return msg

src/model/slot.py

@@ -1,4 +1,7 @@
 class Slot:
     def __init__(self, name, value=None):
         self.name = name
-        self.value = value
+        self.value = value
+
+    def __str__(self) -> str:
+        return f"Name: {self.name}, Value: {self.value}"

src/nlu_example.py

@@ -1,30 +1,30 @@
-from flair.models import SequenceTagger
-from nlu_utils import predict_single, predict_multiple, predict_and_annotate
-
-# Exploratory tests
-frame_model = SequenceTagger.load('frame-model/best-model.pt')
-tests = [
-    'chciałbym zamówić pizzę',
-    'na godzinę 12',
-    'prosiłbym o pizzę z pieczarkami',
-    'to wszystko, jaka cena?',
-    'ile kosztuje pizza',
-    'do widzenia',
-    'tak',
-    'nie dziękuję',
-    'dodatkowy ser',
-    'pizzę barcelona bez cebuli',
-]
-
-# print("=== Exploratory tests - frame model ===")
-for test in tests:
-    print(f"Sentence: {test}")
-    print(f"Single prediction: {predict_single(frame_model, test.split(), 'frame')}")
-    print(f"Multiple predictions: {predict_multiple(frame_model, test.split(), 'frame')}")
-    print(f"Annotated sentence: {predict_and_annotate(frame_model, test.split(), 'frame')}")
-
-print("=== Exploratory tests - slot model ===")
-slot_model = SequenceTagger.load('slot-model/final-model.pt')
-for test in tests:
-    print(f"Sentence: {test}")
+from flair.models import SequenceTagger
+from utils.nlu_utils import predict_single, predict_multiple, predict_and_annotate
+
+# Exploratory tests
+frame_model = SequenceTagger.load('frame-model-prod/best-model.pt')
+tests = [
+    'chciałbym zamówić pizzę',
+    'na godzinę 12',
+    'prosiłbym o pizzę z pieczarkami',
+    'to wszystko, jaka cena?',
+    'ile kosztuje pizza',
+    'do widzenia',
+    'tak',
+    'nie dziękuję',
+    'dodatkowy ser',
+    'pizzę barcelona bez cebuli',
+]
+
+print("=== Exploratory tests - frame model ===")
+for test in tests:
+    print(f"Sentence: {test}")
+    print(f"Single prediction: {predict_single(frame_model, test.split(), 'frame')}")
+    print(f"Multiple predictions: {predict_multiple(frame_model, test.split(), 'frame')}")
+    print(f"Annotated sentence: {predict_and_annotate(frame_model, test.split(), 'frame')}")
+
+print("=== Exploratory tests - slot model ===")
+slot_model = SequenceTagger.load('slot-model-prod/best-model.pt')
+for test in tests:
+    print(f"Sentence: {test}")
     print(f"Prediction: {predict_and_annotate(slot_model, test.split(), 'slot')}")

src/service/dialog_policy.py

@@ -1,11 +1,8 @@
 from model.frame import Frame
 
 class DialogPolicy:
-    def __init__(self, frames: list[Frame]) -> None:
-        self.frames = frames
-
-    def next_dialogue_act(self) -> Frame:
-        if self.frames[-1].act == "welcomemsg":
+    def next_dialogue_act(self, frames: list[Frame]) -> Frame:
+        if frames[-1].act == "welcomemsg":
             return Frame("system", "welcomemsg", [])
         else:
             return Frame("system", "canthelp", [])

src/service/natural_languag_understanding.py

@@ -1,14 +1,92 @@
+from flair.models import SequenceTagger
+from utils.nlu_utils import predict_single, predict_and_annotate
+from model.frame import Frame, Slot
+
+"""
+ACTS:
+    inform/order
+    request/menu
+    inform/address
+    request/price
+    request/ingredients
+    request/sauce
+    inform/phone
+    inform/order-complete
+    request/time
+    request/size
+    welcomemsg
+    affirm
+    inform/delivery
+    inform/payment
+    request/delivery-price
+    bye
+    inform/time
+    request/drinks
+    inform/name
+    negate
+
+SLOTS:
+    food
+    pizza
+    size
+    address
+    quantity
+    ingredient
+    payment-method
+    delivery
+    drink
+    ingredient/neg
+    name
+    phone
+    sauce
+"""
+
 class NaturalLanguageUnderstanding:
+    def __init__(self):
+        print("\n========================================================")
+        print("Models are loading, it may take a moment, please wait...")
+        print("========================================================\n")
 
-    dictionary = {
-        "Cześć," : "welcomemsg()",
-        "imię?" : "request(name)"
-    }
+        self.frame_model = SequenceTagger.load('frame-model-prod/best-model.pt')
+        self.slot_model = SequenceTagger.load('slot-model-prod/best-model.pt')
 
-    def convert_text_to_frame(self, text: str):
-        frame = ""
-        text = text.split(" ")
-        for word in text:
-            if(word in self.dictionary):
-                    frame+=self.dictionary[word]+"&"
-        return frame[0:-1]
+        print("\n========================================================")
+        print("Models loaded. NLU system is ready.")
+        print("========================================================\n")
+    
+    def __predict_intention(self, text: str):
+        return predict_single(self.frame_model, text.split(), 'frame')
+
+    def __predict_slot(self, text: str):
+        anootations = predict_and_annotate(self.slot_model, text.split(), 'slot')
+        current_slot = None
+        current_slot_value = ""
+        slots = []
+
+        for annotation in anootations:
+            form = annotation["form"]
+            slot = annotation["slot"]
+
+            if slot[0:2] == "B-":
+                if current_slot != None:
+                    slots.append(Slot(name=current_slot, value=current_slot_value))
+                current_slot = slot[2:]
+                current_slot_value = form
+            elif slot[0:2] == "I-":
+                current_slot_value = current_slot_value + " " + form
+            elif slot == "O":
+                if current_slot != None:
+                    slots.append(Slot(name=current_slot, value=current_slot_value))
+                    current_slot = None
+                    current_slot_value = ""
+
+        if current_slot != None:
+            slots.append(Slot(name=current_slot, value=current_slot_value))
+
+        return slots
+
+    def process_input(self, text: str):
+        act = self.__predict_intention(text)
+        slots = self.__predict_slot(text)
+        frame = Frame(source = 'user', act = act, slots = slots)
+        return frame

src/service/natural_language_generation.py

@@ -1,5 +1,5 @@
 class NaturalLanguageGeneration:
-    def __init__(self, name):
+    def __init__(self):
         self.name = ["Michał"]
 
     def respond_to_name_query(self, question):

src/utils/nlu_utils.py

@@ -1,101 +1,100 @@
-from flair.data import Sentence
-from flair.datasets import FlairDatapointDataset
-
-def nolabel2o(line, i):
-    return 'O' if line[i] == 'NoLabel' else line[i]
-
-def conllu2flair(sentences, label=None):
-    if label == "frame":
-        return conllu2flair_frame(sentences, label)
-    else:
-        return conllu2flair_slot(sentences, label)
-
-def conllu2flair_frame(sentences, label=None):
-    fsentences = []
-    for sentence in sentences:
-        tokens = [token["form"] for token in sentence]
-        fsentence = Sentence(' '.join(tokens), use_tokenizer=False)
-
-        for i in range(len(fsentence)):
-            fsentence[i:i+1].add_label(label, sentence[i][label])
-
-        fsentences.append(fsentence)
-
-    return FlairDatapointDataset(fsentences)
-
-def conllu2flair_slot(sentences, label=None):
-    fsentences = []
-    for sentence in sentences:
-        fsentence = Sentence(' '.join(token['form'] for token in sentence), use_tokenizer=False)
-        start_idx = None
-        end_idx = None
-        tag = None
-
-        if label:
-            for idx, (token, ftoken) in enumerate(zip(sentence, fsentence)):
-                if token[label].startswith('B-'):
-                    if start_idx is not None:
-                        fsentence[start_idx:end_idx+1].add_label(label, tag)
-                    start_idx = idx
-                    end_idx = idx
-                    tag = token[label][2:]
-                elif token[label].startswith('I-'):
-                    end_idx = idx
-                elif token[label] == 'O':
-                    if start_idx is not None:
-                        fsentence[start_idx:end_idx+1].add_label(label, tag)
-                        start_idx = None
-                        end_idx = None
-                        tag = None
-
-            if start_idx is not None:
-                fsentence[start_idx:end_idx+1].add_label(label, tag)
-
-        fsentences.append(fsentence)
-    return FlairDatapointDataset(fsentences)
-
-def __predict(model, csentence):
-    fsentence = conllu2flair([csentence])[0]
-    model.predict(fsentence)
-    return fsentence
-
-def __csentence(sentence, label_type):
-    if label_type == "frame":
-        return [{'form': word } for word in sentence]
-    else:
-        return [{'form': word, 'slot': 'O'} for word in sentence]
-
-def predict_single(model, sentence, label_type):
-    csentence = __csentence(sentence, label_type)
-    fsentence = __predict(model, csentence)
-    intent = {}
-
-    for span in fsentence.get_spans(label_type):
-        tag = span.get_label(label_type).value
-        if tag in intent:
-            intent[tag] += 1
-        else:
-            intent[tag] = 1
-
-    return max(intent, key=intent.get)
-
-def predict_multiple(model, sentence, label_type):
-    csentence = __csentence(sentence, label_type)
-    fsentence = __predict(model, csentence)
-
-    return set(span.get_label(label_type).value for span in fsentence.get_spans(label_type))
-
-def predict_and_annotate(model, sentence, label_type):
-    csentence = __csentence(sentence, label_type)
-    fsentence = __predict(model, csentence)
-
-    for span in fsentence.get_spans(label_type):
-        tag = span.get_label(label_type).value
-        if label_type == "frame":
-            csentence[span.tokens[0].idx-1]['frame'] = tag
-        else:
-            csentence[span.tokens[0].idx - 1]['slot'] = f'B-{tag}'
-            for token in span.tokens[1:]:
-                csentence[token.idx - 1]['slot'] = f'I-{tag}'
-
+from flair.data import Sentence
+from flair.datasets import FlairDatapointDataset
+
+def nolabel2o(line, i):
+    return 'O' if line[i] == 'NoLabel' else line[i]
+
+def conllu2flair(sentences, label=None):
+    if label == "frame":
+        return conllu2flair_frame(sentences, label)
+    else:
+        return conllu2flair_slot(sentences, label)
+
+def conllu2flair_frame(sentences, label=None):
+    fsentences = []
+    for sentence in sentences:
+        tokens = [token["form"] for token in sentence]
+        fsentence = Sentence(' '.join(tokens), use_tokenizer=False)
+
+        for i in range(len(fsentence)):
+            fsentence[i:i+1].add_label(label, sentence[i][label])
+
+        fsentences.append(fsentence)
+
+    return FlairDatapointDataset(fsentences)
+
+def conllu2flair_slot(sentences, label=None):
+    fsentences = []
+    for sentence in sentences:
+        fsentence = Sentence(' '.join(token['form'] for token in sentence), use_tokenizer=False)
+        start_idx = None
+        end_idx = None
+        tag = None
+
+        if label:
+            for idx, (token, ftoken) in enumerate(zip(sentence, fsentence)):
+                if token[label].startswith('B-'):
+                    if start_idx is not None:
+                        fsentence[start_idx:end_idx+1].add_label(label, tag)
+                    start_idx = idx
+                    end_idx = idx
+                    tag = token[label][2:]
+                elif token[label].startswith('I-'):
+                    end_idx = idx
+                elif token[label] == 'O':
+                    if start_idx is not None:
+                        fsentence[start_idx:end_idx+1].add_label(label, tag)
+                        start_idx = None
+                        end_idx = None
+                        tag = None
+
+            if start_idx is not None:
+                fsentence[start_idx:end_idx+1].add_label(label, tag)
+
+        fsentences.append(fsentence)
+    return FlairDatapointDataset(fsentences)
+
+def __predict(model, csentence):
+    fsentence = conllu2flair([csentence])[0]
+    model.predict(fsentence)
+    return fsentence
+
+def __csentence(sentence, label_type):
+    if label_type == "frame":
+        return [{'form': word } for word in sentence]
+    else:
+        return [{'form': word, 'slot': 'O'} for word in sentence]
+
+def predict_single(model, sentence, label_type):
+    csentence = __csentence(sentence, label_type)
+    fsentence = __predict(model, csentence)
+    intent = {}
+    for span in fsentence.get_spans(label_type):
+        tag = span.get_label(label_type).value
+        if tag in intent:
+            intent[tag] += 1
+        else:
+            intent[tag] = 1
+
+    return max(intent, key=intent.get)
+
+def predict_multiple(model, sentence, label_type):
+    csentence = __csentence(sentence, label_type)
+    fsentence = __predict(model, csentence)
+
+    return set(span.get_label(label_type).value for span in fsentence.get_spans(label_type))
+
+def predict_and_annotate(model, sentence, label_type):
+    csentence = __csentence(sentence, label_type)
+    fsentence = __predict(model, csentence)
+
+    for span in fsentence.get_spans(label_type):
+        tag = span.get_label(label_type).value
+        if label_type == "frame":
+            csentence[span.tokens[0].idx-1]['frame'] = tag
+        else:
+            csentence[span.tokens[0].idx - 1]['slot'] = f'B-{tag}'
+            for token in span.tokens[1:]:
+                csentence[token.idx - 1]['slot'] = f'I-{tag}'
+
     return csentence