XLMRoberta-Alexa-Intents-NER-NLU Open-Source Model - Supports Intent and Entity Recognition in 51 Languages

Xlmroberta Alexa Intents NER NLU

Developed by qanastek

XLM-Roberta-based multilingual natural language understanding model supporting intent recognition and named entity recognition in 51 languages

Sequence Labeling

Transformers

#Multilingual slot filling #Voice assistant NLU #High-precision intent recognition

Downloads 18

Release Time : 5/8/2022

Model Overview

This model is a multilingual sequence labeling model specifically designed for natural language understanding tasks in voice assistants, capable of recognizing 60 intents and 55 slot types.

Model Features

Multilingual support

Supports intent recognition and named entity recognition in 51 languages

Extensive entity coverage

Capable of recognizing 55 different types of named entities

High-precision recognition

Achieves high F1 scores across multiple entity types, such as 0.8593 F1 for time recognition

Voice assistant optimization

Specifically optimized for voice assistant scenarios, covering 60 common intents

Model Capabilities

Intent recognition

Named entity recognition

Slot filling

Multilingual processing

Voice command understanding

Use Cases

Smart voice assistants

Alarm setting

Recognizes the time and date when users set alarms

Time recognition F1 score 0.8593, date recognition F1 score 0.8995

Media playback control

Recognizes songs, artists or podcasts requested by users

Artist name recognition F1 score 0.7757, song name recognition F1 score 0.6433

Information query

Recognizes stock, weather or location information queried by users

Business name recognition F1 score 0.8075, location name recognition F1 score 0.8417

Multilingual applications

Cross-language command understanding

Understands user commands with the same intent across different languages

Supports identical intent recognition in 51 languages

🚀 XLMRoberta-Alexa-Intents-NER-NLU

A token classification model for intent prediction and slot annotation across 51 languages.

🚀 Quick Start

Prerequisites

This project requires the transformers library. You can install it using the following command:

pip install transformers

Usage Example

from transformers import AutoTokenizer, AutoModelForTokenClassification, TokenClassificationPipeline

tokenizer = AutoTokenizer.from_pretrained('qanastek/XLMRoberta-Alexa-Intents-NER-NLU')
model = AutoModelForTokenClassification.from_pretrained('qanastek/XLMRoberta-Alexa-Intents-NER-NLU')
predict = TokenClassificationPipeline(model=model, tokenizer=tokenizer)
res = predict("réveille-moi à neuf heures du matin le vendredi")
print(res)

The output will be similar to the following: English - Hebrew - Spanish

[{'word': '▁neuf', 'score': 0.9911066293716431, 'entity': 'B-time', 'index': 6, 'start': 15, 'end': 19},
{'word': '▁heures', 'score': 0.9200698733329773, 'entity': 'I-time', 'index': 7, 'start': 20, 'end': 26},
{'word': '▁du', 'score': 0.8476170897483826, 'entity': 'I-time', 'index': 8, 'start': 27, 'end': 29},
{'word': '▁matin', 'score': 0.8271021246910095, 'entity': 'I-time', 'index': 9, 'start': 30, 'end': 35},
{'word': '▁vendredi', 'score': 0.9813069701194763, 'entity': 'B-date', 'index': 11, 'start': 39, 'end': 47}]

✨ Features

Multilingual Support: Supports 51 languages, including Afrikaans, Arabic, Chinese, English, French, German, Hebrew, Hindi, Japanese, and more.
Intent Prediction and Slot Annotation: Capable of predicting intents and annotating slots in natural language utterances.
Trained on MASSIVE Dataset: Utilizes the MASSIVE dataset, which contains over 1 million utterances across 51 languages.

📦 Installation

To use this model, you need to install the transformers library:

pip install transformers

📚 Documentation

Training Data

MASSIVE is a parallel dataset of over 1 million utterances across 51 languages. It includes annotations for intent prediction and slot annotation tasks. The utterances cover 60 intents and 55 slot types. MASSIVE was created by localizing the SLURP dataset, which consists of general Intelligent Voice Assistant single-shot interactions.

Named Entities

The model can recognize the following named entities:

O
currency_name
personal_info
app_name
list_name
alarm_type
cooking_type
time_zone
media_type
change_amount
transport_type
drink_type
news_topic
artist_name
weather_descriptor
transport_name
player_setting
email_folder
music_album
coffee_type
meal_type
song_name
date
movie_type
movie_name
game_name
business_type
music_descriptor
joke_type
music_genre
device_type
house_place
place_name
sport_type
podcast_name
game_type
timeofday
business_name
time
definition_word
audiobook_author
event_name
general_frequency
relation
color_type
audiobook_name
food_type
person
transport_agency
email_address
podcast_descriptor
order_type
ingredient
transport_descriptor
playlist_name
radio_name

Evaluation Results

The following table shows the evaluation results of the model:

                      precision    recall  f1-score   support

                   O     0.9537    0.9498    0.9517   1031927
          alarm_type     0.8214    0.1800    0.2953       511
            app_name     0.3448    0.5318    0.4184       660
         artist_name     0.7143    0.8487    0.7757     11413
    audiobook_author     0.7038    0.2971    0.4178      1232
      audiobook_name     0.7271    0.5381    0.6185      5090
       business_name     0.8301    0.7862    0.8075     15385
       business_type     0.7009    0.6196    0.6577      4600
       change_amount     0.8179    0.9104    0.8617      1663
         coffee_type     0.6147    0.8322    0.7071       876
          color_type     0.6999    0.9176    0.7941      2890
        cooking_type     0.7037    0.5184    0.5970      1003
       currency_name     0.8479    0.9686    0.9042      6501
                date     0.8667    0.9348    0.8995     49866
     definition_word     0.9043    0.8135    0.8565      8333
         device_type     0.8502    0.8825    0.8661     11631
          drink_type     0.0000    0.0000    0.0000       131
       email_address     0.9715    0.9747    0.9731      3986
        email_folder     0.5913    0.9740    0.7359       884
          event_name     0.7659    0.7630    0.7645     38625
           food_type     0.6502    0.8697    0.7441     12353
           game_name     0.8974    0.6275    0.7386      4518
   general_frequency     0.8012    0.8673    0.8329      3173
         house_place     0.9337    0.9168    0.9252      7067
          ingredient     0.5481    0.0491    0.0901      1161
           joke_type     0.8147    0.9101    0.8598      1435
           list_name     0.8411    0.7275    0.7802      8188
           meal_type     0.6072    0.8926    0.7227      2282
          media_type     0.8578    0.8522    0.8550     17751
          movie_name     0.4598    0.1856    0.2645       431
          movie_type     0.2673    0.4341    0.3309       364
         music_album     0.0000    0.0000    0.0000       146
    music_descriptor     0.2906    0.3979    0.3359      1053
         music_genre     0.7999    0.7483    0.7732      5908
          news_topic     0.7052    0.5702    0.6306      9265
          order_type     0.6374    0.8845    0.7409      2614
              person     0.8173    0.9376    0.8733     33708
       personal_info     0.7035    0.7444    0.7234      1976
          place_name     0.8616    0.8228    0.8417     38881
      player_setting     0.6429    0.6212    0.6319      5409
       playlist_name     0.5852    0.5293    0.5559      3671
  podcast_descriptor     0.7486    0.5413    0.6283      4951
        podcast_name     0.6858    0.5675    0.6211      3339
          radio_name     0.8196    0.8013    0.8103      9892
            relation     0.6662    0.8569    0.7496      6477
           song_name     0.5617    0.7527    0.6433      7251
          sport_type     0.0000    0.0000    0.0000         0
                time     0.9032    0.8195    0.8593     35456
           time_zone     0.8368    0.4467    0.5824      2823
           timeofday     0.7931    0.8459    0.8187      6140
    transport_agency     0.7876    0.7764    0.7820      1051
transport_descriptor     0.5738    0.2756    0.3723       254
      transport_name     0.8497    0.5149    0.6412      1010
      transport_type     0.9303    0.8980    0.9139      6363
  weather_descriptor     0.8584    0.7466    0.7986     11702

            accuracy                         0.9092   1455270
           macro avg     0.6940    0.6668    0.6613   1455270
        weighted avg     0.9111    0.9092    0.9086   1455270

📄 License

This project is licensed under the CC BY 4.0 license.

👥 People Involved

LABRAK Yanis (1)

Affiliations

LIA, NLP team, Avignon University, Avignon, France.

Featured Recommended AI Models

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご