Few-Shot-Intent-Detection

:bangbang: ❤️ ‼️ **07/18/2023: Check our latest updates on DialogStudio

DialogStudio is a meticulously curated collection of dialogue datasets. These datasets are unified under a consistent format while retaining their original information. We incorporate domain-aware prompts and identify dataset licenses, making DialogStudio an exceptionally rich and diverse resource for dialogue research and model training.**

Few-Shot-Intent-Detection is a repository designed for few-shot intent detection with/without Out-of-Scope (OOS) intents. It includes popular challenging intent detection datasets and baselines. For more details of the new released OOS datasets, please check our paper.

Intent detection datasets

We process data based on previous published resources, all the data are in the same format as DNNC.

Intent detection datasets with OOS queries

What is OOS queires:

OOD-OOS: i.e., out-of-domain OOS. General out-of-scope queries which are not supported by the dialog systems, also called out-of-domain OOS. For instance, requesting an online NBA/TV show service in a banking system.

ID-OOS: i.e., in-domain OOS. Out-of-scope queries which are more related to the in-scope intents, which makes the intent detection task more challenging. For instance, requesting a banking service that is not supported by the banking system.

Data structure:

Datasets/
├── BANKING77
│   ├── train
│   ├── train_10
│   ├── train_5
│   ├── valid
│   └── test
├── CLINC150
│   ├── train
│   ├── train_10
│   ├── train_5
│   ├── valid
│   ├── test
│   ├── oos
│       ├──train
│       ├──valid
│       └──test
├── HWU64
│   ├── train
│   ├── train_10
│   ├── train_5
│   ├── valid
│   └── test
├── SNIPS
│   ├── train
│   ├── valid
│   └── test
├── ATIS
│   ├── train
│   ├── valid
│   └── test
├── BANKING77-OOS
│   ├── train
│   ├── valid
│   ├── test
│   ├── id-oos
│   │   ├──train
│   │   ├──valid
│   │   └──test
│   ├── ood-oos
│       ├──valid
│       └──test
├── CLINC-Single-Domain-OOS
│   ├── banking
│   │   ├── train
│   │   ├── valid
│   │   ├── test
│   │   ├── id-oos
│   │   │   ├──valid
│   │   │   └──test
│   │   ├── ood-oos
│   │       ├──valid
│   │       └──test
│   ├── credit_cards
│   │   ├── train
│   │   ├── valid
│   │   ├── test
│   │   ├── id-oos
│   │   │   ├──valid
│   │   │   └──test
│   │   ├── ood-oos
│   │       ├──valid
└── └──     └──test

Briefly describe the BANKING77-OOS dataset.

• A dataset with a single banking domain, includes both general Out-of-Scope (OOD-OOS) queries and In-Domain but Out-of-Scope (ID-OOS) queries, where ID-OOS queries are semantically similar intents/queries with in-scope intents. BANKING77 originally includes 77 intents. BANKING77-OOS includes 50 in-scope intents in this dataset, and the ID-OOS queries are built up based on 27 held-out semantically similar in-scope intents.

Briefly describe the CLINC-Single-Domain-OOS dataset.

• A dataset with two separate domains, i.e., the "Banking'' domain and the "Credit cards'' domain with both general Out-of-Scope (OOD-OOS) queries and In-Domain but Out-of-Scope (ID-OOS) queries, where ID-OOS queries are semantically similar intents/queries with in-scope intents. Each domain in CLINC150 originally includes 15 intents. Each domain in the new dataset includes ten in-scope intents in this dataset, and the ID-OOS queries are built up based on five held-out semantically similar in-scope intents.

Both datasets can be used to conduct intent detection with and without OOD-OOS and ID-OOS queries

You can easily load the processed data:

class IntentExample:
    def __init__(self, text, label, do_lower_case):
        self.original_text = text
        self.text = text
        self.label = label

        if do_lower_case:
            self.text = self.text.lower()
        
def load_intent_examples(file_path, do_lower_case=True):
    examples = []

    with open('{}/seq.in'.format(file_path), 'r', encoding="utf-8") as f_text, open('{}/label'.format(file_path), 'r', encoding="utf-8") as f_label:
        for text, label in zip(f_text, f_label):
            e = IntentExample(text.strip(), label.strip(), do_lower_case)
            examples.append(e)

    return examples

More details can check code for load data and do random sampling for few-shot learning.

State-of-the art models and baselines

DNNC

Download pre-trained RoBERTa NLI checkpoint:

wget https://storage.googleapis.com/sfr-dnnc-few-shot-intent/roberta_nli.zip

Access to public code: Link

CONVERT

Download pre-trained checkpoint:

wget https://github.com/connorbrinton/polyai-models/releases/download/v1.0/model.tar.gz

Access to public code:

wget https://github.com/connorbrinton/polyai-models/archive/refs/tags/v1.0.zip

CONVBERT

Download pre-trained checkpoints:

Step-1: install AWS CL2: e.g., install MacOS PKG

Step-2:

aws s3 cp s3://dialoglue/ --no-sign-request `Your_folder_name` --recursive

Then the checkpoints are downloaded into Your_folder_name

Few-shot intent detection baselines/leaderboard:

5-shot learning

10-shot learning

Note: the 5-shot learning results of RoBERTa+Classifier, DNNC and CPFT, and the 10-shot learning results of all the models are reported by the paper authors.

Citation

Please cite our paper if you use above resources in your work:

@article{zhang2020discriminative,
  title={Discriminative nearest neighbor few-shot intent detection by transferring natural language inference},
  author={Zhang, Jian-Guo and Hashimoto, Kazuma and Liu, Wenhao and Wu, Chien-Sheng and Wan, Yao and Yu, Philip S and Socher, Richard and Xiong, Caiming},
  journal={EMNLP},
  pages={5064--5082},
  year={2020}
}
@article{zhang2021few,
  title={Few-Shot Intent Detection via Contrastive Pre-Training and Fine-Tuning},
  author={Zhang, Jianguo and Bui, Trung and Yoon, Seunghyun and Chen, Xiang and Liu, Zhiwei and Xia, Congying and Tran, Quan Hung and Chang, Walter and Yu, Philip},
  journal={EMNLP},
  year={2021}
}
@article{zhang2022pretrained,
  title={Are Pretrained Transformers Robust in Intent Classification? A Missing Ingredient in Evaluation of Out-of-Scope Intent Detection},
  author={Zhang, Jian-Guo and Hashimoto, Kazuma and Wan, Yao and Liu, Zhiwei and Liu, Ye and Xiong, Caiming and Yu, Philip S},
  journal={The 4th Workshop on NLP for Conversational AI, ACL 2022},
  year={2022}
}