Detecting Intent Drift in Continuous Conversation via Temporal Transition Accumulation

📂 Dataset and Preprocessing Instructions

We use the following publicly available datasets in our experiments:

ATIS(Airline Travel Information System)

SIM(The Simulated-Dialogue dataset)

Multiwoz(Multi-Domain Wizard-of-Oz Datase)

Due to the large size of the processed MultiWOZ dataset, we only include the preprocessed versions of ATIS and SIM in our repository. However, we provide the necessary scripts to preprocess all datasets, including MultiWOZ.

The preprocessing script is located at:

/data/process_data.py

📌 Before running the script, we recommend merging all dialogues from the original datasets into a single file for each dataset.

Detailed preprocessing instructions can be found in:

/data/README.md

Once the data is prepared, you can run the training and evaluation pipeline using:

/train/train_and_eval_with_DriftHunter.py

And the result on SIM test set will be saved into:

/train/results/AGLCF_DriftHunter/sim

---

🔧 Hyperparameters: α, β, and γ

The optimal values of hyperparameters α, β, and γ are set individually for each dataset and method as follows:

📘 ATIS Dataset

Method	α	β	γ
BERT-DH	0.05	0.15	0.80
RNNContextual-DH	0.05	0.25	0.70
AGLCF-DH	0.05	0.25	0.70
DHLG-DH	0.10	0.20	0.70

📗 SIM Dataset

Method	α	β	γ
BERT-DH	0.05	0.15	0.80
RNNContextual-DH	0.05	0.25	0.70
AGLCF-DH	0.05	0.25	0.70
DHLG-DH	0.10	0.30	0.60

📙 MultiWOZ Dataset

Method	α	β	γ
BERT-DH	0.05	0.15	0.80
RNNContextual-DH	0.25	0.25	0.50
AGLCF-DH	0.05	0.35	0.60
DHLG-DH	0.05	0.35	0.60

---