TransferTransfo: A Transfer Learning Approach for Neural Network Based Conversational Agents

By Thomas Wolf, Victor Sanh, Julien Chaumond, Clement Delangue. 2019

522 citations recorded by the source corpus.

We introduce a new approach to generative data-driven dialogue systems (e.g. chatbots) called TransferTransfo which is a combination of a Transfer learning based training scheme and a high-capacity Transformer model. Fine-tuning is performed by using a multi-task objective which combines several unsupervised prediction tasks. The resulting fine-tuned model shows strong improvements over the current state-of-the-art end-to-end conversational models like memory augmented seq2seq and information-retrieval models. On the privately held PERSONA-CHAT dataset of the Conversational Intelligence Challenge 2, this approach obtains a new state-of-the-art, with respective perplexity, Hits@1 and F1 metrics of 16.28 (45 % absolute improvement), 80.7 (46 % absolute improvement) and 19.5 (20 % absolute improvement).

Subfields: Artificial Intelligence, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Natural Language Processing

Citations in the corpus

Towards a Human-like Open-Domain Chatbot (2020, positive): “We adapt the decoding implementation by Wolf et al. (2019).”
Open-Domain Conversational Agents: Current Progress, Open Problems, and Future Directions (2020, methodology): “Pretraining First, we have found that pretraining is important to performance for nearly every variant of chitchat we have experimented with (Wolf et al., 2019b; Dinan et al., 2020), including both in retrieval (Humeau et al.”
HuggingFace's Transformers: State-of-the-art Natural Language Processing (2019, positive): “The approach used to reach these performances is described in Wolf et al. (2019); Golovanov et al. (2019) and the code and pretrained models, based on the Transformers library, are available online 6 .”
ACUTE-EVAL: Improved Dialogue Evaluation with Optimized Questions and Multi-turn Comparisons (2019, methodology): “In this work, we leverage the human-model chat logs from the ConvAI2 competition for three models: Lost in Conversation (LIC)1, which won the competition, and Hugging Face (HF; Wolf et al., 2019) which won the automatic evaluation track, and the KVMemNN (Miller et al.”
Poly-encoders: Architectures and Pre-training Strategies for Fast and Accurate Multi-sentence Scoring (2019, methodology): “The best performing competitor out of 23 entrants in this task achieved 80.7% accuracy on the test set utilizing a pre-trained Transformer ﬁne-tuned for this task (Wolf et al., 2019).”
BlenderBot 3: a deployed conversational agent that continually learns to responsibly engage (2022, background): “(at the time) pre-trained transformers being used by the top two winning teams (Wolf et al., 2019; Golovanov et al., 2020; Dinan et al., 2020c).”
Survey of Hallucination in Natural Language Generation (2022, background): “The PersonaChat datasets [31, 228] have accelerated research into persona consistency [68, 91, 126, 208, 219, 224, 232].”
Recipes for Building an Open-Domain Chatbot (2020, methodology): “In particular, Wolf et al. (2019) pre-trained via the method of Radford et al. (2018) using the BooksCorpus dataset, resulting in the best per-plexities and F1 scores.”
Exploring Versatile Generative Language Model Via Parameter-Efficient Transfer Learning (2020, methodology): “We implement VLM based on GPT-2-small (124M) (Wolf et al., 2019a), and experiment with varying adapter bottleneck dimensions in { 10 , 50 , 100 , 300 } and pick the best one in each task to trade-off the performance with the parameter efﬁciency.”
HuggingFace's Transformers: State-of-the-art Natural Language Processing (2019, neutral): “2 held at the Thirty-second Annual Conference on Neural Information Processing Systems (NIPS 2018). The approach used to reach 5https://transformer.huggingface.co 5 these performances is described in Wolf et al. (2019); Golovanov et al. (2019) and the code and pretrained models, based on the Transfo”
Large-scale Pretraining for Visual Dialog: A Simple State-of-the-Art Baseline (2019, methodology): “[59], we use di erent segment embeddings for questions and answers to help the model distinguish between the two and understand question”
DIALOGPT : Large-Scale Generative Pre-training for Conversational Response Generation (2019, methodology): “Modelling conversations, however, presents distinct challenges in that human dialogue, which encapsulates the possibly competing goals of two participants, is intrinsically more diverse in the range of potential responses (Li et al., 2016a; Zhang et al., 2018; Gao et al., 2019a,b).”
Learning to Speak and Act in a Fantasy Text Adventure Game (2019, methodology): “Second, the BERT-based Cross-Ranker instead concatenates the context with each candidate ut-terance, similar to Wolf et al. (2019).”
The Dialogue Dodecathlon: Open-Domain Knowledge and Image Grounded Conversational Agents (2019, methodology): “The implementation relies on HuggingFace Transformers (Wolf et al., 2019a).”
Hurdles to Progress in Long-form Question Answering (2021, neutral): “3 We ﬁne-tune our model in a decoder-only fashion (Liu et al., 2018; Wolf et al., 2018) by concatenating the top K retrieved documents to the question [ r i,K , r i,K − 1 ... r i, 1 , q i , a i ] and training the model to predict tokens of the answer a i .”
Neural Generation Meets Real People: Towards Emotionally Engaging Mixed-Initiative Conversations (2020, methodology): “The model was finetuned using the HuggingFace ConvAI code15 (Wolf et al., 2019b) and is hosted on a GPU-enabled EC2 machine with one NVIDIA T4 Tensor Core GPU.”
Challenges in Building Intelligent Open-domain Dialog Systems (2019, methodology): “-trained BERT model or a GPT-2 model, the decoder a GPT-2 model. Both the parameters of the encoder and the decoder are initialized using the pre-trained models and then fine-tuned on a dialog corpus [1, 39, 106, 155, 179]. The fine-tuning process is often tailored to the dialog scenario via encodin”
USR: An Unsupervised and Reference Free Evaluation Metric for Dialog Generation (2020, methodology): “Recent research has highlighted the complementary nature of dialog retrieval and generation with respect to multi-tasking (Wolf et al., 2019b) and pre-training (Mehri et al., 2019).”