Generating Informative and Diverse Conversational Responses via Adversarial Information Maximization

By Yizhe Zhang, Michel Galley, Jianfeng Gao, Zhe Gan, Xiujun Li, Chris Brockett, William B. Dolan. 2018

318 citations recorded by the source corpus.

Responses generated by neural conversational models tend to lack informativeness and diversity. We present Adversarial Information Maximization (AIM), an adversarial learning framework that addresses these two related but distinct problems. To foster response diversity, we leverage adversarial training that allows distributional matching of synthetic and real responses. To improve informativeness, our framework explicitly optimizes a variational lower bound on pairwise mutual information between query and response. Empirical results from automatic and human evaluations demonstrate that our methods significantly boost informativeness and diversity.

Subfields: Artificial Intelligence, Statistical and Nonlinear Physics, 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 Statistical and Nonlinear Physics

In Natural Language Processing

Citations in the corpus

Towards a Human-like Open-Domain Chatbot (2020, background): “Recent works also explore new frameworks such as adversarial learning (Li et al., 2017; Zhang et al., 2018c), variational autoencod-ing (Zhao et al., 2017; Gu et al., 2019), or both (Gao et al., 2019b) at the cost of added complexity and less scalability.”
Towards Coherent and Cohesive Long-form Text Generation (2018, background): “, 2014) to incorporate feedback signals for text generation (Yu et al., 2017; Lin et al., 2017; Zhang et al., 2017; Guo et al., 2017; Fedus et al., 2018; Zhang et al., 2018).”
Red Teaming Language Models with Language Models (2022, methodology): “Thus, we also measure using the entropy of the n-gram distribution, following prior work in dialogue (Zhang et al., 2018).”
Locating and Editing Factual Associations in GPT (2022, methodology): “Finally, we monitor fluency degradations by measuring the weighted average of bi- and tri-gram entropies (Zhang et al., 2018) given by − ∑ k f(k) log2 f(k), where f(·) is the n-gram frequency distribution, which we report as (GE); this quantity drops if text generations are repetitive.”
DIALOGPT : Large-Scale Generative Pre-training for Conversational Response Generation (2019, methodology): “To address this problem, we implement a maximum mutual information (MMI) scoring function (Li et al., 2016a; Zhang et al., 2018).”
Neural Approaches to Conversational AI (2018, neutral): “More recently, researchers (Li et al., 2017c; Xu et al., 2017; Zhang et al., 2018e) have used adversarial training and Generative Adversarial Networks (GAN) (Goodfellow et al.”
Conversing by Reading: Contentful Neural Conversation with On-demand Machine Reading (2019, methodology): “Prior work, e.g, (Ghazvininejad et al., 2018; Zhang et al., 2018a; Huang et al., 2019), uses grounding in the form of independent snippets of text: Foursquare tips and background information about a given speaker.”
Negative Training for Neural Dialogue Response Generation (2019, methodology): “Similar to negative training, our experiments show that positive training (or “teacher forcing” in some literature) is crucial to aid the model to maintain its original performance, during GAN training.”
Locating and Editing Factual Knowledge in GPT (2022, background): “Given some generation x, the n-gram entropy (Zhang et al., 2018) is given by − ∑ k f(k) log2 f(k), where k is an n-gram, and f(k) is its relative frequency.”
Conversations Are Not Flat: Modeling the Dynamic Information Flow across Dialogue Utterances (2021, unknown): “For the lexical diversity, DialoFlow performs similarly with the DialoGPT on Entropy.”
Improving Generative Visual Dialog by Answering Diverse Questions (2019, unknown): “Adding constraints to generate a diverse set of natural language dialog responses have previously been explored in Zhang et al. (2018) via adversarial information maximization, in Gao et al. (2019) by jointly modeling diversity and relevance in a shared latent space, and in Li et al. (2016a) us-ing…”
Diversifying Dialogue Generation with Non-Conversational Text (2020, methodology): “The backward seq2seq model measures the context relevance better than forward models since the latter highly biases generic utterances (Li et al., 2016a; Zhang et al., 2018b)”
Comparison of Diverse Decoding Methods from Conditional Language Models (2019, methodology): “Zhang et al. (2018) instead propose an entropy metric, Ent-k , deﬁned as: where S is the set of all k -grams that appear in candidate responses for an example, and F ( w ) denotes the frequency of w in the candidate responses.”
Jointly Optimizing Diversity and Relevance in Neural Response Generation (2019, background): “While various approaches have been explored to diversify the output of the conversation model, the improvement often comes at the cost of decreased relevance (Zhang et al., 2018).”
Trading Off Diversity and Quality in Natural Language Generation (2020, background): “Zhang et al. (2018) and Xu et al. (2018) use adversarial learning methods to encourage diversity.”
Cluster-based beam search for pointer-generator chatbot grounded by knowledge (2020, methodology): “Other metrics such as DIV-1 and DIV-2 (also known as distinct-1 and distinct2) (Li et al. 2016) and Entropy1-4 (Zhang et al. 2018) were also employed to measure diversity.”
Grounded Conversation Generation as Guided Traverses in Commonsense Knowledge Graphs (2019, methodology): “Nevertheless, the degenerating of irrelevant, off-topic, and non-useful responses is still one of the main challenges in conversational generation (Rosset et al., 2020; Tang et al., 2019; Zhang et al., 2018; Gao et al., 2019).”
Diversifying Content Generation for Commonsense Reasoning with Mixture of Knowledge Graph Experts (2022, background): “Entropy-k (Zhang et al., 2018) reflects how evenly the empirical k -gram distribution is for a given sentence when word frequency is considered.”
Multi-turn Dialogue Response Generation with Autoregressive Transformer Models (2019, methodology): “…limitations of the MLE objective alone (Li et al. 2016a; Li et al. 2016b; Li et al. 2017; Yu et al. 2017; Che et al. 2017; Zhang et al. 2017; Xu et al. 2017; Serban et al. 2017b; Zhang et al. 2018b; Olabiyi et al. 2018; Olabiyi, Khazan, and Mueller 2018; Olabiyi et al. 2019a; Olabiyi et al. 2019b).”
Moving the Eiffel Tower to ROME: Tracing and Editing Facts in GPT (2021, background): “One might rec-852 ognize E n ( g ) as the n -gram entropy of g , which penalizes generations with high repetition (Zhang et al., 2018).”