[Paper Review] A Persona-Based Neural Conversation Model

Jiwei Li, Michel Galley+4 more|arXiv (Cornell University)|Mar 19, 2016

Topic ModelingComputer Science32 references181 citations

TL;DR

Introduces speaker and speaker-addressee persona embeddings in Seq2Seq conversation models to improve consistency and fluency, with gains in perplexity and Bleu on open-domain Twitter and TV-series dialogue data.

ABSTRACT

We present persona-based models for handling the issue of speaker consistency in neural response generation. A speaker model encodes personas in distributed embeddings that capture individual characteristics such as background information and speaking style. A dyadic speaker-addressee model captures properties of interactions between two interlocutors. Our models yield qualitative performance improvements in both perplexity and BLEU scores over baseline sequence-to-sequence models, with similar gains in speaker consistency as measured by human judges.

Motivation & Objective

• Address speaker consistency in neural response generation.
• Learn and inject persona representations into Seq2Seq decoders.
• Model both a single-speaker and a dyadic speaker–addressee setting to capture personal and interactive styles.

Proposed method

• Encode speaker personas as trainable embeddings v_i and inject into the decoder LSTM at each time step.
• Extend to a Speaker-Addressee model by constructing an interaction vector V_{i,j} = tanh(W1 v_i + W2 v_j) and incorporating it into decoding.
• Train end-to-end on large conversational corpora (Twitter, TV-series scripts) with backpropagation.
• Use beam search (B=200) for decoding and rerank N-best lists with a scoring function that combines p(R|M,v), p(M|R), and length penalty.
• Evaluate using perplexity and Bleu, with human judgments for consistency.

Experimental results

Research questions

• RQ1 Can persona embeddings improve response consistency and diversity in open-domain dialogue?
• RQ2 Do speaker and speaker–addressee models yield better perplexity and Bleu than non-persona Seq2Seq baselines?
• RQ3 How does the dyadic interaction model affect generation when speakers interact with different addressees?
• RQ4 What is the impact of decoding and reranking strategies on output quality and consistency?

Key findings

• Speaker Model reduces perplexity by about 10% over standard Seq2Seq on the Twitter Persona dataset (perplexity 42.2 vs. 47.2).
• Bleu gains for the Speaker Model over standard Seq2Seq reach +21.7% under MLE and +11.7% under MMI on the Twitter Persona dataset.
• On the TV-series dataset, both Speaker and Speaker-Addressee models reduce perplexity by about 7–8% relative to standard Seq2Seq and improve Bleu by roughly 10–14%.
• Qualitative analyses show the Speaker Model yields diverse, speaker-specific responses and some consistency reductions, while the Speaker-Addressee model demonstrates dyadic sensitivity (e.g., personalized addressee responses).
• Human evaluation indicates the persona models are judged more consistent than the baseline in a majority of comparisons (56.7% clearly more consistent; 6.1% substantially more consistent when excluding ties).