Source-linked AI summary
A Persona-Based Neural Conversation Model
Jiwei Li, Michel Galley, Chris Brockett, Georgios P. Spithourakis, Jianfeng Gao, Bill Dolan
TL;DR
Neural conversation systems can produce vague or inconsistent responses because likelihood-based generation favors consensus behavior. The paper introduces speaker and speaker-addressee persona embeddings within SEQ2SEQ models, achieving improvements in BLEU, perplexity, and human-judged speaker consistency. The models capture individual characteristics and dyadic interaction patterns, while their effectiveness involves trade-offs and data-collection constraints.
Problem
Data-driven conversation systems need coherent personas to address vague or inconsistent responses produced by likelihood-based generation.
Method
The paper encodes speaker characteristics and speaker-addressee interaction patterns as embeddings within SEQ2SEQ response models.
Results
The models outperform baseline SEQ2SEQ systems in BLEU, perplexity, and human judgments of speaker consistency.
Takeaways & Limitations
Distributed persona representations capture characteristics such as speaking style and background information, with evidence of benefit from modeling dyadic interactions.
Takeaways & Limitations
Speaker-addressee modeling requires large-scale data covering speakers’ conversations with varied interlocutors, and persona embeddings trade off persona-specific and general conversational modeling.
Abstract
from arXiv · showhide
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.
1 Introduction
Data-driven conversation systems often produce generic or inconsistent responses because likelihood-based generation favors consensus behavior. This paper addresses speaker consistency by encoding personas and speaker–addressee interactions in neural response models.
- Likelihood-based systems tend to select consensus responses, producing vague or inconsistent outputs.
- The paper defines persona as an agent’s identity, language behavior, and interaction style, which can adapt to different interlocutors.
- Neural conversation models can incorporate personas through distributed embeddings.
- The proposed Speaker Model represents one speaker, while the Speaker-Addressee Model represents interactions between two interlocutors within SEQ2SEQ.
- Up to 20% relative BLEU improvement and 12% perplexity improvement were achieved, alongside human-judged gains in consistency.
2 Related Work
Related work moves from rule-based and statistical dialogue generation toward end-to-end neural response generation. This paper extends that direction by learning persona vectors from conversational data and side information.
- Earlier dialogue systems commonly relied on hand-coded rules, templates, or authored labels for generation.
- Ritter et al. introduced an end-to-end, purely data-driven SMT model without an explicit dialogue-structure model.
- Neural SEQ2SEQ models were applied directly to response generation, while hierarchical models targeted extended conversation history.
- Mutual-information objectives were used to reduce generic responses typical of SEQ2SEQ systems.
- The present work enriches SEQ2SEQ models by training persona vectors from conversational data and relevant side information.
3 Sequence-to-Sequence Models
The SEQ2SEQ framework uses LSTMs to encode input sequences and predict output tokens sequentially. Separate input and output LSTMs model their distinct compositional patterns, with decoding ending at EOS.
- An LSTM processes each input time step using input, memory, and output gates.The hidden vector combines the current text-unit representation with the previous hidden state.
- SEQ2SEQ tasks pair an input sequence X with an output sequence Y to be predicted.
- The LSTM predicts output tokens sequentially using a softmax-defined distribution.
- Input and output sequences use separate LSTMs with separate parameters to capture different compositional patterns.
- Decoding terminates when the model predicts the special EOS token, using either greedy prediction or beam search.
4 Personalized Response Generation
The paper introduces speaker and speaker-addressee persona representations for personalized response generation. These representations condition sequence-to-sequence decoding on respondent identity and interlocutor-specific interaction patterns.
- Speaker Model: The Speaker Model represents each respondent with an embedding encoding speaker-specific information that influences response content and style.The attributes are not explicitly annotated; the model learns representations from conversational responses.
- Speaker Model: Speaker information is injected into the target hidden layer at every generation step to help predict personalized responses.Speaker embeddings are shared across conversations involving the same speaker and learned through backpropagated word-prediction errors.
- Speaker-Addressee Model: The Speaker-Addressee Model combines respondent and addressee embeddings into an interaction representation that conditions target-side LSTM states.This representation models how speaker i responds to addressee j rather than modeling the respondent alone.
- Speaker-Addressee Model: The same speaker can respond differently to different interlocutors because the interaction representation depends on both speaker and addressee.Embedding-based generalization can help even when the test-time pair has not appeared together in training.
- Decoding: Beam search uses B = 200 and a maximum candidate length of 20, followed by reranking to address generic responses from sequence-to-sequence models.The reranking score combines a length penalty with the log likelihood of the source given the target, with weights optimized for BLEU on development data.
5 Datasets
The paper evaluates persona models using Twitter conversations and television-script dialogues, with distinct datasets and training strategies for each setting. Dataset design also constrains which model can be tested and how BLEU comparisons should be interpreted.
- Twitter Persona Dataset: 24,725,711 3-turn conversational sequences were extracted from Twitter responses by 74,003 users with repeated conversational activity.Users participated in 60–164 conversational turns, averaging 92.24 and with a median of 90.
- Twitter Persona Dataset: 12,000 additional Twitter conversations were reserved equally for development, validation, and testing using single-reference responses.Single references avoid confounding persona differences that would arise from collecting responses from different people.
- Twitter Persona Dataset: Experiments on the Twitter Persona Dataset were limited to the Speaker Model because only response speaker IDs were available.The dataset therefore did not support evaluating the dyadic Speaker-Addressee Model.
- Twitter Sordoni Dataset: The Twitter Sordoni Dataset contains 2,114 test contexts and messages with up to 10 references per message for comparison with prior work.Its multi-reference structure makes its BLEU scores incomparable with the single-reference Twitter Persona Dataset.
- Television Series Transcripts: The Speaker-Addressee experiments used 69,565 turns from 13 main characters in Friends and The Big Bang Theory, with development and test sets of about 2,000 turns each.Because this corpus was relatively small, the model used domain adaptation from a much larger OpenSubtitles dataset.
6 Experiments
Experiments show that persona conditioning improves response-model performance and produces speaker- and addressee-specific outputs across Twitter and TV-series data. Human judges also favored persona-model outputs for consistency over a strong baseline.
- Twitter Persona results: About 10% lower perplexity was observed for the Speaker model than the standard SEQ2SEQ model on the Twitter Persona dataset.The reported comparison covers the development set.
- Twitter Persona results: 21% higher BLEU was obtained with the Speaker model under MLE, while the MMI setting showed an 11.7% increase on the Twitter Persona dataset.Persona models were more beneficial to MLE models than to MMI models.
- TV-series results: 25.4 and 25.0 perplexity were achieved by the Speaker and Speaker-Addressee models, versus 27.3 for Standard LSTM on the TV-series dataset.These correspond to reductions of 7.0% and 8.4%, respectively.
- Qualitative analysis: Persona models generated speaker-specific factual responses and addressee-sensitive language, including names and gendered pronouns in TV-series examples.The Speaker model used randomly selected users, while the Speaker-Addressee model varied responses with interlocutor identity.
- Human evaluation: 56.7% of non-tied human judgments rated the persona model more or somewhat more consistent than the baseline.Ignoring somewhat-more judgments, the persona model won 6.1% of cases versus 1.6% for the baseline.
- Caveats: The experiments retain important data limitations, including a simplifying speaker-turn assumption for OSDb and incomplete consistency in some generated answers.The OSDb assumption treats each subtitle line as a full speaker turn, despite adjacent lines not necessarily sharing a scene or speaker.
7 Conclusions
The paper presents persona-based models that capture individual characteristics and dyadic interactions in open-domain conversation generation. These models outperform baseline SEQ2SEQ systems across automatic metrics and human judgments, while mood and emotion remain outside the paper’s scope.
- Two persona-based response generation models encode speaker characteristics for open-domain conversation.The models capture personal characteristics such as speaking style and background information, with the Speaker-Addressee model additionally representing dyadic interactions.
- Mood and emotion, among other dimensions of speaker behavior, remain beyond the paper’s scope and are left for future work.
- The models outperform baseline SEQ2SEQ systems in BLEU, perplexity, and human judgments of speaker consistency.The reported gains are described as present but not spectacular.
- The Speaker-Addressee model provides evidence that capturing dyadic interactions is beneficial.
- The ultimate goal is to emulate an arbitrary individual’s persona from that person’s profile, including linguistic response behavior and other salient characteristics.