Source-linked AI summary
Long Time No See! Open-Domain Conversation with Long-Term Persona Memory
Xinchao Xu, Zhibin Gou, Wenquan Wu, Zheng-Yu Niu, Hua Wu, Haifeng Wang, Shihang Wang
TL;DR
Open-domain dialogue models struggle with long-term human-bot conversations because they lack effective understanding and memory of long-term dialogue history. The paper introduces the LeMon task, DuLeMon dataset, and PLATO-LTM, which dynamically manages both parties’ persona memories. On DuLeMon, PLATO-LTM improves long-term dialogue consistency, engagingness, and coherence, although small-data fine-tuning can hurt coherence.
Problem
Long-term persona ability is less studied because existing systems lack long-term dialogue-history memory, while existing datasets often ignore or cannot dynamically update the user’s persona.
Method
The paper builds DuLeMon for mutual-persona long-term conversation and adds PLATO-LTM, which extracts, stores, updates, and retrieves user and chatbot persona information in real time.
Results
PLATO-LTM effectively uses both parties’ persona information from dialogue history to enhance long-term dialogue consistency and engagingness.
Takeaways & Limitations
Dynamic memory of both user and chatbot personas supports more consistent and engaging long-term conversations without training on long-session data.
Takeaways & Limitations
Fine-tuning on the small-scale dataset can hurt dialogue coherence, and long-session retrieval-based approaches require expensive, difficult-to-annotate training data.
Abstract
from arXiv · showhide
Most of the open-domain dialogue models tend to perform poorly in the setting of long-term human-bot conversations. The possible reason is that they lack the capability of understanding and memorizing long-term dialogue history information. To address this issue, we present a novel task of Long-term Memory Conversation (LeMon) and then build a new dialogue dataset DuLeMon and a dialogue generation framework with Long-Term Memory (LTM) mechanism (called PLATO-LTM). This LTM mechanism enables our system to accurately extract and continuously update long-term persona memory without requiring multiple-session dialogue datasets for model training. To our knowledge, this is the first attempt to conduct real-time dynamic management of persona information of both parties, including the user and the bot. Results on DuLeMon indicate that PLATO-LTM can significantly outperform baselines in terms of long-term dialogue consistency, leading to better dialogue engagingness.
1 Introduction
Existing open-domain dialogue systems struggle to sustain long-term connections because they lack long-term dialogue-history understanding and memory. The paper introduces LeMon, DuLeMon, and PLATO-LTM to model and use both speakers’ evolving persona information.
- Motivation: Long-term persona ability is underexplored because existing systems lack the capability to understand and memorize long-term dialogue history.Without this ability, systems tend to talk to people like strangers in long-term conversations.
- Limitations of prior work: Existing persona datasets focus on chatbot-persona consistency, ignore user-persona memory, and typically use fixed persona information.MSC adds multi-session summaries, but its stored documents are not dynamically modified and grow as conversations continue.
- Task and dataset: LeMon defines long-term mutual-persona conversation, while DuLeMon annotates both speakers’ grounding personas and assumes prior interaction between them.The dataset focuses on both chatbot-persona consistency and active construction and use of the user’s persona.
- Framework: PLATO-LTM adds a plug-and-play long-term memory mechanism that extracts, stores, updates, and retrieves both parties’ persona information in real time.Its Persona Extractor filters irrelevant information, separate memories store each interlocutor’s persona, and retrieved persona sentences are concatenated with dialogue context for generation.
- Task and dataset: DuLeMon is presented as the largest available multi-turn Chinese mutual-persona chat dataset and supports the LeMon task.The task and dataset are designed for Chinese long-term conversations.
- Results: Automatic and human evaluations show that PLATO-LTM significantly improves long-conversation consistency while making responses more engaging and coherent.The framework studies long-term persona conversations without relying on long-session dialogue data for training.
2 Related Work
Prior persona-dialogue research includes implicit and explicit persona models, while dialogue systems have also used external memory to retain user-related information. The paper positions DuLeMon as requiring active memory and use of the user’s persona in long-term interaction.
- Persona dialogue: Persona-dialogue approaches include implicit semantic persona representations and explicit persona models.Implicit representations are less easy to interpret and control during target-response generation.
- External memory: Earlier external-memory dialogue systems used rule-based memory mechanisms to integrate persona or user interests and rewrite responses.These approaches stored or retrieved user-related information for later response generation.
- Paper positioning: DuLeMon requires the chatbot to actively remember and use the user’s persona rather than treating the speakers as strangers exchanging information.This setting is intended to improve conversational engagement and intimacy in long-term interactions.
3 Data Collection
DuLeMon is designed for conversations where the chatbot tracks persona information for both participants and links responses to the persona information they use. The dataset combines persona-aware dialogue collection with explicit persona-grounding annotations and distinct visibility settings.
- Task definition: The task predicts the chatbot response from dialogue context together with user and chatbot persona descriptions.The formalization defines separate persona sets for the user and chatbot and seeks the corresponding persona before predicting the response.
- Dataset design: DuLeMon represents both user and chatbot personas and identifies the persona associated with each response.The dataset exposes both parties’ persona information during conversation and explicitly labels response-associated persona information.
- Persona collection: DuLeMon includes user persona information already known to the chatbot and information learned through historical conversations.The collection procedure separates user persona into known and initially unknown portions.
- Dialogue collection: Crowd-workers were paired to role-play users and chatbots using assigned personas, while keeping dialogue relevant to and consistent with those personas.The chatbot was instructed to use known user persona information for deeper conversation, and the dialogue could not conflict with the given persona.
- Persona grounding: Persona-grounding labeling records whether each response uses persona information and whether the grounding persona comes from the chatbot or user.This annotation establishes the direct relationship between responses and the persona information used.
- Dataset variants: The dataset contains DuLeMon-SELF, where the bot knows only its own persona, and DuLeMon-BOTH, where it also knows part of the user’s persona.These two settings vary the visibility of user persona information to the chatbot.
4 Model Architecture
PLATO-LTM combines persona extraction, dynamic long-term memory, retrieval, and response generation to manage user and chatbot persona information. Its architecture writes extracted personas into memory, retrieves context-relevant entries, and uses role-aware generation to produce responses.
- Framework overview: PLATO-LTM has three core parts: a persona extractor, long-term persona memory, and a generation module.The framework stores, updates, and reads user and chatbot persona information through explicit memory operations.
- Persona extraction: The persona extractor classifies utterances or text spans for persona information and collects positively classified clauses as persona sentences.It uses a supervised ERNIE-CNN architecture, with clause segmentation performed during inference.
- Memory architecture: The long-term memory stores user and chatbot persona histories, while context-persona matching controls memory reads and writes.Context and persona encoders represent the current context and persona entries for matching.
- Memory matching: Memory matching is trained with triplet loss using personas relevant to the current utterance and response as positives and remaining session personas as negatives.The context and persona encoders are initialized with ERNIE and trained on DuLeMon.
- Memory writing: Writing removes duplicates by replacing highly similar stored personas or adding new personas when similarity remains below the duplication threshold.Each stored persona is paired with its encoded representation.
- Memory reading: Reading retrieves candidate personas by dense-vector similarity, reranks them for context relevance, and filters low-similarity entries before generation.The system uses the top k candidates from both user and chatbot memories, subject to threshold sc.
- Response generation: The generator is built on PLATO-2’s first-stage architecture and models responses with autoregressive negative log-likelihood training.Response tokens attend only to preceding generated tokens, while context uses bidirectional attention.
- Role disambiguation: Role embeddings and role tokens distinguish user and chatbot personas to reduce confusing persona usage during generation.The strategies mark persona roles with distinct embeddings and textual prefixes such as “system persona” and “user persona.”
5 Experiments
Experiments evaluate persona extraction, memory retrieval, generation, and self-chat performance. PLATO-LTM improves long-term dialogue consistency and engagingness, while small-dataset fine-tuning can reduce coherence.
- Experimental setup: The experiments compare PLATO-2, PLATO-FT, PLATO-LTM, and PLATO-LTM w/o PE across automatic and human evaluation settings.The evaluation includes persona classification, memory ranking, response generation, and self-chat metrics.
- Persona extractor: The persona extractor’s F1 exceeds 0.9, indicating effective recognition and storage of persona information from dialogue history.The two-stage pc-stage2 model outperforms pc-stage1 and is adopted in the system.
- Long-term memory retrieval: The memory-ranking model achieves AUC 0.76 and recall@5 0.83 on the automatic test set.These results indicate that the model retrieves relevant persona information from long-term memory.
- Generative model: The 32L generative model lowers PPL by 4.4 and increases F1 by 2.5 relative to the 12L model.PLATO-FT 32L with role embedding and role token strategies is selected for the final system.
- Human evaluation: PLATO-FT scores 1.59 for dialogue coherence, below the non-fine-tuned PLATO baseline, suggesting small-dataset fine-tuning can hurt coherence.The authors attribute this to broad open-domain topics in self-play conversations and reduced response appropriateness.
6 Conclusion
The paper introduces LeMon and DuLeMon for long-term persona modeling and adds a plug-in Long-Term Memory component to large-scale dialogue models. PLATO-LTM uses both parties’ persona information to enhance dialogue consistency and engagingness.
- Contributions: LeMon defines a long-term persona conversation task, and DuLeMon provides its corresponding dialogue dataset.The task brings long-term persona modeling into large-scale generative dialogue models.
- Long-Term Memory: LTM maintains separate user and chatbot memories that extract, remember, and continuously update persona information from dialogue history.The user memory models information mentioned by the user, while chatbot memory updates the chatbot’s persona over time.
- Results: PLATO-LTM effectively uses both parties’ persona information to improve dialogue consistency and engagingness.
7 Ethical Considerations
The authors state that DuLeMon collection followed source-use, intellectual-property, and privacy requirements and received institutional review board approval. They also describe steps intended to reduce dataset-quality risks.
- Compliance: The DuLeMon collection is described as consistent with source terms of use and the intellectual-property and privacy rights of original authors.
- Oversight: The project received IRB approval and documents dataset characteristics and quality safeguards intended to reduce potential risks.
A Details of Data Collection
DuLeMon is collected through paired crowdworker conversations constrained by persona consistency, minimum dialogue length, and natural interaction guidelines. Workers are encouraged to elaborate on persona information rather than copy it verbatim.
- Collection process: Paired crowdworkers enter the chat interface, with role 1 initiating the conversation.
- Dialogue constraints: Chat content includes greetings, self-introductions, persona-consistent exchanges, questions, and answers using the assigned personal information.
- Dialogue constraints: Each dialogue contains at least 8 turns, with each participant producing at least 8 utterances.
- Quality guidelines: Workers are instructed to make conversations natural and varied, avoid repetition, and avoid copying persona sentences directly.Persona information should be expressed richly and discussed through related content rather than reproduced verbatim.
B Details of Models
The generation model follows PLATO-2 and uses fixed context, persona, vocabulary, optimization, and hardware settings. Long-term memory is configured for both user and chatbot memory.
- The generation model follows PLATO-2 as its base architecture.
- Context, user persona, and chatbot persona lengths are set to 384, 76, and 52, respectively.
- The vocabulary contains 30K Chinese BPE tokens, and optimization uses Adam with batches of B = 16384 tokens and learning rate lr = 5e −5.
- Experiments run on NVIDIA V100 32GB and A100 48GB GPUs.
- Long-term memory is configured for both user memory and chatbot memory.
C Cases of PLATO-LTM
The paper presents a cherry-picked episode conversation between PLATO-LTM and PLATO-2 to demonstrate long-term persona ability. The exchange revisits personal details across topics including writing, education, relationships, work, and investments.
- Figure 4 presents a cherry-picked example of one episode conversation between PLATO-LTM and PLATO-2.
- The conversation discusses the speaker’s campus novel, university year, major, relationship status, and institution.
- The dialogue also covers funds, an equity-fund purchase of $10,000, and questions about the progress and content of a novel.
- Later turns mention a shift from sales to administration, reading a self-written novel, a literature major, and clerical work.