CoChat
Sign Up
Cited by 815Speech and dialogue systemsAI in Service InteractionsSocial Robot Interaction and HRI

The DOI matches its Crossref record; 73 of the 73 references it cites were found in a registry or an index; none of the 27 sources its claims cite is checked for support yet.

The Design and Implementation of XiaoIce, an Empathetic Social Chatbot

Open access
Zhou, Li, Gao, Jianfeng, Li, Di, Shum, Heung-Yeung·2020·Computational Linguistics
DOI 10.1162/coli_a_00368
Open this paper in CoChat — free account

Abstract

This article describes the development of Microsoft XiaoIce, the most popular social chatbot in the world. XiaoIce is uniquely designed as an artifical intelligence companion with an emotional connection to satisfy the human need for communication, affection, and social belonging. We take into account both intelligent quotient and emotional quotient in system design, cast human–machine social chat as decision-making over Markov Decision Processes, and optimize XiaoIce for long-term user engagement, measured in expected Conversation-turns Per Session (CPS). We detail the system architecture and key components, including dialogue manager, core chat, skills, and an empathetic computing module. We show how XiaoIce dynamically recognizes human feelings and states, understands user intent, and responds to user needs throughout long conversations. Since the release in 2014, XiaoIce has communicated with over 660 million active users and succeeded in establishing long-term relationships with many of them. Analysis of large-scale online logs shows that XiaoIce has achieved an average CPS of 23, which is significantly higher than that of other chatbots and even human conversations.

Overview

AI snapshot

Written from the paper’s full text.

TL;DR

This paper describes Microsoft XiaoIce, an empathetic social chatbot designed as an AI companion with IQ, EQ, and personality, optimized for long-term user engagement measured by expected Conversation-turns Per Session (CPS). It reports system architecture, dialogue manager, empathetic computing, Core Chat, skills, and large-scale deployment results, including 660 million active users and an average CPS of 23.

Key findings

  • XiaoIce was launched in May 2014 and by July 2018 had over 660 million active users, was deployed in five countries on more than 40 platforms, and achieved an average CPS of 23 in its 6th generation; the paper compares this with CPS 9 for human conversations in its user study and CPS 14.6 for the latest Amazon Alexa systems per Khatri et al. [2018].
  • The empathetic computing module’s A/B test on Weibo showed no significant CPS change but increased NAU from 0.5 million to 5.1 million in 3 months; it was released in July 2018 and became the most important feature in the 6th generation.
  • Core Chat’s hybrid approach improved expected CPS by 0.5 in two weeks over a retrieval-only baseline; incorporating the neural generator and unpaired-data retrieval improved response coverage by 20% and 10%, respectively. Topic Manager improved Core Chat expected CPS by 0.5, and Image Commenting doubled expected CPS across all dialogues containing images.
  • The persona-based neural response generator achieved perplexity 25.0 (−8.4% vs. seq2seq) and BLEU 1.9% (+18.8% vs. seq2seq, +11.8% vs. LSTM-MMI) on the TV series dataset; the XiaoIce Image Commenting prototype scored BLEU-4 4.53, ROUGE-L 15.33, CIDEr-D 3.21, METEOR 15.51, and SPICE 4.82, outperforming four state-of-the-art image captioning systems.

Methods

  • Design: XiaoIce is a modular system with user experience, conversation engine, and data layers; social chat is cast as hierarchical decision-making over options in Markov Decision Processes, with a dialogue manager, empathetic computing module, Core Chat, and 230 dialogue skills, optimized for expected CPS.
  • Data and training: paired conversational data included more than 30 billion conversation pairs as of May 2018, with 70% of XiaoIce’s responses retrieved from her own past conversations; classifiers for user understanding were trained on 10K manually labeled dialogue sessions, the response ranker on 50K manually labeled dialogues, neural generation and ranking features on 50 million human dialogues, and Image Commenting on 28 million images each paired with 6 text comments.
  • Evaluation: A/B tests used expected CPS and NAU; pilot studies used human ratings on a 3-level quality scale and benchmark metrics such as perplexity, BLEU, METEOR, CIDEr, ROUGE-L, and SPICE; large-scale online logs were analyzed over long periods.

Limitations

  • The paper states that automatic evaluation of open-domain dialogue systems is “by no means a solved problem,” that common automatic metrics have limitations, and that a unified modeling framework for XiaoIce “remains to be proved” effective.
  • The paper states that fully grounding all conversations in the physical world for goal-oriented interactions remains a non-trivial challenge, and that the fundamental mechanism of human-level intelligence is not yet fully understood.
  • The paper discusses ethical concerns including privacy, who is in control, unrealistic user expectations, and potential misuse; evident from the text, many key results come from proprietary A/B tests and online logs without reported confidence intervals or full statistical details, and the comparison to human conversations relies on an unspecified user study.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1162/coli_a_00368

Cites

73 of 73 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

0 of 27 cited sources checked · 0 supported · 0 flagged

Show all 73 claims

  • Probably not a substantial claim. 66
  • It may state two separate claims as one. 43
  • The quote only partly states this claim. 2
89 of 89

  • The authors state this as an assumption.

    Human-machine social chat can be modeled as a hierarchical decision-making process over Markov Decision Processes.

    “we mathematically cast human-machine social chat as a hierarchical decision-making process, and optimize XiaoIce for long-term user engagement, measured in expected CPS.”
    “cast human-machine social chat as decision-making over Markov Decision Processes (MDPs)”

  • The authors state this as an assumption.

    Expected CPS is a valid measure of long-term user engagement for social chatbots.

    “we propose to use expected Conversation-turns Per Session (CPS) as the success metric for social chatbots.”
    “The larger the CPS is, the better engaged the social chatbot is.”

  • The paper relies on this without support.

    An AI companion can satisfy the human need for communication, affection, and social belonging.

    “XiaoIce is uniquely designed as an AI companion with an emotional connection to satisfy the human need for communication, affection, and social belonging.”

  • The paper relies on this without support.

    The machine learning classifiers used for user understanding accurately detect user intents, emotions, topics, opinions, and persona from dialogue.

    “These key-value pairs are generated using a set of machine learning classifiers as follows.”

  • The paper relies on this without support.

    The TV series dataset used in the first pilot study is representative enough to validate the persona-based neural response generator for XiaoIce.

    “In the first pilot study reported in Li et al. [2016b], we compare the persona model against two baseline models, using the TV series dataset for model training and evaluation.”

  • The paper relies on this without support.

    The 3-level quality scale labels assigned by human judges reliably measure response quality.

    “Each generated response is labeled on the 3-level quality scale by three human judges.”

  • The paper relies on this without support.

    The KG triples that co-occur often in the same conversations represent related topics humans often discuss.

    “We include in the XiaoIce KG a Satori triple ( h, r, t ) only if h and t co-occur often enough in the same conversations. Such a triple contains a pair of related topics ( h, t ) that humans often discuss in one conversation”

  • The paper relies on this without support.

    The unpaired database sentences from public lectures and quotes are suitable as candidate responses for XiaoIce.

    “The unpaired database we have used in XiaoIce consists of sentences collected from public lectures and quotes in news articles and reports. These sentences are considered candidate responses R .”

  • The paper relies on this without support.

    Any individual dialogue skill is unlikely to have a significant impact on CPS.

    “Since any individual skill is unlikely to have a significant impact on CPS, we measure the user satisfaction of a skill by monitoring its active users and skill triggering rate (i.e., the number of times a skill is activated by users within a day or a week).”

  • The paper relies on this without support.

    Traditionally, social chatbots are designed for chitchat scenarios where bots are expected to mimic human user conversations but not to interact with the user's environment.

    “Traditionally, social chatbots are designed for chitchat scenarios where the bots are expected to mimic human user conversations but not to interact with the user's environment.”

  • The paper relies on this without support.

    Happiness and meaningful conversations often go hand in hand.

    “Psychological studies show that happiness and meaningful conversations often go hand in hand.”

  • The paper relies on this without support.

    The long-term success of a social chatbot needs to be measured by its user engagement.

    “the long-term success of a social chatbot needs to be measured by its user engagement”

  • The paper relies on this without support.

    The most reliable evaluation is to deploy the chatbot to users and monitor user feedback and engagement over a long period.

    “There is no doubt that the most reliable evaluation is to deploy the chatbot to users and monitor the user feedback and engagement, measured by user ratings, NAU, CPS, and so on, over a long period of time.”

  • The paper relies on this without support.

    In open-domain dialog systems, the same input may have many plausible responses that differ in topics or contents significantly.

    “Nevertheless, in open-domain dialog systems, the same input may have many plausible responses that differ in topics or contents significantly.”

  • The paper relies on this without support.

    The number of reference responses in the test set is always limited.

    “as the number of reference responses is always limited in the test set”
15 of 15

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Pilot-study dataset is reported as both 4K and 5K sessions

    Minor

    The description of the second pilot study says its evaluation data consists of 4K dialogue sessions, but the caption of the resulting Table 3 calls it a 5K dialogue dataset.

  • Image Commenting generator direction is reversed

    Minor

    The Image Commenting section describes a generator that produces text from images, but its evaluation calls the trained component a text-to-image generator.

2 of 2

Citations

Cites

Flagged first, then by citations

Research works printed but not linked: 23

  1. Albrecht Joshua, 2007, Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics, 880
  2. Asimov I., 1984, “The Bicentennial Man” in I. Asimov, The Bicentennial Man and Other Stories
  3. Banerjee Satanjeev, 2005, Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, 65
  4. Brahnam Sheryl, 2005, Abuse: The Darker Side of Human Computer Interaction, 62
  5. Cuayáhuitl Heriberto, 2018, NeurIPS Workshop on Conversational AI
  6. Dinan Emily, 2018, CoRR
  7. Fang Hao, 2017, Alexa Prize Proceedings
  8. Fung Pascale, 2016, CoRR
  9. Gan Chuang, 2017, Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 3137
  10. Ghazvininejad Marjan, 2018, Proceedings of AAAI, 5111
  11. Lin Chin-Yew, 2004, Proceedings of the ACL workshop, 74
  12. Mathews Alexander Patrick, 2016, AAAI, 3574
  13. Misu Teruhisa, 2012, Proceedings of the 13th Annual Meeting of the Special Interest Group on Discourse and Dialogue, 84
  14. Mostafazadeh Nasrin, 2017, Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 462
  15. Papineni Kishore, 2002, Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, 311
  16. Sai Ananya, 2019, Proceedings of AAAI 2019, 33, 7281
  17. Serban Iulian Vlad, 2016, AAAI, 16, 3776
  18. Serban Iulian Vlad, 2017, AAAI, 3295
  19. Shum Heung-Yeung, 2018, CoRR
  20. Sutskever Ilya, 2014, NIPS, 3104
  21. Vinyals Oriol, 2015, ICML Deep Learning Workshop
  22. Wu Bowen, 2016, Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, 652
  23. Xing Chen, 2017, AAAI, 17, 3351

Cited by

Sorted by citations

See all 815

Related

See all

Metadata

Authors
Zhou, L., Gao, J., Li, D., & Shum, H.-Y.
Affiliations
Microsoft (United States) (US)
ORCID
Venue
Computational Linguistics
Publisher
MIT Press
Published
2020-03
Volume
46
Issue
1
Pages
53-93
ISSN
  • 0891-2017 (ISSN-L)
  • 1530-9312
DOI
10.1162/coli_a_00368
OpenAlex
W2906579211
Semantic Scholar
66b7d31527f980bb2eecc23629f08ba6037facc4
Cited by
777
References
66
Citations by year
2026: 97 · 2025: 109 · 2024: 106 · 2023: 113 · 2022: 120 · 2021: 112 · 2020: 94
FWCI
43.11
Type
journal article
Language
en
Open access
Licence
cc-by-nc-nd
Locations
  • Computational Linguistics (publishedVersion, cc-by-nc-nd)
  • arXiv (Cornell University) (submittedVersion)
  • Rare & Special e-Zone (The Hong Kong University of Science and Technology) (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • arXiv (Cornell University)
Funders
  • Microsoft Research
  • Microsoft Research Asia
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Speech and dialogue systems
  • Social Sciences › Business, Management and Accounting › Marketing › AI in Service Interactions
  • Social Sciences › Psychology › Social Psychology › Social Robot Interaction and HRI
Keywords
social chatbots, Markov decision process, dialogue management, user engagement, emotion recognition, user intent understanding

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-08, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-08 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review