Personalizing Dialogue Agents: I have a dog, do you have pets too?

Saizheng ZhangEmily DinanJack UrbanekArthur SzlamDouwe KielaJason Weston

article2018ACL1,780 citations

Introduces a persona-conditioned framework and benchmark dataset that trains open-domain dialogue agents to exhibit consistent personalities and actively learn about their interlocutors.

Listen

Conversational artificial intelligence often struggles in open-ended social settings. Standard chat agents regularly suffer from vague answers, lack of long-term memory, and shifting personalities because they are trained on vast, uncurated internet datasets representing thousands of different voices. These weaknesses lead to impersonal and disengaging interactions, presenting a major barrier for deploying conversational agents in customer-facing and relationship-building roles.

The article evaluates whether providing conversational models with explicit, textual persona profiles improves dialogue consistency and engagement. The researchers also investigate whether agents can learn to discover information about their conversation partner through dialogue.

To conduct this evaluation, the researchers created a dedicated dataset of over 160,000 conversational turns across nearly 11,000 paired human chats. Each participant was assigned a predefined persona consisting of several short descriptive sentences, along with paraphrased versions designed to prevent simple word-matching shortcuts. The authors then tested and compared both generative systems and retrieval-based ranking systems, including memory-augmented neural networks, under conditions where models were given access to their own profile, their partner's profile, or no profile at all. In addition to automated performance metrics, human evaluators interacted directly with the models to rate fluency, consistency, and engagement.

The study produced several key findings. First, conditioning models on persistent persona profiles significantly improves response prediction and perceived conversational consistency. Second, ranking models that retrieve and score candidate responses substantially outperformed generative models in automated selection benchmarks, achieving correct response identification rates exceeding 50% when conditioned on persona data compared to under 13% for generative approaches. Third, human evaluation demonstrated that agents trained on the structured persona dataset scored significantly higher in fluency and engagement (averaging scores around 3.5 to 4.0 out of 5) compared to models trained on standard movie subtitles or social media data (which averaged roughly 1.8 to 2.2 for engagement). Finally, human conversational partners successfully recognized the persona traits of the models in up to 81% of test dialogues, confirming that the agents maintained identifiable characteristics.

These findings demonstrate that endowing conversational agents with explicit, interpretable personas addresses fundamental deficits in coherence and user engagement without requiring massive scale-ups in training data. For organizations deploying conversational systems, this structured approach reduces the risk of inconsistent responses and lowers development overhead by focusing on targeted profile conditioning rather than opaque, unguided text generation.

Teams developing open-ended dialogue systems should adopt structured persona framing and memory-augmented architectures when designing user-facing conversational agents. For operational deployment, organizations should prioritize retrieval and ranking architectures where consistency and accuracy are critical, while exploring hybrid models for novel text generation. Future work should integrate profile prediction with response generation so agents can dynamically build and remember user preferences over extended multi-session interactions.

The research is bounded by its crowdsourced experimental environment, where human interactions were relatively short (averaging six to eight turns) and based on artificial personas rather than organic, long-term personal histories. Additionally, human evaluation metrics showed moderate variance due to subjective rating differences among evaluators. Nevertheless, the underlying evidence strongly supports the conclusion that explicit profile conditioning provides a practical and effective framework for improving conversational AI consistency.

Meta Mila – Québec Artificial Intelligence Institute dialogue multiagent system Dialogue Models Pro

arXiv: 1801.07243

Paper: A Neural Conversational Model, Oriol Vinyals et al. (2015). This foundational work establishes the sequence-to-sequence neural conversational framework that the source paper directly builds upon and modifies with explicit persona conditioning.
Paper: A Diversity-Promoting Objective Function for Neural Conversation Models, Jiwei Li et al. (2016). This paper identifies the generic response issue in neural chit-chat and introduces diversity-promoting training objectives, framing the core dialogue limitations that the source addresses via persona conditioning.

Paper: Learning Personalized Agents from Human Feedback, Kaiqu Liang et al. (2026). This work extends persona-conditioned dialogue by introducing interactive human-in-the-loop feedback mechanisms to dynamically adapt and refine agent profiles during live conversations.
Paper: Generative Agents: Interactive Simulacra of Human Behavior, Joon Sung Park et al. (2023). This research generalizes static persona conditioning into persistent, reflective agent architectures capable of sustaining coherent personalities and social relationships in multi-agent environments.
Paper: CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society, Guohao Li et al. (2023). This paper applies persona and role conditioning to multi-agent collaborative frameworks, exploring how structured communicative roles enable autonomous task completion.
Paper: LaMDA: Language Models for Dialog Applications, R. Thoppilan et al. (2022). This study advances open-domain conversational agents by scaling dialogue models with fine-tuning for safety, factual grounding, and quality metrics like specificity and interestingness.

Cover for Personalizing Dialogue Agents: I have a dog, do you have pets too?

Abstract

Chit-chat models are known to have several problems: they lack specificity, do not display a consistent personality and are often not very captivating. In this work we present the task of making chit-chat more engaging by conditioning on profile information. We collect data and train models to (i) condition on their given profile information; and (ii) information about the person they are talking to, resulting in improved dialogues, as measured by next utterance prediction. Since (ii) is initially unknown our model is trained to engage its partner with personal topics, and we show the resulting dialogue can be used to predict profile information about the interlocutors.

Topics

dialogue multiagent system

Keywords

Dialogue Models Pro Persona Model arXiv Memory Personalizing Dialogue Agents Personas Dog

1 Introduction
2 Related Work
3 The persona-chat Dataset
3.1 Personas
3.2 Revised Personas
3.3 Persona Chat
3.4 Evaluation
4 Models
4.1 Baseline ranking models
4.2 Ranking Profile Memory Network
4.3 Key-Value Profile Memory Network
4.4 Seq2Seq
4.5 Generative Profile Memory Network
5 Experiments
5.1 Automated metrics
5.2 Human Evaluation
5.3 Profile Prediction
6 Conclusion & Discussion
References
A Next Utterance Prediction Additional Evaluation Metrics
B Example Dialogs between Humans and Models
C Human Evaluation Measures
D Profile Prediction

Knowls

Knowl 1 — The PERSONA-CHAT Dataset
definition
The PERSONA-CHAT dataset is a multi-turn chit-chat corpus curated to train and evaluate conversational models conditioned on explicit textual persona profiles. Data collection was performed in three stages:
1. Persona Collection: Crowdworkers on Amazon Mechanical Turk created $1{,}155$ distinct character profiles. Each profile contains at least $5$ natural, descriptive sentences (capped at a maximum of $15$ words per sentence) outlining background, personal traits, or hobbies (e.g., "I am a vegetarian", "I like swimming"). Profiles do not contain real personal data of the annotators. $100$ personas are held out for validation and $100$ for test.
2. Persona Revision: To prevent models from exploiting superficial token matching, a separate set of crowdworkers rewrote each of the $1{,}155$ profiles. Rewritten sentences describe related characteristics of the same persona via paraphrasing, generalization, or specialization (e.g., "I like basketball" revised to "I am a big fan of Michael Jordan"). Direct copying of non-stop words was prohibited via automated entry validation.
3. Dialogue Collection: Pairs of crowdworkers were each assigned a distinct persona from the original pool and instructed to converse naturally while adopting their assigned persona, asking and answering questions to get to know each other. Messages were turn-based with a limit of $15$ words per message and a target dialog length of $6$ to $8$ turns per speaker. Verbatim copying of persona sentences into chat messages was blocked by string-matching validation.
The dataset consists of $164{,}356$ utterances across $10{,}981$ dialogues, partitioned into a training set ( $133{,}532$ utterances / $9{,}013$ dialogues), a validation set ( $15{,}705$ utterances / $968$ dialogues), and a test set ( $15{,}119$ utterances / $1{,}000$ dialogues).
Knowl 2 — Ranking Profile Memory Network
model/method
The Ranking Profile Memory Network is a retrieval-based dialogue model that scores candidate responses by conditioning on both dialogue context and a set of persona profile sentences via memory attention.

Let $W \in \mathbb{R}^{D \times d}$ be an embedding matrix for a vocabulary of size $D$ and embedding dimension $d$ . The dialogue history query is represented as a bag-of-words vector $q = \sum_{w \in \text{context}} W_w$ . Given $M$ persona sentences, each profile sentence $p_i$ ( $i = 1, \dots, M$ ) is represented as the sum of its constituent word embeddings $p_i = \sum_{w \in p_i} W_w$ .

The network computes attention weights $s_i$ over the profile sentences using cosine similarity $\text{sim}(u, v) = \frac{u \cdot v}{\|u\|_2 \|v\|_2}$ :

$s_i = \text{Softmax}(\text{sim}(q, p_i)) = \frac{\exp(\text{sim}(q, p_i))}{\sum_{j=1}^M \exp(\text{sim}(q, p_j))}$

$q^+ = q + \sum_{i=1}^M s_i p_i$

Given a set of candidate responses $c'$ (each embedded as $c' = \sum_{w \in c'} W_w$ ), the model ranks each candidate according to $\text{sim}(q^+, c')$ . The parameters $W$ are optimized using a margin ranking loss over ground truth and $k$ negative candidate utterances sampled from the training set.
Knowl 3 — Key-Value Profile Memory Network for Dialogue
model/method
The Key-Value (KV) Profile Memory Network extends profile memory ranking by incorporating dialogue histories and their corresponding replies from the training set as an external key-value memory.

The model executes two sequential memory hops:
1. First Hop (Profile Memory): The query vector $q \in \mathbb{R}^d$ (sum of word embeddings of the dialogue history) attends over profile sentence embeddings $p_i$ :
$s_i^{(1)} = \text{Softmax}(\text{sim}(q, p_i)), \quad q^+ = q + \sum_i s_i^{(1)} p_i$

where $\text{sim}(u, v)$ is cosine similarity.
1. Second Hop (Key-Value Dialogue Memory): $q^+$ attends over external memory pairs $(k_j, v_j)$ , where each key $k_j$ is a dialogue history from the training corpus and each value $v_j$ is the corresponding next-utterance reply:
$s_j^{(2)} = \text{Softmax}(\text{sim}(q^+, k_j)), \quad q^{++} = q^+ + \sum_j s_j^{(2)} v_j$

Candidate replies $c'$ are scored and ranked according to $\text{sim}(q^{++}, c')$ . Because direct end-to-end training over large sets of key-value dialogue pairs is computationally intensive, the model can utilize the embedding weights learned by the Ranking Profile Memory Network at test time.
Knowl 4 — Generative Profile Memory Network
model/method
The Generative Profile Memory Network is an encoder-decoder model that incorporates individual persona profile sentences into the generative decoding process via memory-augmented attention.

The input dialogue history $x = (x_1, \dots, x_T)$ is encoded via an LSTM encoder $h_t^e = \text{LSTM}_{\text{enc}}(x_t \mid h_{t-1}^e)$ , with the final state $h_T^e$ initializing the decoder state $h_0^d$ . Each persona profile entry $p_i = (p_{i,1}, \dots, p_{i,n_i})$ is represented as a weighted bag-of-words vector:

$f(p_i) = \sum_{j=1}^{n_i} \alpha_{i,j} W_{p_{i,j}}, \quad \alpha_{i,j} = \frac{1}{1 + \log(1 + \text{tf}_{i,j})}$

where $W_{p_{i,j}}$ are word embeddings and $\text{tf}_{i,j}$ denotes term frequency approximated from the GloVe index $\text{idx}$ via Zipf's law: $\text{tf} = 10^6 \cdot \text{idx}^{-1.07}$ .

Let $F$ be the matrix containing the memory representations $\{f(p_i)\}$ . At decoder step $t$ , the decoder hidden state $h_t^d$ computes an attention distribution $a_t$ over memory representations:

$a_t = \text{softmax}(F W_a h_t^d)$

$c_t = a_t^\top F$

$\hat{x}_t = \tanh(W_c [c_{t-1}, x_t])$

where $W_a$ and $W_c$ are learnable projection matrices, $[\cdot, \cdot]$ denotes concatenation, and $\hat{x}_t$ is provided as input to the decoder LSTM step. The model is trained end-to-end using negative log-likelihood of target tokens.
Knowl 5 — StarSpace Supervised Embedding Baseline for Dialogue Ranking
model/method
The StarSpace model is a supervised embedding approach applied to dialogue next-utterance prediction. Both query context and candidate replies are embedded by summing word vectors from a shared lookup matrix $W \in \mathbb{R}^{D \times d}$ :

$q = \sum_{w \in \text{context}} W_w, \quad c' = \sum_{w \in \text{candidate}} W_w$

When persona profile sentences $P = \{p_1, \dots, p_M\}$ are provided, they are concatenated into the query bag-of-words representation:

$q = \sum_{w \in \text{context} \cup P} W_w$

The matching score is computed via cosine similarity:

$\text{sim}(q, c') = \frac{q \cdot c'}{\|q\|_2 \|c'\|_2}$

The parameters $W$ are optimized using a margin ranking loss with $k$ -negative candidate sampling:

$\mathcal{L} = \sum_{(q, c^+)} \sum_{i=1}^k \max\left(0, \gamma - \text{sim}(q, c^+) + \text{sim}(q, c_i^-)\right)$

where $c^+$ is the true dialogue response, $c_i^-$ are randomly sampled distractor candidates, and $\gamma > 0$ is a fixed margin.
Knowl 6 — Evaluation Protocols for Persona-Conditioned Dialogue
experimental setup
Dialogue next-utterance prediction on PERSONA-CHAT is evaluated across four persona conditioning configurations and two primary evaluation paradigms:
1. Conditioning Modes:
  - No Persona: Only the dialogue history is provided.
  - Self Persona: Conditioned on the speaker's own persona profile (using either Original or Revised sentences).
  - Their Persona: Conditioned on the conversation partner's profile.
  - Both: Conditioned simultaneously on both speakers' profiles.
2. Automated Metrics:
  - Perplexity (ppl): Exponentiated negative log-likelihood per word of the target utterance (generative models).
  - Hits@1: Accuracy of selecting the true next utterance from a candidate set comprising the ground truth response and $19$ random distractor responses drawn from other dialogues ( $N=19$ ).
  - F1 Score: Token-level overlap between predicted and ground-truth utterances.
3. Human Evaluation Metrics: Crowdworkers engage in multi-turn dialogues ( $6$ to $8$ turns per participant) with models or human baselines under assigned test personas. Post-chat evaluations collect:
  - Fluency (1–5 scale): Grammatical correctness and naturalness.
  - Engagingness (1–5 scale): How captivating and interactive the partner is.
  - Consistency (1–5 scale): Coherence and self-consistency of persona/statements.
  - Persona Detection (accuracy $\in [0, 1]$ ): Turkers select which of two displayed persona profiles corresponds to their conversation partner.

Knowl 7 — Automated Next Utterance Prediction Performance on PERSONA-CHAT

data/table

The next utterance prediction performance of generative and ranking models across persona conditioning settings on PERSONA-CHAT is reported below:

Method	No Persona		Original Persona		Revised Persona
	ppl	hits@1	ppl	hits@1	ppl	hits@1
Generative Models
Seq2Seq	38.08	0.092	40.53	0.084	40.65	0.082
Profile Memory	38.08	0.092	34.54	0.125	38.21	0.108
Ranking Models
IR baseline	-	0.214	-	0.410	-	0.207
Starspace	-	0.318	-	0.491	-	0.322
Profile Memory	-	0.318	-	0.509	-	0.354
KV Profile Memory	-	0.349	-	0.511	-	0.351

Ranking models achieve substantially higher hits@1 than generative models because they directly optimize candidate discrimination. Conditioning on Original Persona yields large improvements across all ranking models and the generative Profile Memory model (reducing ppl from 38.08 to 34.54 and increasing hits@1 from 0.092 to 0.125). On Revised Personas (which remove lexical overlap), performance gains over the No Persona baseline remain significant for Profile Memory networks (0.354 hits@1 vs. 0.318), demonstrating that the model learns semantic alignment rather than simple keyword matching.

Knowl 8 — Human Evaluation of Persona-Conditioned and Baseline Dialogue Models

data/table

Human evaluation of conversational quality (Fluency, Engagingness, Consistency on a 1–5 scale) and Persona Detection accuracy ( $100$ dialogues per model, test personas) is reported below:

Model	Profile	Fluency	Engagingness	Consistency	Persona Detection
Human	Self	4.31 (1.07)	4.25 (1.06)	4.36 (0.92)	0.95 (0.22)
Generative PersonaChat Models
Seq2Seq	None	3.17 (1.10)	3.18 (1.41)	2.98 (1.45)	0.51 (0.50)
Profile Memory	Self	3.08 (1.40)	3.13 (1.39)	3.14 (1.26)	0.72 (0.45)
Ranking PersonaChat Models
KV Memory	None	3.81 (1.14)	3.88 (0.98)	3.36 (1.37)	0.59 (0.49)
KV Profile Memory	Self	3.97 (0.94)	3.50 (1.17)	3.44 (1.30)	0.81 (0.39)
External Baselines
Twitter LM	None	3.21 (1.54)	1.75 (1.04)	1.95 (1.22)	0.57 (0.50)
OpenSubtitles 2018 LM	None	2.85 (1.46)	2.13 (1.07)	2.15 (1.08)	0.35 (0.48)
OpenSubtitles 2009 LM	None	2.25 (1.37)	2.12 (1.33)	1.96 (1.22)	0.38 (0.49)
OpenSubtitles 2009 KV Memory	None	2.14 (1.20)	2.22 (1.22)	2.06 (1.29)	0.42 (0.49)

Models trained on PERSONA-CHAT exhibit markedly higher engagingness ( $3.13$ -- $3.88$ ) and consistency ( $2.98$ -- $3.44$ ) than models trained on OpenSubtitles or Twitter ( $1.75$ -- $2.22$ engagingness; $1.95$ -- $2.15$ consistency). Furthermore, profile conditioning enables human evaluators to detect the model's persona profile with significantly higher accuracy ( $0.72$ for generative Profile Memory vs. $0.51$ for Seq2Seq; $0.81$ for KV Profile Memory vs. $0.59$ for KV Memory), approaching the human performance upper bound of $0.95$ .

Knowl 9 — Interlocutor Profile Prediction from Dialogue History
empirical result
In addition to next utterance prediction, PERSONA-CHAT can be utilized to predict an interlocutor's persona profile from the generated dialogue transcript.

Automated profile extraction experiments demonstrate:
1. A human speaker's profile can be predicted from their own dialogue utterances with $94.3\%$ accuracy, matching human rater detection capability ( $95\%$ accuracy).
2. An interlocutor's profile can be extracted from the model's dialogue responses (achieving $23\%$ accuracy when conversing with KV Profile Memory), indicating that profile-conditioned models actively elicit and attend to their dialogue partner's topics.
3. Profile prediction accuracy monotonically increases as conversation length progresses across successive turns.

Coverage note — None was omitted; all primary contributions, model architectures, dataset design stages, automated metrics, human evaluations, and profile prediction experiments from the paper are represented.

References

1.Antoine Bordes and Jason Weston. 2016. Learning end-to-end goal-oriented dialog. arXiv preprint arXiv:1605.07683.
2.Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading wikipedia to answer open-domain questions. arXiv preprint arXiv:1704.00051.
3.Jesse Dodge, Andreea Gane, Xiang Zhang, Antoine Bordes, Sumit Chopra, Alexander Miller, Arthur Szlam, and Jason Weston. 2015. Evaluating prerequisite qualities for learning end-to-end dialog systems. arXiv preprint arXiv:1511.06931.
4.Robin IM Dunbar, Anna Marriott, and Neil DC Duncan. 1997. Human conversational behavior. Human nature, 8(3):231–246.
5.Chaitanya K Joshi, Fei Mi, and Boi Faltings. 2017. Personalization in goal-oriented dialog. arXiv preprint arXiv:1706.07503.
6.Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2015. A diversity-promoting objective function for neural conversation models. arXiv preprint arXiv:1510.03055.
7.Jiwei Li, Michel Galley, Chris Brockett, Georgios P Spithourakis, Jianfeng Gao, and Bill Dolan. 2016a. A persona-based neural conversation model. arXiv preprint arXiv:1603.06155.
8.Jiwei Li, Will Monroe, Alan Ritter, Michel Galley, Jianfeng Gao, and Dan Jurafsky. 2016b. Deep reinforcement learning for dialogue generation. arXiv preprint arXiv:1606.01541.
9.Chia-Wei Liu, Ryan Lowe, Iulian V Serban, Michael Noseworthy, Laurent Charlin, and Joelle Pineau. 2016. How not to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for dialogue response generation. arXiv preprint arXiv:1603.08023.
10.Ryan Lowe, Nissan Pow, Iulian Serban, and Joelle Pineau. 2015. The ubuntu dialogue corpus: A large dataset for research in unstructured multi-turn dialogue systems. arXiv preprint arXiv:1506.08909.
11.JM Lucas, F Fernández, J Salazar, J Ferreiros, and R San Segundo. 2009. Managing speaker identity and user profiles in a spoken dialogue system. Procesamiento del Lenguaje Natural, 43:77–84.
12.Alexander Miller, Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes, and Jason Weston. 2016. Key-value memory networks for directly reading documents. arXiv preprint arXiv:1606.03126.
13.Mor Naaman, Jeffrey Boase, and Chih-Hui Lai. 2010. Is it really about me?: message content in social awareness streams. In Proceedings of the 2010 ACM conference on Computer supported cooperative work, pages 189–192. ACM.
14.Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532–1543.
15.Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad: 100,000+ questions for machine comprehension of text. arXiv preprint arXiv:1606.05250.
16.Iulian V Serban, Chinnadhurai Sankar, Mathieu Germain, Saizheng Zhang, Zhouhan Lin, Sandeep Subramanian, Taesup Kim, Michael Pieper, Sarath Chandar, Nan Rosemary Ke, et al. 2017a. A deep reinforcement learning chatbot. arXiv preprint arXiv:1709.02349.
17.Iulian Vlad Serban, Ryan Lowe, Laurent Charlin, and Joelle Pineau. 2015. A survey of available corpora for building data-driven dialogue systems. arXiv preprint arXiv:1512.05742.
18.Iulian Vlad Serban, Ryan Lowe, Laurent Charlin, and Joelle Pineau. 2016. Generative deep neural networks for dialogue: A short review. arXiv preprint arXiv:1611.06216.
19.Iulian Vlad Serban, Alessandro Sordoni, Ryan Lowe, Laurent Charlin, Joelle Pineau, Aaron C Courville, and Yoshua Bengio. 2017b. A hierarchical latent variable encoder-decoder model for generating dialogues.
20.Alessandro Sordoni, Michel Galley, Michael Auli, Chris Brockett, Yangfeng Ji, Margaret Mitchell, Jian-Yun Nie, Jianfeng Gao, and Bill Dolan. 2015. A neural network approach to context-sensitive generation of conversational responses. arXiv preprint arXiv:1506.06714.
21.Sainbayar Sukhbaatar, Jason Weston, Rob Fergus, et al. 2015. End-to-end memory networks. In Advances in neural information processing systems, pages 2440–2448.
22.Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In Advances in neural information processing systems, pages 3104–3112.
23.Oriol Vinyals and Quoc Le. 2015. A neural conversational model. arXiv preprint arXiv:1506.05869.
24.Ellen M Voorhees et al. 1999. The trec-8 question answering track report. In Trec, volume 99, pages 77–82.
25.Joseph Weizenbaum. 1966. Elizaa computer program for the study of natural language communication between man and machine. Communications of the ACM, 9(1):36–45.
26.Ledell Wu, Adam Fisch, Sumit Chopra, Keith Adams, Antoine Bordes, and Jason Weston. 2017. Starspace: Embed all the things! arXiv preprint arXiv:1709.03856.
27.Steve Young, Milica Gašić, Blaise Thomson, and Jason D Williams. 2013. Pomdp-based statistical spoken dialog systems: A review. Proceedings of the IEEE, 101(5):1160–1179.
28.Steve J Young. 2000. Probabilistic methods in spoken–dialogue systems. Philosophical Transactions of the Royal Society of London A: Mathematical, Physical and Engineering Sciences, 358(1769):1389–1402.

Citation

MLA

Zhang, S., et al. “Personalizing Dialogue Agents: I Have a Dog, Do You Have Pets Too?”. arXiv, 2018, http://arxiv.org/abs/1801.07243v5.

APA

Zhang, S., Dinan, E., Urbanek, J., Szlam, A., Kiela, D., & Weston, J. (2018). Personalizing Dialogue Agents: I have a dog, do you have pets too?. arXiv. http://arxiv.org/abs/1801.07243v5

Chicago

Zhang, S., E. Dinan, J. Urbanek, A. Szlam, D. Kiela, and J. Weston. 2018. “Personalizing Dialogue Agents: I Have a Dog, Do You Have Pets Too?”. arXiv. http://arxiv.org/abs/1801.07243v5.

Harvard

Zhang, S. et al. (2018) “Personalizing Dialogue Agents: I have a dog, do you have pets too?”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1801.07243v5.

Vancouver

1. Zhang S, Dinan E, Urbanek J, Szlam A, Kiela D, Weston J (2018) Personalizing Dialogue Agents: I have a dog, do you have pets too?. arXiv

BibTeX

@article{zhang2018personalizing,
  title = {Personalizing Dialogue Agents: I have a dog, do you have pets too?},
  author = {Zhang, Saizheng and Dinan, Emily and Urbanek, Jack and Szlam, Arthur and Kiela, Douwe and Weston, Jason},
  year = {2018},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1801.07243v5},
  eprint = {1801.07243}
}

Metadata:arXiv

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: https://creativecommons.org/licenses/by/4.0/

Personalizing Dialogue Agents: I have a dog, do you have pets too?

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — The PERSONA-CHAT Dataset

Knowl 2 — Ranking Profile Memory Network

Knowl 3 — Key-Value Profile Memory Network for Dialogue

Knowl 4 — Generative Profile Memory Network

Knowl 5 — StarSpace Supervised Embedding Baseline for Dialogue Ranking

Knowl 6 — Evaluation Protocols for Persona-Conditioned Dialogue