ar
X
iv
:1
80
1.
07
24
3v
5
[
cs
.A
I]
2
5
S
ep
2
01
8
Personalizing Dialogue Agents: I have a dog, do you have pets too?
Saizheng Zhang†,1, Emily Dinan‡, Jack Urbanek‡, Arthur Szlam‡, Douwe Kiela‡, Jason Weston‡
† Montreal Institute for Learning Algorithms, MILA
‡ Facebook AI Research
saizheng. , {edinan,jju,aszlam,dkiela,jase}@fb.com
Abstract
Chit-chat models are known to have sev-
eral problems: they lack specificity, do not
display a consistent personality and are of-
ten not very captivating. In this work we
present the task of making chit-chat more
engaging by conditioning on profile infor-
mation. We collect data and train models
to (i) condition on their given profile in-
formation; and (ii) information about the
person they are talking to, resulting in im-
proved dialogues, as measured by next ut-
terance 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 inter-
locutors.
1 Introduction
Despite much recent success in natural language
processing and dialogue research, communication
between a human and a machine is still in its in-
fancy. It is only recently that neural models have
had sufficient capacity and access to sufficiently
large datasets that they appear to generate mean-
ingful responses in a chit-chat setting. Still, con-
versing with such generic chit-chat models for
even a short amount of time quickly exposes their
weaknesses (Serban et al., 2016; Vinyals and Le,
2015).
Common issues with chit-chat models include:
(i) the lack of a consistent personality (Li et al.,
2016a) as they are typically trained over many di-
alogs each with different speakers, (ii) the lack of
an explicit long-term memory as they are typically
trained to produce an utterance given only the
recent dialogue history (Vinyals and Le, 2015);
1
Work done while at Facebook AI Research.
and (iii) a tendency to produce non-specific an-
swers like “I don’t know” (Li et al., 2015). Those
three problems combine to produce an unsatisfy-
ing overall experience for a human to engage with.
We believe some of those problems are due to
there being no good publicly available dataset for
general chit-chat.
Because of the low quality of current con-
versational models, and because of the diffi-
culty in evaluating these models, chit-chat is
often ignored as an end-application. Instead,
the research community has focused on task-
oriented communication, such as airline or restau-
rant booking (Bordes et al., 2016), or else single-
turn information seeking, i.e. question answer-
ing (Rajpurkar et al., 2016). Despite the success
of the latter, simpler, domain, it is well-known
that a large quantity of human dialogue centers
on socialization, personal interests and chit-chat
(Dunbar et al., 1997). For example, less than 5%
of posts on Twitter are questions, whereas around
80% are about personal emotional state, thoughts
or activities, authored by so called “Meformers”
(Naaman et al., 2010).
In this work we make a step towards more
engaging chit-chat dialogue agents by endowing
them with a configurable, but persistent persona,
encoded by multiple sentences of textual descrip-
tion, termed a profile. This profile can be stored
in a memory-augmented neural network and then
used to produce more personal, specific, consis-
tent and engaging responses than a persona-free
model, thus alleviating some of the common is-
sues in chit-chat models. Using the same mecha-
nism, any existing information about the persona
of the dialogue partner can also be used in the
same way. Our models are thus trained to both
ask and answer questions about personal topics,
and the resulting dialogue can be used to build a
model of the persona of the speaking partner.
http://arxiv.org/abs/1801.07243v5
To support the training of such models, we
present the PERSONA-CHAT dataset, a new dia-
logue dataset consisting of 162,064 utterances be-
tween crowdworkers who were randomly paired
and each asked to act the part of a given provided
persona (randomly assigned, and created by an-
other set of crowdworkers). The paired workers
were asked to chat naturally and to get to know
each other during the conversation. This produces
interesting and engaging conversations that our
agents can try to learn to mimic.
Studying the next utterance prediction task dur-
ing dialogue, we compare a range of models: both
generative and ranking models, including Seq2Seq
models and Memory Networks (Sukhbaatar et al.,
2015) as well as other standard retrieval baselines.
We show experimentally that in either the gener-
ative or ranking case conditioning the agent with
persona information gives improved prediction of
the next dialogue utterance. The PERSONA-CHAT
dataset is designed to facilitate research into al-
leviating some of the issues that traditional chit-
chat models face, and with the aim of making such
models more consistent and engaging, by endow-
ing them with a persona. By comparing against
chit-chat models built using the OpenSubtitles and
Twitter datasets, human evaluations show that our
dataset provides more engaging models, that are
simultaneously capable of being fluent and consis-
tent via conditioning on a persistent, recognizable
profile.
2 Related Work
Traditional dialogue systems consist of building
blocks, such as dialogue state tracking compo-
nents and response generators, and have typi-
cally been applied to tasks with labeled internal
dialogue state and precisely defined user intent
(i.e., goal-oriented dialogue), see e.g. (Young,
2000). The most successful goal-oriented dia-
logue systems model conversation as partially ob-
servable Markov decision processes (POMDPs)
(Young et al., 2013). All those methods typically
do not consider the chit-chat setting and are more
concerned with achieving functional goals (e.g.
booking an airline flight) than displaying a per-
sonality. In particular, many of the tasks and
datasets available are constrained to narrow do-
mains (Serban et al., 2015).
Non-goal driven dialogue systems go back
to Weizenbaum’s famous program ELIZA
(Weizenbaum, 1966), and hand-coded systems
have continued to be used in applications to this
day. For example, modern solutions that build an
open-ended dialogue system to the Alexa chal-
lenge combine hand-coded and machine-learned
elements (Serban et al., 2017a). Amongst the
simplest of statistical systems that can be used in
this domain, that are based on data rather than
hand-coding, are information retrieval models
(Sordoni et al., 2015), which retrieve and rank
responses based on their matching score with the
recent dialogue history. We use IR systems as a
baseline in this work.
End-to-end neural approaches are a class of
models which have seen growing recent interest.
A popular class of methods are generative re-
current systems like seq2seq applied to dialogue
(Sutskever et al., 2014; Vinyals and Le, 2015;
Sordoni et al., 2015; Li et al., 2016b; Serban et al.,
2017b). Rooted in language modeling, they are
able to produce syntactically coherent novel re-
sponses, but their memory-free approach means
they lack long-term coherence and a persistent
personality, as discussed before. A promising di-
rection, that is still in its infancy, to fix this is-
sue is to use a memory-augmented network in-
stead (Sukhbaatar et al., 2015; Dodge et al., 2015)
by providing or learning appropriate memories.
Serban et al. (2015) list available corpora for
training dialogue systems. Perhaps the most
relevant to learning chit-chat models are ones
based on movie scripts such as OpenSubtitles
and Cornell Movie-Dialogue Corpus, and dia-
logue from web platforms such as Reddit and
Twitter, all of which have been used for train-
ing neural approaches (Vinyals and Le, 2015;
Dodge et al., 2015; Li et al., 2016b; Serban et al.,
2017b). Naively training on these datasets leads to
models with the lack of a consistent personality as
they will learn a model averaged over many dif-
ferent speakers. Moreover, the data does little to
encourage the model to engage in understanding
and maintaining knowledge of the dialogue part-
ner’s personality and topic interests.
According to Serban et al. (2015)’s survey, per-
sonalization of dialogue systems is “an important
task, which so far has not received much atten-
tion”. In the case of goal-oriented dialogue some
work has focused on the agent being aware of the
human’s profile and adjusting the dialogue accord-
ingly, but without a personality to the agent it-
self (Lucas et al., 2009; Joshi et al., 2017). For
the chit-chat setting, the most relevant work is
(Li et al., 2016a). For each user in the Twitter cor-
pus, personas were captured via distributed em-
beddings (one per speaker) to encapsulate individ-
ual characteristics such as background information
and speaking style, and they then showed using
those vectors improved the output of their seq2seq
model for the same speaker. Their work does not
focus on attempting to engage the other speaker by
getting to know them, as we do here. For that rea-
son, our focus is on explicit profile information,
not hard-to-interpret latent variables.
3 The PERSONA-CHAT Dataset
The aim of this work is to facilitate more en-
gaging and more personal chit-chat dialogue.
The PERSONA-CHAT dataset is a crowd-sourced
dataset, collected via Amazon Mechanical Turk,
where each of the pair of speakers condition their
dialogue on a given profile, which is provided.
The data collection consists of three stages:
(i) Personas: we crowdsource a set of 1155 pos-
sible personas, each consisting of at least 5 profile
sentences, setting aside 100 never seen before per-
sonas for validation, and 100 for test.
(ii) Revised personas: to avoid modeling that
takes advantage of trivial word overlap, we crowd-
source additional rewritten sets of the same 1155
personas, with related sentences that are rephrases,
generalizations or specializations, rendering the
task much more challenging.
(iii) Persona chat: we pair two Turkers and as-
sign them each a random (original) persona from
the pool, and ask them to chat. This resulted in a
dataset of 162,064 utterances over 10,907 dialogs,
15,602 utterances (1000 dialogs) of which are set
aside for validation, and 15,024 utterances (968 di-
alogs) for test.
The final dataset and its corresponding data col-
lection source code, as well as models trained on
the data, are all available open source in ParlAI2.
In the following, we describe each data collec-
tion stage and the resulting tasks in more detail.
3.1 Personas
We asked the crowdsourced workers to create a
character (persona) description using 5 sentences,
providing them only a single example:
2
http://parl.ai
“I am a vegetarian. I like swimming. My father
used to work for Ford. My favorite band is Ma-
roon5. I got a new job last month, which is about
advertising design.”
Our aim was to create profiles that are natural
and descriptive, and contain typical topics of hu-
man interest that the speaker can bring up in con-
versation. Because the personas are not the real
profiles of the Turkers, the dataset does not con-
tain personal information (and they are told specif-
ically not to use any). We asked the workers to
make each sentence short, with a maximum of 15
words per sentence. This is advantageous both for
humans and machines: if they are too long, crowd-
sourced workers are likely to lose interest, and for
machines the task could become more difficult.
Some examples of the personas collected are
given in Table 1 (left).
3.2 Revised Personas
A difficulty when constructing dialogue datasets,
or text datasets in general, is that in order
to encourage research progress, the task must
be carefully constructed so that is neither too
easy nor too difficult for the current technology
(Voorhees et al., 1999). One issue with condition-
ing on textual personas is that there is a danger
that humans will, even if asked not to, unwittingly
repeat profile information either verbatim or with
significant word overlap. This may make any sub-
sequent machine learning tasks less challenging,
and the solutions will not generalize to more diffi-
cult tasks. This has been a problem in some re-
cent datasets: for example, the dataset curation
technique used for the well-known SQuAD dataset
suffers from this word overlap problem to a certain
extent (Chen et al., 2017).
To alleviate this problem, we presented the orig-
inal personas we collected to a new set of crowd-
workers and asked them to rewrite the sentences
so that a new sentence is about “a related char-
acteristic that the same person may have”, hence
the revisions could be rephrases, generalizations
or specializations. For example “I like basketball”
can be revised as “I am a big fan of Michael Jor-
dan” not because they mean the same thing but
because the same persona could contain both.
In the revision task, workers are instructed not
to trivially rephrase the sentence by copying the
original words. However, during the entry stage
if a non-stop word is copied we issue a warning,
http://parl.ai
Original Persona Revised Persona
I love the beach. To me, there is nothing like a day at the seashore.
My dad has a car dealership My father sales vehicles for a living.
I just got my nails done I love to pamper myself on a regular basis.
I am on a diet now I need to lose weight.
Horses are my favorite animal. I am into equestrian sports.
I play a lot of fantasy videogames. RPGs are my favorite genre.
I have a computer science degree. I also went to school to work with technology.
My mother is a medical doctor The woman who gave birth to me is a physician.
I am very shy. I am not a social person.
I like to build model spaceships. I enjoy working with my hands.
Table 1: Example Personas (left) and their revised versions (right) from the PERSONA-CHAT dataset.
The revised versions are designed to be characteristics that the same persona might have, which could be
rephrases, generalizations or specializations.
Persona 1 Persona 2
I like to ski I am an artist
My wife does not like me anymore I have four children
I have went to Mexico 4 times this year I recently got a cat
I hate Mexican food I enjoy walking for exercise
I like to eat cheetos I love watching Game of Thrones
[PERSON 1:] Hi
[PERSON 2:] Hello ! How are you today ?
[PERSON 1:] I am good thank you , how are you.
[PERSON 2:] Great, thanks ! My children and I were just about to watch Game of Thrones.
[PERSON 1:] Nice ! How old are your children?
[PERSON 2:] I have four that range in age from 10 to 21. You?
[PERSON 1:] I do not have children at the moment.
[PERSON 2:] That just means you get to keep all the popcorn for yourself.
[PERSON 1:] And Cheetos at the moment!
[PERSON 2:] Good choice. Do you watch Game of Thrones?
[PERSON 1:] No, I do not have much time for TV.
[PERSON 2:] I usually spend my time painting: but, I love the show.
Table 2: Example dialog from the PERSONA-CHAT dataset. Person 1 is given their own persona (top left)
at the beginning of the chat, but does not know the persona of Person 2, and vice-versa. They have to get
to know each other during the conversation.
and ask them to rephrase, guaranteeing that the
instructions are followed. For example, “My fa-
ther worked for Ford.” can be revised to “My dad
worked in the car industry”, but not “My dad was
employed by Ford.” due to word overlap.
Some examples of the revised personas col-
lected are given in Table 1 (right).
3.3 Persona Chat
After collecting personas, we then collected the di-
alogues themselves, conditioned on the personas.
For each dialogue, we paired two random crowd-
workers, and gave them the instruction that they
will chit-chat with another worker, while playing
the part of a given character. We then provide them
with a randomly chosen persona from our pool,
different to their partners. The instructions are on
purpose quite terse and simply ask them to “chat
with the other person naturally and try to get to
know each other”. In an early study we noticed
the crowdworkers tending to talk about themselves
(their own persona) too much, so we also added
the instructions “both ask questions and answer
questions of your chat partner” which seemed to
help. We also gave a bonus for high quality di-
alogs. The dialog is turn-based, with a maximum
of 15 words per message. We again gave instruc-
tions to not trivially copy the character descrip-
tions into the messages, but also wrote explicit
code sending them an error if they tried to do so,
using simple string matching. We define a mini-
mum dialogue length which is randomly between
6 and 8 turns each for each dialogue. An example
dialogue from the dataset is given in Table 2.
3.4 Evaluation
We focus on the standard dialogue task of pre-
dicting the next utterance given the dialogue his-
tory, but consider this task both with and without
the profile information being given to the learn-
ing agent. Our goal is to enable interesting direc-
tions for future research, where chatbots can for
instance have personalities, or imputed personas
could be used to make dialogue more engaging to
the user.
We consider this in four possible scenarios:
conditioning on no persona, your own persona,
their persona, or both. These scenarios can be
tried using either the original personas, or the re-
vised ones. We then evaluate the task using three
metrics: (i) the log likelihood of the correct se-
quence, measured via perplexity, (ii) F1 score, and
(iii) next utterance classification loss, following
Lowe et al. (2015). The latter consists of choos-
ing N random distractor responses from other di-
alogues (in our setting, N=19) and the model se-
lecting the best response among them, resulting in
a score of one if the model chooses the correct re-
sponse, and zero otherwise (called hits@1 in the
experiments).
4 Models
We consider two classes of model for next utter-
ance prediction: ranking models and generative
models. Ranking models produce a next utterance
by considering any utterance in the training set as a
possible candidate reply. Generative models gen-
erate novel sentences by conditioning on the dia-
logue history (and possibly, the persona), and then
generating the response word-by-word. Note one
can still evaluate the latter as ranking models by
computing the probability of generating a given
candidate, and ranking candidates by those scores.
4.1 Baseline ranking models
We first consider two baseline models, an IR
baseline (Sordoni et al., 2015) and a supervised
embedding model, Starspace (Wu et al., 2017)3.
While there are many IR variants, we adopt the
simplest one: find the most similar message in
the (training) dataset and output the response
from that exchange. Similarity is measured by
the tf-idf weighted cosine similarity between the
bags of words. Starspace is a recent model that
3
github.com/facebookresearch/StarSpace
also performs information retrieval but by learn-
ing the similarity between the dialog and the
next utterance by optimizing the embeddings di-
rectly for that task using the margin ranking loss
and k-negative sampling. The similarity function
sim(q, c′) is the cosine similarity of the sum of
word embeddings of the query q and candidate c′.
Denoting the dictionary of D word embeddings as
W which is a D× d matrix, where Wi indexes the
ith word (row), yielding its d-dimensional embed-
ding, it embeds the sequences q and c′.
In both methods, IR and StarSpace, to incor-
porate the profile we simply concatenate it to the
query vector bag of words.
4.2 Ranking Profile Memory Network
Both the previous models use the profile infor-
mation by combining it with the dialogue history,
which means those models cannot differentiate be-
tween the two when deciding on the next utter-
ance. In this model we instead use a memory
network with the dialogue history as input, which
then performs attention over the profile to find rel-
evant lines from the profile to combine with the
input, and then finally predicts the next utterance.
We use the same representation and loss as in the
Starspace model, so without the profile, the two
models are identical. When the profile is available
attention is performed by computing the similarity
of the input q with the profile sentences pi, com-
puting the softmax, and taking the weighted sum:
q+ = q+
∑
sipi, si = Softmax(sim(q, pi))
where Softmax(zi) = e
zi/
∑
j
ezj . One can then
rank the candidates c′ using sim(q+, c′). One can
also perform multiple “hops” of attention over the
profile rather than one, as shown here, although
that did not bring significant gains in our parame-
ter sweeps.
4.3 Key-Value Profile Memory Network
The key-value (KV) memory network
(Miller et al., 2016) was proposed as an im-
provement to the memory network by performing
attention over keys and outputting the values
(instead of the same keys as in the original),
which can outperform memory networks depen-
dent on the task and definition of the key-value
pairs. Here, we apply this model to dialogue,
and consider the keys as dialog histories (from
the training set), and the values as the next
github.com/facebookresearch/StarSpace
dialogue utterances, i.e., the replies from the
speaking partner. This allows the model to have
a memory of past dialogues that it can directly
use to help influence its prediction for the current
conversation. The model we choose is identical to
the profile memory network just described in the
first hop over profiles, while in the second hop,
q+ is used to attend over the keys and output a
weighted sum of values as before, producing q++.
This is then used to rank the candidates c′ using
sim(q++, c′) as before. As the set of (key-value)
pairs is large this would make training very slow.
In our experiments we simply trained the profile
memory network and used the same weights
from that model and applied this architecture at
test time instead. Training the model directly
would presumably give better results, however
this heuristic already proved beneficial compared
to the original network.
4.4 Seq2Seq
The input sequence x is encoded by applying
het = LSTMenc(xt | h
e
t−1). We use GloVe
(Pennington et al., 2014) for our word embed-
dings. The final hidden state, het , is fed into the
decoder LSTMdec as the initial state h
d
0
. For each
time step t, the decoder then produces the proba-
bility of a word j occurring in that place via the
softmax, i.e.,
p(yt,j = 1 | yt−1, . . . , y1) =
exp(wjh
d
t )∑K
j′=1
exp(wj′h
d
t )
.
The model is trained via negative log likelihood.
The basic model can be extended to include
persona information, in which case we simply
prepend it to the input sequence x, i.e., x = ∀p ∈
P || x, where || denotes concatenation. For the
OpenSubtitles and Twitter datasets trained in Sec-
tion 5.2 we found training a language model (LM),
essentially just the decoder part of this model,
worked better and we report that instead.
4.5 Generative Profile Memory Network
Finally, we introduce a generative model that en-
codes each of the profile entries as individual
memory representations in a memory network.
As before, the dialogue history is encoded via
LSTMenc, the final state of which is used as the
initial hidden state of the decoder. Each entry pi =
〈pi,1, . . . , pi,n〉 ∈ P is then encoded via f(pi) =∑|pi|
j
αipi,j . That is, we weight words by their in-
verse term frequency: αi = 1/(1 + log(1 + tf))
where tf is computed from the GloVe index via
Zipf’s law4. Let F be the set of encoded memo-
ries. The decoder now attends over the encoded
profile entries, i.e., we compute the mask at, con-
text ct and next input x̂t as:
at = softmax(FWah
d
t ),
ct = a
⊺
tF ; x̂t = tanh(Wc[ct−1, xt]).
If the model has no profile information, and hence
no memory, it becomes equivalent to the Seq2Seq
model.
5 Experiments
We first report results using automated evalua-
tion metrics, and subsequently perform an extrin-
sic evaluation where crowdsourced workers per-
form a human evaluation of our models.
5.1 Automated metrics
The main results are reported in Table 3. Overall,
the results show the following key points:
Persona Conditioning Most models improve
significantly when conditioning prediction on their
own persona at least for the original (non-revised)
versions, which is an easier task than the re-
vised ones which have no word overlap. For
example, the Profile Memory generation model
has improved perplexity and hits@1 compared to
Seq2Seq, and all the ranking algorithms (IR base-
line, Starspace and Profile Memory Networks) ob-
tain improved hits@1.
Ranking vs. Generative. Ranking models are
far better than generative models at ranking. This
is perhaps obvious as that is the metric they are
optimizing, but still the performance difference is
quite stark. It may be that the word-based proba-
bility which generative models use works well, but
is not calibrated well enough to give a sentence-
based probability which ranking requires. Human
evaluation is also used to compare these methods,
which we perform in Sec. 5.2.
Ranking Models. For the ranking models, the
IR baseline is outperformed by Starspace due to
its learnt similarity metric, which in turn is out-
performed by Profile Memory networks due to the
attention mechanism over the profiles (as all other
parts of the models are the same). Finally KV Pro-
file Memory networks outperform Profile Memory
Networks in the no persona case due to the ability
4
tf = 1e6 ∗ 1/(idx1.07)
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
Table 3: Evaluation of dialog utterance prediction with various models in three settings: without
conditioning on a persona, conditioned on the speakers given persona (“Original Persona”), or a revised
persona that does not have word overlap.
Method Persona
Model Profile Fluency Engagingness Consistency 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)
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)
Table 4: Human Evaluation of various PERSONA-CHAT models, along with a comparison to human per-
formance, and Twitter and OpenSubtitles based models (last 4 rows), standard deviation in parenthesis.
to consider neighboring dialogue history and next
utterance pairs in the training set that are similar to
the current dialogue, however when using persona
information the performance is similar.
Revised Personas. Revised personas are much
harder to use. We do however still see some
gain for the Profile Memory networks compared
to none (0.354 vs. 0.318 hits@1). We also tried
two variants of training: with the original personas
in the training set or the revised ones, a compari-
son of which is shown in Table 6 of the Appendix.
Training on revised personas helps, both for test
examples that are in original form or revised form,
likely due to the model be forced to learn more
than simple word overlap, forcing the model to
generalize more (i.e., learn semantic similarity of
differing phrases).
Their Persona. We can also condition a model
on the other speaker’s persona, or both personas
at once, the results of which are in Tables 5 and 6
in the Appendix. Using “Their persona” has less
impact on this dataset. We believe this is because
most speakers tend to focus on themselves when
it comes to their interests. It would be interest-
ing how often this is the case in other datasets.
Certainly this is skewed by the particular instruc-
tions one could give to the crowdworkers. For
example if we gave the instructions “try not to
talk about yourself, but about the other’s interests’
likely these metrics would change.
5.2 Human Evaluation
As automated metrics are notoriously poor for
evaluating dialogue (Liu et al., 2016) we also per-
form human evaluation using crowdsourced work-
ers. The procedure is as follows. We perform al-
most exactly the same setup as in the dataset col-
lection process itself as in Section 3.3. In that
setup, we paired two Turkers and assigned them
each a random (original) persona from the col-
lected pool, and asked them to chat. Here, from
the Turker’s point of view everything looks the
same except instead of being paired with a Turker
they are paired with one of our models instead
(they do not know this). In this setting, for both
the Turker and the model, the personas come from
the test set pool.
After the dialogue, we then ask the Turker some
additional questions in order to evaluate the qual-
ity of the model. We ask them to evaluate fluency,
engagingness and consistency (scored between 1-
5). Finally, we measure the ability to detect the
other speaker’s profile by displaying two possi-
ble profiles, and ask which is more likely to be
the profile of the person the Turker just spoke to.
More details of these measures are given in the
Appendix.
The results are reported in Table 4 for the best
performing generative and ranking models, in both
the No Persona and Self Persona categories, 100
dialogues each. We also evaluate the scores of hu-
man performance by replacing the chatbot with a
human (another Turker). This effectively gives us
upper bound scores which we can aim for with our
models. Finally, and importantly, we compare our
models trained on PERSONA-CHAT with chit-chat
models trained with the Twitter and OpenSubtitles
datasets (2009 and 2018 versions) instead, follow-
ing Vinyals and Le (2015). Example chats from a
few of the models are shown in the Appendix in
Tables 7, 8, 9, 10, 11 and 12.
Firstly, we see a difference in fluency, engag-
ingness and consistency between all PERSONA-
CHAT models and the models trained on OpenSub-
titles and Twitter. PERSONA-CHAT is a resource
that is particularly strong at providing training data
for the beginning of conversations, when the two
speakers do not know each other, focusing on ask-
ing and answering questions, in contrast to other
resources. We also see suggestions of more sub-
tle differences between the models, although these
differences are obscured by the high variance of
the human raters’ evaluations. For example, in
both the generative and ranking model cases, mod-
els endowed with a persona can be detected by the
human conversation partner, as evidenced by the
persona detection accuracies, whilst maintaining
fluency and consistency compared to their non-
persona driven counterparts.
Finding the balance between fluency, engage-
ment, consistency, and a persistent persona re-
mains a strong challenge for future research.
5.3 Profile Prediction
Two tasks could naturally be considered using
PERSONACHAT: (1) next utterance prediction
during dialogue, and (2) profile prediction given
dialogue history. The main study of this work has
been Task 1, where we have shown the use of pro-
file information. Task 2, however, can be used to
extract such information. While a full study is be-
yond the scope of this paper, we conducted some
preliminary experiments, the details of which are
in Appendix D. They show (i) human speaker’s
profiles can be predicted from their dialogue with
high accuracy (94.3%, similar to human perfor-
mance in Table 4) or even from the model’s di-
alogue (23% with KV Profile Memory) showing
the model is paying attention to the human’s inter-
ests. Further, the accuracies clearly improve with
further dialogue, as shown in Table 14. Combining
Task 1 and Task 2 into a full system is an exciting
area of future research.
6 Conclusion & Discussion
In this work we have introduced the PERSONA-
CHAT dataset, which consists of crowd-sourced di-
alogues where each participant plays the part of an
assigned persona; and each (crowd-sourced) per-
sona has a word-distinct paraphrase. We test vari-
ous baseline models on this dataset, and show that
models that have access to their own personas in
addition to the state of the dialogue are scored as
more consistent by annotators, although not more
engaging. On the other hand, we show that models
trained on PERSONA-CHAT (with or without per-
sonas) are more engaging than models trained on
dialogue from other resources (movies, Twitter).
We believe PERSONA-CHAT will be a useful re-
source for training components of future dialogue
systems. Because we have paired human gener-
ated profiles and conversations, the data aids the
construction of agents that have consistent per-
sonalities and viewpoints. Furthermore, predict-
ing the profiles from a conversation moves chit-
chat tasks in the direction of goal-directed dia-
logue, which has metrics for success. Because we
collect paraphrases of the profiles, they cannot be
trivially matched; indeed, we believe the original
and rephrased profiles are interesting as a semantic
similarity dataset in their own right. We hope that
the data will aid training agents that can ask ques-
tions about users’ profiles, remember the answers,
and use them naturally in conversation.
References
Antoine Bordes, Y-Lan Boureau and Jason Weston.
2016. Learning end-to-end goal-oriented dialog.
arXiv preprint arXiv:1605.07683.
Danqi Chen, Adam Fisch, Jason Weston, and An-
toine Bordes. 2017. Reading wikipedia to an-
swer open-domain questions. arXiv preprint
arXiv:1704.00051.
Jesse Dodge, Andreea Gane, Xiang Zhang, Antoine
Bordes, Sumit Chopra, Alexander Miller, Arthur
Szlam, and Jason Weston. 2015. Evaluating prereq-
uisite qualities for learning end-to-end dialog sys-
tems. arXiv preprint arXiv:1511.06931.
Robin IM Dunbar, Anna Marriott, and Neil DC Dun-
can. 1997. Human conversational behavior. Human
nature, 8(3):231–246.
Chaitanya K Joshi, Fei Mi, and Boi Faltings. 2017.
Personalization in goal-oriented dialog. arXiv
preprint arXiv:1706.07503.
Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao,
and Bill Dolan. 2015. A diversity-promoting objec-
tive function for neural conversation models. arXiv
preprint arXiv:1510.03055.
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.
Jiwei Li, Will Monroe, Alan Ritter, Michel Galley,
Jianfeng Gao, and Dan Jurafsky. 2016b. Deep re-
inforcement learning for dialogue generation. arXiv
preprint arXiv:1606.01541.
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 met-
rics for dialogue response generation. arXiv preprint
arXiv:1603.08023.
Ryan Lowe, Nissan Pow, Iulian Serban, and Joelle
Pineau. 2015. The ubuntu dialogue corpus: A large
dataset for research in unstructured multi-turn dia-
logue systems. arXiv preprint arXiv:1506.08909.
JM Lucas, F Fernández, J Salazar, J Ferreiros, and
R San Segundo. 2009. Managing speaker iden-
tity and user profiles in a spoken dialogue system.
Procesamiento del Lenguaje Natural, 43:77–84.
Alexander Miller, Adam Fisch, Jesse Dodge, Amir-
Hossein Karimi, Antoine Bordes, and Jason We-
ston. 2016. Key-value memory networks for
directly reading documents. arXiv preprint
arXiv:1606.03126.
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 coopera-
tive work, pages 189–192. ACM.
Jeffrey Pennington, Richard Socher, and Christopher
Manning. 2014. Glove: Global vectors for word
representation. In Proceedings of the 2014 confer-
ence on empirical methods in natural language pro-
cessing (EMNLP), pages 1532–1543.
Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and
Percy Liang. 2016. Squad: 100,000+ questions
for machine comprehension of text. arXiv preprint
arXiv:1606.05250.
Iulian V Serban, Chinnadhurai Sankar, Mathieu Ger-
main, Saizheng Zhang, Zhouhan Lin, Sandeep Sub-
ramanian, Taesup Kim, Michael Pieper, Sarath
Chandar, Nan Rosemary Ke, et al. 2017a. A
deep reinforcement learning chatbot. arXiv preprint
arXiv:1709.02349.
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.
Iulian Vlad Serban, Ryan Lowe, Laurent Charlin, and
Joelle Pineau. 2016. Generative deep neural net-
works for dialogue: A short review. arXiv preprint
arXiv:1611.06216.
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 di-
alogues.
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 gen-
eration of conversational responses. arXiv preprint
arXiv:1506.06714.
Sainbayar Sukhbaatar, Jason Weston, Rob Fergus, et al.
2015. End-to-end memory networks. In Advances
in neural information processing systems, pages
2440–2448.
Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014.
Sequence to sequence learning with neural net-
works. In Advances in neural information process-
ing systems, pages 3104–3112.
Oriol Vinyals and Quoc Le. 2015. A neural conversa-
tional model. arXiv preprint arXiv:1506.05869.
Ellen M Voorhees et al. 1999. The trec-8 question an-
swering track report. In Trec, volume 99, pages 77–
82.
Joseph Weizenbaum. 1966. Elizaa computer program
for the study of natural language communication be-
tween man and machine. Communications of the
ACM, 9(1):36–45.
Ledell Wu, Adam Fisch, Sumit Chopra, Keith
Adams, Antoine Bordes, and Jason Weston. 2017.
Starspace: Embed all the things! arXiv preprint
arXiv:1709.03856.
Steve Young, Milica Gašić, Blaise Thomson, and Ja-
son D Williams. 2013. Pomdp-based statistical spo-
ken dialog systems: A review. Proceedings of the
IEEE, 101(5):1160–1179.
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.
A Next Utterance Prediction Additional
Evaluation Metrics
In Table 5 and Table 6 we show additional results
on next utterance prediction. In particular we give
results for conditioning on the other person’s per-
sona (”Their”) instead of one’s own (”Self”) or the
combination (”Both”). In particular, we see that,
knowing the other’s persona does not help preci-
sion on this data with these models. Finally, we
show in Table 6 for the ranking models the per-
formance difference from training on original per-
sonas versus revised ones. Revised personas give
improved results, perhaps because the models are
forced to learn more than trivial word overlap (i.e.
semantic similarity of differing phrases).
B Example Dialogs between Humans
and Models
In Tables 7, 8, 9, 11, 12 we show example dia-
logues between Turkers and various models that
were collected as part of the human evaluation of
Section 5.2.
C Human Evaluation Measures
After dialogues between humans and a model, we
then ask the Turker some additional questions in
order to evaluate the quality of the model. They
are, in order:
• Fluency: We ask them to judge the fluency
of the other speaker as a score from 1 to 5,
where 1 is “not fluent at all”, 5 is “extremely
fluent”, and 3 is “OK”.
• Engagingness: We ask them to judge the en-
gagingness of the other speaker disregarding
fluency from 1-5, where 1 is “not engaging
at all”, 5 is “extremely engaging”, and 3 is
“OK”.
• Consistency: We ask them to judge the con-
sistency of the persona of the other speaker,
where we give the example that “I have a
dog” followed by “I have no pets” is not con-
sistent. The score is again from 1-5.
• Profile Detection: Finally, we display two
possible profiles, and ask which is more
likely to be the profile of the person the
Turker just spoke to. One profile is chosen
at random, and the other is the true persona
given to the model.
D Profile Prediction
While the main study of this work is the ability
to improve next utterance classification by condi-
tioning on a persona, one could naturally consider
two tasks: (1) next utterance prediction during di-
alogue, and (2) profile prediction given dialogue
history. In the main paper we show that Task 1 can
be improved by using profile information. Task 2,
however, can be used to extract such information.
In this section we conduct a preliminary study
of the ability to predict the persona of a speaker
given a set of dialogue utterances. We consider the
dialogues between humans (PERSON 0) and our
best performing model, the retrieval-based Key-
Value Profile Memory Network (PERSON 1) from
Section 5.2. We tested the ability to predict the
profile information of the two speakers from the
dialogue utterances of each speaker, considering
all four combinations. We employ the same IR
baseline model used in the main paper to predict
profiles: it ranks profile candidates, either at the
entire profile level (considering all the sentences
that make up the profile as a bag) or at the sen-
tence level (each sentence individually). We con-
sider 100 negative profile candidates for each pos-
itive profile, and compute the error rate of predict-
ing the true profile averaged over all dialogues and
candidates. The results are given in Table 13, both
for the model conditioned on profile information,
and the same KV Memory model that is not. The
results indicate the following:
• It is possible to predict the humans profile
from their dialogue utterances (PERSON 0,
Profile 0) with high accuracy at both the pro-
file and sentence level, independent of the
model they speaking to.
• Similarly the model’s profile can be predicted
with high accuracy from its utterances (PER-
SON 1, Profile 1) when it is conditioned on
the profile, otherwise this is chance level (w/o
Profile).
• It is possible to predict the model’s profile
from the human’s dialogue, but with a lower
accuracy (PERSON 0, Profile 1) as long as
the model is conditioned on its own pro-
file. This indicates the human responds to the
model’s utterances and pays attention to the
model’s interests.
Persona Method
Original Revised
ppl hits@1 F1 ppl hits@1 F1
No Persona 38.08 0.092 0.168 38.08 0.092 0.168
Self Persona
Seq2Seq 40.53 0.084 0.172 40.65 0.082 0.171
Profile Memory 34.54 0.125 0.172 38.21 0.108 0.170
Their Persona
Seq2Seq 41.48 0.075 0.168 41.95 0.074 0.168
Profile Memory 36.42 0.105 0.167 37.75 0.103 0.167
Both Personas
Seq2Seq 40.14 0.084 0.169 40.53 0.082 0.166
Profile Memory 35.27 0.115 0.171 38.48 0.106 0.168
Table 5: Evaluation of dialog utterance prediction with generative models in four settings: condi-
tioned on the speakers persona (“self persona”), the dialogue partner’s persona (“their persona”), both
or none. The personas are either the original source given to Turkers to condition the dialogue, or the
revised personas that do not have word overlap. In the “no persona” setting, the models are equivalent,
so we only report once.
No Persona Self Persona Their Persona Both Personas
Method Orig Rewrite Orig Rewrite Orig Rewrite Orig Rewrite
IR baseline 0.214 0.214 0.410 0.207 0.181 0.181 0.382 0.188
Training on original personas
Starspace 0.318 0.318 0.481 0.295 0.245 0.235 0.429 0.258
Profile Memory 0.318 0.318 0.473 0.302 0.283 0.267 0.438 0.266
Training on revised personas
Starspace 0.318 0.318 0.491 0.322 0.271 0.261 0.432 0.288
Profile Memory 0.318 0.318 0.509 0.354 0.299 0.294 0.467 0.331
KV Profile Memory 0.349 0.349 0.511 0.351 0.291 0.289 0.467 0.330
Table 6: Evaluation of dialog utterance prediction with ranking models using hits@1 in four settings:
conditioned on the speakers persona (”self persona”), the dialogue partner’s persona (”their persona”),
both or none. The personas are either the original source given to Turkers to condition the dialogue, or
the rewritten personas that do not have word overlap, explaining the poor performance of IR in that case.
• Similarly, the human’s profile can be pre-
dicted from the model’s dialogue, but with
lower accuracy. Interestingly, the model
without profile conditioning is better at this,
perhaps because it does not concentrate on
talking about itself, and pays more attention
to responding to the human’s interests. There
appears to be a tradeoff that needs to be ex-
plored and understood here.
We also study the performance of profile pre-
diction as the dialogue progresses, by computing
error rates for dialogue lengths 1 to 8 (the longest
length we consider in this work). The results,
given in Table 14, show the error rate of predict-
ing the persona decreases in all cases as dialogue
length increases.
Overall, the results in this section show that it is
plausible to predict profiles given dialogue utter-
ances, which is an important extraction task. Note
that better results could likely be achieved with
more sophisticated models.
Persona 1 Persona 2 (Not used by model)
I am a doctor I eat sardines for breakfast daily
I have one brother I just had surgery
My mom is a secretary I love playing Frisbee
I like to eat a lot of meat I have two large dogs
I am a bodybuilder I never finished high school
[PERSON 1:] Hi I am Dr. Jenson. What is your name?
[PERSON 2:] Miss Dawes.
[PERSON 1:] Hi Miss Dawes, tell me about your family!
[PERSON 2:] Miss Baggoli, I know it must show on your records that my name is Mary, but everyone
has referred to me as Lola since I was a squealing infant in my mother’ s arms.
[PERSON 1:] fascinating. My secretary mother would have made note of all of that.
[PERSON 2:] In addition to the remainder, it writes my autobiography.
[PERSON 1:] Oh yeah? well I enjoy a nice steak. how about you?
[PERSON 2:] Father always used to say that if we ever had the money you have…
we’ d have steak and ice cream three times every day!
[PERSON 1:] I eat all the steak and chicken too, even bacon
[PERSON 2:] Father always used to say that if we ever had the money you have…
we’ d have steak and ice cream three times every day!
[PERSON 1:] yeah you said that already
[PERSON 2:] Dynamic structure can absorb shock.
[PERSON 1:] yes it can. I know about that as a bodybuilder
[PERSON 2:] Ladies and gentlemen… the strongest man in the world is blowing up a hot water bottle.
Table 7: Example dialog between a human (Person 1) and the OpenSubtitles KV Memory Network
model (Person 2).
Persona 1 Persona 2 (Not used by model)
I am afraid of dogs I am a carpenter
I used to work at a carnival I used to drink but I stopped five years ago
I like to drink scotch to relax I have a wife and three kids
I write short stories for a living I like playing poker
I like dark superhero movies I dont have many friends
[PERSON 1:] Hi are you relaxing on this fine evening?
[PERSON 2:] I am doing good and you?
[PERSON 1:] I am well, just drinking some scotch and working on my writing
[PERSON 2:] I love to cook and bake but I rarely make time for it
[PERSON 1:] what do you usually bake?
[PERSON 2:] I do not cook, I bake cakes at a local restaurant. how about you?
[PERSON 1:] I used to be a juggler at a carnival but now I am an author
[PERSON 2:] I am a dancer, what do you do?
[PERSON 1:] I just told you. Do you have any pets?
[PERSON 2:] no, I am allergic to pets unfortunately.
[PERSON 1:] I understand. I am not allergic but I am terrified of dogs!
[PERSON 2:] I have a cat. she has been with me through three wives, haha.
[PERSON 1:] Funny that you just told me you have no pets
[PERSON 2:] I do not have any animals.
Table 8: Example dialog between a human (Person 1) and the Seq2Seq model (Person 2).
Persona 1 Persona 2
I am engaged to be married I am a writer
I used to play sports in high school I write romance novels
My favorite sport is soccer I live on a faraway island
I am a college graduate I am an introvert
I love watching sports on TV I have two cats
[PERSON 1:] I just got done playing soccer. What have you been up to?
[PERSON 2:] I do not have pets. I spend my free time writing. what about you?
[PERSON 1:] Just graduated from college. Looking for work.
[PERSON 2:] yes I graduated, in writing and poetry
[PERSON 1:] Have you had anything published?
[PERSON 2:] I mainly write fiction novels.
[PERSON 1:] Do you like to watch sports?
[PERSON 2:] do you like kings of leon my favorite by them is use somebody
[PERSON 1:] Are you married? I will be married soon.
[PERSON 2:] haha, no time. I have got a novel to finish.
[PERSON 1:] What is your novel about?
[PERSON 2:] it is a love story. I really enjoy writing. do you write or read?
[PERSON 1:] Like to read. Spend most time watching football.
[PERSON 2:] yes I do I am and author for a living
Table 9: Example dialog between a human (Person 1) and the Key-Value Profile Memory Network with
Self Persona.
Persona 1 Persona 2
I dont like exercise or physical activity My favorite type of food is Italian
My favorite food is ice cream I have a big pet dog
I am a homebody I want to travel to Europe
I have a husband who I love very much I love reading books about travel
I love watching movies and TV I like to walk long distances
[PERSON 1:] Hello, how are you doing today?
[PERSON 2:] I am good. I just got finished walking my dog.
[PERSON 1:] Oh cool! I am just sitting watching a movie.
[PERSON 2:] I have been watching that one before. I love it.
[PERSON 1:] Me too. I don’t like going out much so I enjoy movies and TV.
[PERSON 2:] that is a great idea. I am going to read a book on writing.
[PERSON 1:] Oh, are you a writer?
[PERSON 2:] yes, I love to read. I am a great author.
[PERSON 1:] What sorts of topics do you write about?
[PERSON 2:] I could read all of those. I am sure you enjoy reading.
[PERSON 1:] I do, but I prefer movies.
[PERSON 2:] I have not, I can not do that. I have to go to the east coast.
[PERSON 1:] Why can’t you watch movies?
[PERSON 2:] I have not. I have always had that done.
Table 10: Example dialog between a human (Person 1) and the Generative Profile Memory Network with
Self Persona.
Persona 1 Persona 2 (Not used by model)
I am very active I am an engineering student
I live in the desert I work at Eureka
I enjoy the outdoors I like to celebrate with friends
I was named something pretty common I enjoy spending time with my toes in the sand
I instruct people how to get fit as a job I study to be a businessman
[PERSON 1:] Hello there, how are you?
[PERSON 2:] I’m fine.
[PERSON 1:] How is the weather, it is hot here
[PERSON 2:] It’s so hot.
[PERSON 1:] oh yes tell me about it especially when you live in a desert.
[PERSON 2:] the sea is shining.
[PERSON 1:] you at the beach, i love the beach
[PERSON 2:] wait, wait, wait…
[PERSON 1:] what’s up, you going somewhere
[PERSON 2:] you can be a good friend.
[PERSON 1:] yes, if you want me to
[PERSON 2:] I’m sorry.
Table 11: Example dialog between a human (Person 1) and the Language Model trained on the Open-
Subtitles 2018 dataset (does not use persona).
Persona 1 Persona 2 (Not used by model)
I love ordering from Amazon I really enjoy watching classic films over the new ones
I am a mother My mother is the greatest chef in the world
I have a very active imagination I stand out among my surroundings
Winter holidays have always been my favorite My son and daughter lived with me full time after I
I love the Harry Potter books separated from my husband
[PERSON 1:] Good afternoon! How are you?
[PERSON 2:] i’m not sure. I was just talking about the other day.
[PERSON 1:] Do you like to read? I like reading Harry Potter.
[PERSON 2:] I love you so much
[PERSON 1:] That’s funny. You don’t know me.
[PERSON 2:] I think you have to be a good person.
[PERSON 1:] Where do you live? Are you in the US?
[PERSON 2:] yes, I just found out that there were only 10 people who were in the city.
[PERSON 1:] Do you have any children? I am a mother to 1 cat.
[PERSON 2:] this is the first time in history, but not a few.
[PERSON 1:] Is it cold where you are?
[PERSON 2:] I don’t even know what I’m talking about.
Table 12: Example dialog between a human (Person 1) and the Language Model trained on the Twitter
dataset (does not use persona).
Speaker Profile
Profile Level Sentence Level
KV Profile KV w/o Profile KV Profile KV w/o Profile
PERSON 0 Profile 0 0.057 0.017 0.173 0.141
PERSON 0 Profile 1 0.234 0.491 0.431 0.518
PERSON 1 Profile 0 0.254 0.112 0.431 0.349
PERSON 1 Profile 1 0.011 0.512 0.246 0.530
Table 13: Profile Prediction. Error rates are given for predicting either the persona of speaker 0 (Profile
0) or of speaker 1 (Profile 1) given the dialogue utterances of speaker 0 (PERSON 0) or speaker 1
(PERSON 1). This is shown for dialogues between humans (PERSON 0) and either the KV Profile
Memory model (“KV Profile”) which conditions on its own profile, or the KV Memory model (“KV w/o
Profile”) which does not.
Speaker Profile
Dialogue Length
1 2 3 4 5 6 7 8
PERSON 0 Profile 0 0.76 0.47 0.35 0.29 0.23 0.19 0.17 0.17
PERSON 0 Profile 1 0.51 0.39 0.32 0.29 0.27 0.27 0.25 0.25
PERSON 1 Profile 0 0.57 0.52 0.48 0.46 0.45 0.43 0.43 0.43
PERSON 1 Profile 1 0.81 0.58 0.48 0.47 0.45 0.44 0.43 0.43
Table 14: Profile Prediction By Dialog Length. Error rates are given for predicting either the persona
of speaker 0 (Profile 0) or of speaker 1 (Profile 1) given the dialogue utterances of speaker 0 (PERSON
0) or speaker 1 (PERSON 1). This is shown for dialogues between humans (PERSON 0) and the KV
Profile Memory model averaged over the first N dialogue utterances from 100 conversations (where N
is the “Dialogue Length”). The results show the accuracy of predicting the persona improves in all cases
as dialogue length increases.