Hints of Independence in a Pre-scripted World:

On Controlled Usage of Open-domain Language Models for Chatbots in

Highly Sensitive Domains

Erkan Bas¸ar

, Iris Hendrickx

, Emiel Krahmer

, Gert-Jan de Bruijn

4,5

and Tibor Bosse

Behavioural Science Institute, Radboud University, Nijmegen, The Netherlands

Centre for Language Studies, Radboud University, Nijmegen, The Netherlands

Tilburg School of Humanities and Digital Sciences, Tilburg University, Tilburg, The Netherlands

Department of Communication Studies, University of Antwerp, Antwerp, Belgium

Faculty of Social and Behavioural Sciences, University of Amsterdam, Amsterdam, The Netherlands

Keywords:

Hybrid Conversational Agents, Task-oriented Dialogue Systems, Multi-turn Response Selection, Natural

Language Generation.

Abstract:

Open-domain large language models have progressed to generating natural-sounding and coherent text. Even

though the generated texts appear human-like, the main stumbling block is that their output is never fully pre-

dictable, which runs the risk of resulting in harmful content such as false statements or inﬂammatory language.

This makes it difﬁcult to apply these models in highly sensitive domains including personal health counselling.

Hence, most of the chatbots for highly sensitive domains are developed using pre-scripted approaches. Al-

though pre-scripted approaches are highly controlled, they suffer from repetitiveness and scalability issues.

In this paper, we explore the possibility of combining the best of both worlds. We propose and describe in

detail a new, ﬂexible expert-driven hybrid architecture for harnessing the beneﬁts of large language models in

a controlled manner for highly sensitive domains and discuss the expectations and challenges.

1 INTRODUCTION

The attention for conversational agents (or chatbots)

has been on the rise in recent years. Technologi-

cal advancements allowed the development of chat-

bot applications in many domains, ranging from cus-

tomer services and entertainment, to smart personal

assistants. Health-related domains and counselling

also beneﬁt from this technology for a number of rea-

sons. The chatbots are always accessible, have in-

ﬁnite patience, and can simultaneously interact with

multiple users (Krahmer et al., 2021). However,

personal health is a highly sensitive domain where

mishaps in the communication are not acceptable.

Counselling chatbots are required to be fair and non-

discriminatory against their users. As a result, cre-

ating engaging chatbots for highly sensitive domains

becomes a challenge.

Highly sensitive domain chatbots are often built

on pre-scripted methods (such as rules- and template-

based systems) and retrieval-based methods, because

these methods are highly controllable. They typically

operate on human-authored utterances. This lowers

the risk of uttering harmful content (such as false

statements or inﬂammatory language) to almost none.

However, the same feature makes them suffer from

repetitiveness and the lack of ﬂexibility. These is-

sues decrease the user experience, especially during

long-term interactions where users can have multiple

sessions with the chatbot. In contrast, the latest open-

domain large language models can automatically gen-

erate coherent responses based on any given conver-

sation context. Yet their output is never fully pre-

dictable, and they contain the risk of generating harm-

ful content.

In this study, we address these problems by bridg-

ing the gap between the advancements in the natural

language generation (NLG) area and real-world coun-

selling chatbot applications. We focus on developing

personal health counselling chatbots with an attention

to long-term interactions, and explore the possibili-

ties of combining the ﬂexibility of neural generation

models with the low-risk and structured advantages of

retrieval-based and pre-scripted chatbot models.

Ba¸sar, E., Hendrickx, I., Krahmer, E., de Bruijn, G. and Bosse, T.

Hints of Independence in a Pre-scripted World: On Controlled Usage of Open-domain Language Models for Chatbots in Highly Sensitive Domains.

DOI: 10.5220/0010914300003116

In Proceedings of the 14th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2022) - Volume 1, pages 401-407

ISBN: 978-989-758-547-0; ISSN: 2184-433X

401

We introduce a hybrid approach consisting of

three state-of-the-art open-domain neural genera-

tion models (namely GPT-3, BERT, DialoGPT), a

novel retrieval-based approach using domain-speciﬁc

human-authored utterances, a dialogue state tracker,

and a neural response selection model which selects

the best ﬁtting response from a set of both human-

authored and automatically generated candidates. We

expect the open-domain NLG models to add ﬂexibil-

ity to the chatbot and prevent repetitiveness, while the

dialogue state tracker and retrieval-based model pro-

vide control similar to pre-scripted approaches.

2 OPEN-DOMAIN NEURAL

LANGUAGE GENERATION

Lately, arguably the most inﬂuential advancement

in the natural language processing (NLP) area has

been large-scale neural language modelling with deep

neural networks, or foundation models (Bommasani

et al., 2021), such as ELMo (Peters et al., 2018),

BERT (Devlin et al., 2019), and GPT (Brown et al.,

2020). These models have rapidly displayed a sig-

niﬁcant increase in performance over the state-of-the-

art in a broad range of NLP tasks. With conspicuous

results, they have quickly gained the attention of the

NLP community and have reached various commer-

cial successes (such as Google using BERT for ma-

chine translation).

Typically, these large neural language models con-

sist of a special type of artiﬁcial neural network

architecture called Sequence-2-Sequence (Seq2Seq;

Sutskever et al. 2014). They typically incorporate bil-

lions of parameters and are trained on large corpora

in an unsupervised fashion, that is, without the re-

quirement of any labelling by human annotators (Gu-

rurangan et al., 2020). Because the training approach

is unsupervised, it is possible to include a massive

amount of data from various online sources, such as

Wikipedia and CommonCrawl, with minimal human

labour required. Hence, these neural language models

can create accurate grammatical and semantic repre-

sentations for words (known as word embeddings) by

analysing the context in which the words occur in the

training data (Mikolov et al., 2013; Peters et al., 2018;

Devlin et al., 2019). Later, these pre-trained language

models can be reused to generate word embeddings,

which are used as feature vectors in supervised train-

ing for downstream NLP tasks.

Seq2Seq neural networks are also suitable to be

trained as text-to-text natural language generation

models (Vinyals and Le, 2015). NLG is the task of au-

tomatically producing unstructured text in natural lan-

guages that humans can understand (Gatt and Krah-

mer, 2018), and the text-to-text approach only needs

unstructured text as their input. Previous studies show

that these neural generation models can produce very

natural and ﬂuent utterances, leading to the creation

of human-like chatbots (Tao et al., 2018; Zhou et al.,

2020).

The main advantage of these neural generation

models is the ﬂexibility to generate coherent and

unique responses to any given conversation context

(Yang et al., 2019). Likewise, writing domain spe-

ciﬁc rules or utterances is not required for these mod-

els, which reduces human labour for domain adapta-

tions. These features combined make neural genera-

tion models more attractive than traditional methods.

The underlying mechanisms of such models, how-

ever, are still based on statistical generalizations. This

results in generating stylistically intelligent-looking,

grammatically correct texts, but the models lack any

actual understanding or meaning behind what is gen-

erated (Floridi and Chiriatti, 2020). That is to say, the

way these models work is not concerned with whether

the generated content is semantically right or wrong.

Moreover, the generalization process may also result

in human biases, foul language usage, and hate speech

(Caliskan et al., 2017), because the models have been

trained on large corpora from the internet without any

interception from a human annotator. Consequently,

the models carry the risk of producing misinforming,

harmful and offensive content such as racist and sexist

slurs as well as implicit discriminatory text (Bender

et al., 2021; Schlesinger et al., 2018).

3 HIGHLY CONTROLLED

CHATBOTS

Many scientiﬁc projects focus on developing chatbots

for highly sensitive domains (Xu and Zhuang, 2020).

Most commercial implementations or real-world ap-

plications, however, tend to use traditional meth-

ods such as rule-based or retrieval-based approaches

(Chen et al., 2017; Gatt and Krahmer, 2018). The

main advantage of these approaches comes from the

use of human-authored utterances. In the retrieval-

based approach, the chatbot’s responses are selected

from corpora of utterances, based on their relevance

to the corresponding conversation context. In rule-

based systems, the dialogue ﬂows are deterministic;

from the start of the conversation until the end, every

exchange with the chatbot is pre-scripted. These as-

pects not only make the chatbots speak with natural

human utterances, but they also make it easier to keep

them under control (Gatt and Krahmer, 2018).

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

402

On the one hand, traditional methods do not carry

any risk of exposing users to harmful contents be-

cause the responses are carefully crafted by humans.

On the other hand, these systems often have scala-

bility issues and labour-intensive domain adaptations.

Furthermore, due to the ﬁnite number of utterances,

dialogues with these chatbots may become repetitive.

Issues like repetition in conversations is disliked by

humans and can weaken the user engagement (See

et al., 2019). Consequently, users stop talking to the

chatbot after one session in a multi-session, long-term

setup.

In comparison, neural generation models do not

share the limited corpora and repetitiveness issues of

traditional approaches. However, the risk of harmful

content generation is an expensive problem for highly

sensitive domains. The currently most acclaimed of

these models, namely GPT-3 (Brown et al., 2020),

has been tested in medical support system scenarios.

It has produced convincingly well-written texts that

provide inconsistent and unreliable information, and

incidentally its suggestions become inappropriate for

medical applications (Rousseau et al., 2020). As a

result, we cannot yet fully embrace an uninterrupted

implementation of open-domain pre-trained genera-

tion models for chatbot applications in highly sensi-

tive domains.

4 HYBRID SOLUTION

We argue that the solution is a careful combination

of all approaches, in which the dialogue ﬂow is de-

termined by domain experts. Previous studies (Yang

et al., 2019; Song et al., 2018) have demonstrated po-

tential beneﬁts of combining retrieval and generation

approaches. Following these studies, we propose a

hybrid architecture consisting of (1) a retrieval model

based on human-authored utterances, (2) multiple

open-domain neural generation models pre-trained on

various online datasets, (3) a dialogue state tracker,

and (4) a neural response selection model trained on

manually annotated data.

Our proposed method starts with domain experts

who create clusters of chatbot utterances and assign

each cluster to a dialogue state. The retrieval model

is responsible for selecting the highest ranked utter-

ances, but only from the cluster corresponding to the

current dialogue state. Meanwhile, the neural gen-

eration models generate multiple response candidates

for the given conversation context. Then, the system

is completed with the response selection model, se-

lecting the best ﬁtting candidate from a set of both

human-authored and generated response candidates.

For each conversation context, both human-

authored and automatically generated response can-

didates are taken into account, and the best ﬁt-

ting response is uttered by the chatbot. However,

the dialogue state tracker contains a set of rules to

block the generation models under certain circum-

stances. For instance, at the end of each dialogue

state, human-authored utterances are prioritized to ac-

complish seamless transitions between the states.

The dialogue states and human-authored utter-

ances may provide the domain experts control over

the dialogue ﬂow, similar to pre-scripted systems.

Meanwhile, the inclusion of generation models and

the possibility of delivering automatically generated

responses increase the variation of utterances. Like-

wise, they add the value of responding to unexpected

conversation contexts. Hence, with generation mod-

els, we explore the possibilities to counteract the main

limitations of pre-scripted systems, such as repeti-

tiveness and weak user engagement. Moreover, cur-

rent open-domain generation models are advanced

enough to provide the basic capabilities that are ex-

pected from chatbots in general, such as chit-chatting,

without any human-authoring. This means that the

labour of the domain experts is focused only on craft-

ing domain-speciﬁc utterances, without having to deal

with the chatbots’ capacity of handling basic interac-

tions.

Because of the highly sensitive nature of the per-

sonal health domain, the open-domain language mod-

els need to be constrained. From where we stand,

the extent to which this is necessary is an empirical

problem. Nonetheless, solving it is essential. We

hope to reduce the chance of mishaps by implement-

ing checks at 3 different places; (1) dialogue state

tracking, (2) harmful content ﬁltering, and (3) neural

response selection. The monitoring by the dialogue

state tracker reduces deviations from the domain-

speciﬁc track. The harmful content ﬁltering provides

a direct intervention to potentially harmful content as

soon as it is generated. Finally, the response selection

model is trained for selecting the response candidate

that is best ﬁtting to the goals of the highly sensitive

domain, consequently, ignoring the harmful contents.

5 METHODOLOGY

Here we discuss the details of our proposed hybrid

framework for long-term multi-session counselling

chatbots. The framework consists of four modules,

explained below. The connection between these mod-

ules is displayed in Figure 1.

Hints of Independence in a Pre-scripted World: On Controlled Usage of Open-domain Language Models for Chatbots in Highly Sensitive

Domains

403

Figure 1: An overview of the pipeline within our proposed hybrid architecture.

• Dialogue State Tracker: This module is respon-

sible for tracking the progress of the dialogue.

Each state s ∈ S corresponds to a pre-deﬁned ques-

tion, and possible reﬂections and follow-ups.

• Retrieval Module: This module focuses on re-

trieving a set of response candidates, from a pre-

constructed dataset, R, consisting of chatbot utter-

ances. The retrieval algorithm matches the con-

text u

of the i-th conversation, with pre-scripted

utterances, R

⊂ R, given the dialogue state s. It

then returns the highest ranked response candi-

dates set R

for the given context.

• Generation Module: We propose an ensemble

of three pre-trained open-domain generation mod-

els trained on a diversity of training data. Given

the conversation context, u

, each language model

generates multiple response candidates. After we

apply a ﬁltering process to eliminate harmful con-

tent, the highest ranked candidates are collected in

a single set, G

• Response Selection Module: Given the gener-

ated and retrieved response candidates for the i-th

conversation, G

and R

, this module aims to select

the best response as the ﬁnal chatbot output r

. To

do so, all candidates, X

= G

∪R

, are ranked with

a neural network trained on data obtained by man-

ual annotations.

5.1 Dialogue State Tracker

It is important for human counsellors to collect in-

formation about their patients during an interview.

Hence, these counsellors mostly ask questions at cer-

tain times, listen to the patient’s answers and follow

up by giving reﬂections on them. In this context, we

can deﬁne the relevant chatbot responses as correctly-

timed questions, and appropriate reﬂections on the

user answers. The dialogue state tracker is a rule-

based model that aims to replicate this highly orga-

nized aspect of human-human counselling.

The dialogue states, in this project, represent

micro-dialogues that we want the conversational

agent to conduct at one point in the conversation.

Each dialogue state, s ∈ S, has corresponding poten-

tial questions, reﬂections, and follow-ups. Assigning

at least one human-authored utterance to each state is

required to guide the dialogue ﬂow. In the context of

counselling, we can expect a state to start with a coun-

selling question. For instance, the question “How

many cigarettes do you usually smoke in a day?” can

be asked with the hope that it initiates a discussion be-

tween the chatbot and the user on the user’s frequency

of smoking.

We control when a state ends and another begins

via the rules in the dialogue state tracker. We mon-

itor which states have been used in the conversation

and what user information is not available yet to de-

cide on the next dialogue state. Within a state, the

micro-dialogue between the chatbot and the user can

be as short as a single “question-answer-reﬂection”

triplet, or the reﬂection can be followed by more ex-

changes until the conversation reaches a limitation

preset. Moreover, we implement an intent classiﬁer

to detect whether the user expresses explicit interest

in proceeding to the next state. If the limit is reached

or the user’s interest is detected, the chatbot proceeds

with the question of the next dialogue state.

5.2 Retrieval Module

The information retrieval-based conversational mod-

els aim to return the appropriate response to the user’s

ﬁnal query from pre-constructed conversation cor-

pora. Unlike the common retrieval-based approaches,

the pre-constructed dataset in this project only con-

sists of conversational agent utterances which are

hand-crafted by the subject domain experts.

We divide the dataset into subsets corresponding

to the dialogue states. For a given dialogue state s,

there is a subset of the chatbot utterances, R

⊂ R,

containing the potential questions, reﬂections, and

follow-ups. The utterances in R

can only be pre-

sented to the user during the dialogue state s. Hence,

at the i-th conversation with a given dialogue state s,

the set of highest-ranked response candidates R

can

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

404

only be selected from the dialogue state utterances set

, and so making R

⊂ R

The retrieval of the utterance candidates set R

can

be performed by an information retrieval algorithm,

which is BM25 in our case. For each conversation

context u

during the dialogue state s, the retrieval al-

gorithm matches u

with the pre-scripted utterances

in the set R

. The retrieval algorithm then returns the

most relevant k number of utterances, R

5.3 Generation Module

The retrieval-based approach assumes that the appro-

priate response is pre-existing in the conversation cor-

pora. The generation-based approach, however, can

construct coherent and unique responses even for an

unseen conversational context. In this study, we ex-

plore the potential of 3 state-of-the-art pre-trained

open-domain neural generation models, namely GPT-

3, BERT, and DialoGPT.

In order to guide the pre-trained models to pro-

duce responses that are relevant and speciﬁc to the

conversation context, we implement a prompt gener-

ator that prepares the input to be given to the gener-

ation models. For each conversation context u

, the

prompt p

is generated from: (1) the combination of

the conversation context, u

, (2) the retrieved candi-

date utterances, R

, and (3) a selection of informa-

tion from the user database. Evidently, the pre-trained

models do not only take the conversation context into

account. Their generation process is also inﬂuenced

by the retrieved candidate utterances, R

, which are

hand-crafted by the domain experts. By this addition,

we anticipate that the generation models will adapt

to a style similar to the human experts (Song et al.,

2018).

In the long-term interaction setup, ﬁtting the entire

dialogue history within the vector of the conversation

context u

would become unfeasible. To preserve the

coherence of the dialogue, we explore the outcomes

of feeding the generation models with a distilled ver-

sion of the history. Hence, we include a selection of

user information extracted from the dialogue.

As explained in Section 2, employing pre-trained

generation models without any precautions carries the

risk of exposing users to harmful contents. We accept

that we are unable to know and control what these

generations models may generate. Hence, we propose

to implement a classiﬁer to detect harmful content

and eventually eliminate such cases from the response

candidate sets. We see this implementation as the ﬁrst

necessary step towards keeping the pre-trained mod-

els in check, with the hope that the constrained setting

reduces the chance of mishaps.

Given the prompt p

for the i-th conversation, each

pre-trained model generates a response candidate set;

GPT 3

, G

BERT

, and G

DialoGPT

. We combine the sets of

response candidates to create the set of all generated

response candidates G

= G

GPT 3

∪ G

BERT

∪ G

DialoGPT

Finally, we remove the unwanted content detected by

the harmful content classiﬁer from the response can-

didates set G

before the set is sent to the response

selection module.

5.4 Response Selection Module

For each i-th conversation, the retrieved response can-

didates R

and the generated response candidates G

are combined into the set of all response candidates

= R

∪ G

. The response selection module ranks all

response candidates in X

and returns the highest scor-

ing candidate r

as the ﬁnal response which should be

uttered by the conversational agent.

Following previous work on domain-adapted

multi-turn response selection (Li et al., 2021; Gu

et al., 2020; Whang et al., 2020), we adopt a recur-

rent neural network approach based on the pre-trained

BERT model. The BERT model is mainly used to en-

code the utterances in the conversation context u

and

the response candidates in X

. The model is trained

with the next utterance prediction strategy on a corpus

that has been manually created by the domain experts.

Likewise, the BERT model is ﬁne-tuned for domain

adaptation on the same dataset.

6 RELATED WORK

Chatbots in personal health-related domains typically

make use of pre-scripted approaches. He et al. (2022)

designed the dialogue ﬂow of their smoking coun-

selling chatbot on the basis of commercial software

by using rules, human-authored utterances and intent

classiﬁers. Brixey et al. (2017) used a classiﬁer to

select the highest ranked response candidate from a

database of linked questions and answers for their

HIV counselling chatbot. They addressed the repeti-

tion problem by randomly skipping the highest ranked

candidate. Likewise, Denecke et al. (2021) built their

mental health support chatbot on a system that re-

trieves response candidates from a human-authored

utterances corpus, based on their semantic and syn-

tactic similarities to the user input.

Recently, natural language generation approaches

have become popular for health-related chatbot ap-

plications and for other domains. Saha et al. (2021)

trained a variety of classiﬁers, based on a Seq2Seq ar-

chitecture and reinforcement learning applications, on

Hints of Independence in a Pre-scripted World: On Controlled Usage of Open-domain Language Models for Chatbots in Highly Sensitive

Domains

405

a manually annotated dialogue corpus for their mental

health chatbot application.

Methodology-wise, we follow the recent work

that has been done on hybrid retrieval-generation

models. Song et al. (2018) created an ensemble of a

retrieval-based model and a generation-based model.

They proposed multi-seq2seq as their generation-

based model, which incorporated the highest ranked

retrieval-based candidates into the generation process.

Likewise, Yang et al. (2019) proposed another hy-

brid retrieval-generation conversation model, where

the generation model also takes the relevant facts re-

trieved from external sources.

7 DISCUSSION

In the best-case scenario, the generation module pro-

duces responses that are more relevant and speciﬁc

to the conversation context than the human-authored

responses. This way, the response selection mod-

ule will favour the generated responses, resulting in

a unique experience per user. The human-authored

counselling questions are still enforced by the dia-

logue state tracker, but only at the beginning of each

state. The role of the domain experts becomes guiding

the dialogue by initiating the discussion points and

asking questions. They can set up the dialogue states

and author the ﬁrst utterances, but the rest of the dia-

logue would consist of generated responses.

Ultimately, the performance of the system relies

on two properties; the performance of the generation

module and the performance of the response selection

module. In the case that only the generation mod-

ule underperforms, the generated responses would be

ignored by the response selection module, causing

the system to only follow the dialogue states and the

human-authored utterances. The chatbot would be-

come a pre-scripted chatbot with a retrieval-based re-

sponse selection.

If only the response selection module underper-

forms, the selection between the human-authored and

generated responses would be arbitrary. Given that

the dialogue state tracker is still in control of the di-

alogue ﬂow and interrupts based on a limited num-

ber of exchanges, the chatbot could turn into its pre-

scripted version.

The case where both the generation module and

the response selection module underperform deﬁnes

our worst-case scenario. The response selection mod-

ule may not be able to distinguish an appropriate re-

sponse from a harmful one. Hence, combined with

an underperforming generation module, the risk of

harmful content exposure may reach levels that are

intolerable for highly sensitive domains.

We propose not replacing the pre-scripted dia-

logue ﬂows completely, but enhancing them by in-

cluding open-domain NLG models. Hence, this sys-

tem does not reduce the labour of manually authoring

domain-speciﬁc utterances. Our proposed approach

can be seen as an improvement over pre-scripted chat-

bots in an attempt to counteract their main limita-

tions. From that perspective, it reduces the repetitive-

ness, increases the ﬂexibility, and handles mundane

chatbot tasks without any human labour. Addition-

ally, we acknowledge that the current technological

advancements cannot fully replace a human counsel-

lor. Keeping that in mind, our efforts aim to create

a support system that people in need may hopefully

beneﬁt from.

To understand how effective our approach is, we

plan to evaluate it by human evaluation methods, as

we prioritize understanding the engagement of our

chatbot with humans (Novikova et al., 2017; van der

Lee et al., 2021). We are going to conduct a user

study where we measure naturalness, relevance, con-

sistency, quality, and enjoyment. Naturalness will be

measuring how natural sounding the utterances are,

while relevance will be based on relevancy of the ut-

terances to their corresponding conversation contexts.

Finally, we will ask the users about the overall consis-

tency and quality of the entire conversation, alongside

how much they enjoyed talking to the chatbot.

8 CONCLUSIONS

In this paper, we present a new approach for develop-

ing multi-session long-term counselling chatbots for

highly sensitive domains such as personal health. We

explore the usage of open-domain NLG models to in-

crease enjoyability and to counteract the main limita-

tions of pre-scripted chatbots. Our proposed solution

is a hybrid approach that consists of a dialogue state

tracker, a retrieval module, three open-domain NLG

models, and a neural response selection module.

In taking this approach, we aim for the best of two

worlds: we expect the retrieval model and the dia-

logue state tracker to establish a controlled and effec-

tive dialogue ﬂow, while the neural generation models

add variety and the ﬂexibility to respond to any con-

versation context.

ACKNOWLEDGEMENTS

This project is partly ﬁnanced by the Dutch Research

Council (NWO) with project number 406.DI.19.054.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

406

REFERENCES

Bender, E. M., Gebru, T., McMillan-Major, A., and

Shmitchell, S. (2021). On the dangers of stochastic

parrots: Can language models be too big? . In Proc.

of FAccT 2021.

Bommasani, R., Hudson, D. A., Adeli, E., Altman, R.,

Arora, S., von Arx, S., Bernstein, M. S., and et al.

(2021). On the opportunities and risks of foundation

models. CoRR.

Brixey, J., Hoegen, R., Lan, W., Rusow, J., Singla, K., Yin,

X., Artstein, R., and Leuski, A. (2017). SHIHbot:

A Facebook chatbot for sexual health information on

HIV/AIDS. In Proc. of the 18th Annual SIGdial Meet-

ing on Discourse and Dialogue. ACL.

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D.,

Dhariwal, P., Neelakantan, A., and et al. (2020). Lan-

guage models are few-shot learners. In NeurIPS.

Caliskan, A., Bryson, J., and Narayanan, A. (2017). Se-

mantics derived automatically from language corpora

contain human-like biases. Science.

Chen, H., Liu, X., Yin, D., and Tang, J. (2017). A survey on

dialogue systems: Recent advances and new frontiers.

SIGKDD Explor. Newsl.

Denecke, K., Vaaheesan, S., and Arulnathan, A. (2021). A

mental health chatbot for regulating emotions (sermo)

- concept and usability test. IEEE Transactions on

Emerging Topics in Computing.

Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K.

(2019). BERT: Pre-training of deep bidirectional

transformers for language understanding. In Proc. of

the NAACL 2019: Human Language Technologies.

Floridi, L. and Chiriatti, M. (2020). Gpt-3: Its nature, scope,

limits, and consequences. Minds and Machines.

Gatt, A. and Krahmer, E. (2018). Survey of the state of the

art in natural language generation: Core tasks, appli-

cations and evaluation. JAIR.

Gu, J.-C., Li, T., Liu, Q., Ling, Z.-H., Su, Z., Wei, S., and

Zhu, X. (2020). Speaker-aware bert for multi-turn re-

sponse selection in retrieval-based chatbots. In Proc.

of the CIKM’20. ACM.

Gururangan, S., Marasovi

c, A., Swayamdipta, S., Lo, K.,

Beltagy, I., Downey, D., and Smith, N. A. (2020).

Don’t stop pretraining: Adapt language models to do-

mains and tasks. In Proc. of the 58th ACL.

He, L., Basar, E., Wiers, R., Antheunis, M., and Krah-

mer, E. (2022). Can chatbots support smoking ces-

sation? a study on the effectiveness of motivational

interviewing on engagement and therapeutic alliance.

Manuscript submitted for publication.

Krahmer, E., Bosse, T., and Bruijn, G.-J. (2021). Chatbots

and health: General. The International Encyclopedia

of Health Communication.

Li, L., Li, C., and Ji, D. (2021). Deep context modeling

for multi-turn response selection in dialogue systems.

Information Processing & Management.

Mikolov, T., Sutskever, I., Chen, K., Corrado, G., and Dean,

J. (2013). Distributed representations of words and

phrases and their compositionality. In Proc. of the

26th International NeurIPS.

Novikova, J., Du

sek, O., Cercas Curry, A., and Rieser, V.

(2017). Why we need new evaluation metrics for

NLG. In Proc. of EMNLP 2017. ACL.

Peters, M. E., Neumann, M., Iyyer, M., Gardner, M., Clark,

C., Lee, K., and Zettlemoyer, L. (2018). Deep contex-

tualized word representations. In Proc. of NAACL.

Rousseau, A.-L., Baudelaire, C., and Riera, K.

(2020). Doctor gpt-3: Hype or reality?

https://www.nabla.com/blog/gpt-3/. Accessed:

2021-10-18.

Saha, T., Chopra, S., Saha, S., Bhattacharyya, P., and Ku-

mar, P. (2021). A large-scale dataset for motivational

dialogue system: An application of natural language

generation to mental health. In IJCNN 2021.

Schlesinger, A., O’Hara, K., and Taylor, A. S. (2018). Let’s

talk about race: Identity, chatbots, and ai. In ACM

Conference on Human Factors in Computing Systems.

See, A., Roller, S., Kiela, D., and Weston, J. (2019). What

makes a good conversation? how controllable at-

tributes affect human judgments. In Proc. of NAACL:

Human Language Technologies. ACL.

Song, Y., Li, C.-T., Nie, J.-Y., Zhang, M., Zhao, D., and

Yan, R. (2018). An ensemble of retrieval-based and

generation-based human-computer conversation sys-

tems. In Proc. of IJCAI-18.

Sutskever, I., Vinyals, O., and Le, Q. V. (2014). Sequence to

sequence learning with neural networks. In NeurIPS.

Tao, C., Gao, S., Shang, M., Wu, W., Zhao, D., and Yan,

R. (2018). Get the point of my utterance! learning

towards effective responses with multi-head attention

mechanism. In Proc. of IJCAI-18.

van der Lee, C., Gatt, A., van Miltenburg, E., and Krahmer,

E. (2021). Human evaluation of automatically gener-

ated text: Current trends and best practice guidelines.

Computer Speech & Language.

Vinyals, O. and Le, Q. V. (2015). A neural conversational

model. In ICML Deep Learning Workshop.

Whang, T., Lee, D., Lee, C., Yang, K., Oh, D., and Lim,

H. (2020). An effective domain adaptive post-training

method for bert in response selection. In Proc. Inter-

speech 2020.

Xu, B. and Zhuang, Z. (2020). Survey on psychotherapy

chatbots. Concurrency and Computation: Practice

and Experience.

Yang, L., Hu, J., Qiu, M., Qu, C., Gao, J., Croft, W. B., Liu,

X., Shen, Y., and Liu, J. (2019). A hybrid retrieval-

generation neural conversation model. In Proc. of

CIKM’19. ACM.

Zhou, L., Gao, J., Li, D., and Shum, H.-Y. (2020). The

design and implementation of xiaoice, an empathetic

social chatbot. Computational Linguistics.

Hints of Independence in a Pre-scripted World: On Controlled Usage of Open-domain Language Models for Chatbots in Highly Sensitive

Domains

407