Negotiation Dialogue System Using a Deep Learning-Based Parser

Kenjiro Morimoto, Katsuhide Fujita and Ken Watanabe

Tokyo University of Agriculture and Technology, Tokyo, Japan

Keywords: Negotiation Dialogue System, Deep Learning, Parser.

Abstract: In recent years, there has been substantial research on negotiation dialogue agents. A notable study introduced

a method that decoupled strategy from generation using dialogue acts that encapsulated the intent behind

utterances. This approach has enhanced both the task success rate and the human-like quality of the generated

responses. However, the rule-based implementation of the parser limits the types of sentences it can process

for dialogue acts. Thus, this paper presents annotated training data based on the proposed dialogue acts and

introduces a deep learning-based parser. The deep learning-based parser achieved a dialogue act classification

accuracy of approximately 83% and effectively reduced the occurrence of unknown dialogue acts.

Additionally, negotiation dialogue systems using deep learning-based parsers have demonstrated improved

performance in terms of utility and fairness.

1 INTRODUCTION

Negotiation is a crucial skill in human

communication for resolving conflicts and achieving

beneficial agreements (Fisher et al ., 2011) (Lewicki

et al ., 2011). Recently, there has been extensive

research on negotiation dialogue systems, which

focus on creating intelligent dialogue agents capable

of negotiating with humans through natural language

(Zhan et al ., 2020) (Basave et al ., 2016). The goal of

these negotiation dialogue systems is to facilitate

conflict resolution and enhance mutual benefits by

generating context-appropriate negotiation dialogues.

One study on negotiation dialogue systems

presents a structural model that incorporates dialogue

actions, known as dialogue acts, into the natural

language understanding and generation process (He

et al ., 2018). This study proposed a framework

comprising three modules: a parser that transforms

input utterances into dialogue acts, a manager that

produces response dialogue acts based on dialogue

act history and the dialogue scenario, and a generator

that converts the dialogue acts generated by the

manager into natural language responses. This

approach differentiates the formulation of negotiation

strategies from the dialogue generation process,

enhancing human-like interaction and increasing task

completion rates. However, this framework has room

for improvement because its parser is rule-based and

unable to assign dialogue acts accurately based on the

meaning of natural language sentences.

The aim of this study is to introduce a more

accurate method for estimating dialogue acts using

deep learning for a parser that assesses dialogue acts

corresponding to input sentences. While rule-based

methods offer the benefits of high explainability and

ease of implementation, they struggle with data that

fall outside predefined rules, and it is challenging to

establish rules for all possible scenarios. In this paper,

we present innovative dialogue acts and a deep

learning-based method for a parser. We further

illustrate the effectiveness of the proposed method

through comparative experiments with previous

studies. The parser model is created by fine-tuning

several pretrained models using training data

annotated with the proposed dialogue acts. We

integrate the proposed parser along with the previous

method into negotiation dialogue systems and

conduct human-agent negotiation experiments with

participants to assess the deep learning-based parser.

Our findings indicate that a deep learning-based

parser can learn data features that a rule-based parser

cannot handle and can accurately infer dialogue acts

across a broader range of data. Furthermore, a

negotiation dialogue system that incorporates a deep

learning-based parser has been shown to enhance

performance in areas such as utility and fairness

indices.

Morimoto, K., Fujita, K. and Watanabe, K.

Negotiation Dialogue System Using a Deep Learning-Based Parser.

DOI: 10.5220/0013245200003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 1, pages 135-143

ISBN: 978-989-758-737-5; ISSN: 2184-433X

135

2 RELATED WORKS

2.1 Dialogue Act

A dialogue act is a structural model that represents the

actions in a dialogue. Its purpose is to classify each

utterance based on the speaker’s intention (Żelasko et

al ., 2021). Each dialogue act comprises an intent that

conveys the meaning of the utterance and an argument

(e.g., price). Dialogue acts are essential for

understanding spoken language, particularly in the

fields of linguistics and artificial intelligence. In

artificial intelligence, it is crucial to establish

appropriate dialogue acts and perform annotation

according to the type of dialogue data being processed.

Below are examples of dialogue acts found in the

negotiation dialogues examined in this study.

 Utterance = Hi, I'm interested in your bike.

 Dialogue Act = greet

 Utterance = I have it listed for $220.

 Dialogue Act = init-price (220).

2.2 Craigslist Negotiation Dataset

CRAIGSLISTBARGAIN is a dataset comprising price

negotiation conversations for items listed on Craigs-

list, an American classified advertising community

site (He et al ., 2018). In contrast to many earlier

negotiation dialogue datasets that were gathered from

limited dialogue domains, such as games (Lewis et al

., 2017) (Asher et al ., 2016), C

RAIGSLISTBARGAIN

includes negotiation dialogues that feature side offers

and casual discussions, offering scenarios that closely

resemble real-life negotiation settings.

In a two-party negotiation in C

RAIGSLISTBARGAIN,

two agents take on the roles of buyer and seller,

respectively. Each agent receives photos,

descriptions, and listed prices of items available on

Craigslist. The seller aims to negotiate a sale at the

listed price, while the buyer seeks to purchase at an

undisclosed target price. Either agent has the

discretion to make a price offer, which can be

accepted or rejected by the other party. Additionally,

agents have the option to terminate negotiations and

end the task without reaching an agreement. The

negotiation scenarios are centered around the six most

popular categories of Craigslist posts: housing,

furniture, cars and bikes, phones, and electronics.

This dataset consists of 6,682 person-to-person

conversations collected via Amazon Mechanical

Turk. We use the C

RAIGSLISTBARGAIN dataset for the

learning and evaluation experiments of the

negotiation dialogue system in this study.

2.3 BERT

Bidirectional Encoder Representations from

Transformers (BERT) is a natural language

processing model introduced by Google in 2018

(Devlin et al ., 2018). Unlike traditional pretraining

approaches that focus on unidirectional context,

BERT uses two pretraining methods to analyze

context in both directions. This pretraining process

uses a substantial collection of unlabeled sentences,

allowing for fine-tuning the model for various natural

language processing tasks simply by adding an output

layer. The paper reports enhanced accuracy across 11

different natural language processing tasks, and this

technology continues to be used today in numerous

applications, including search engines and chatbots.

2.4 Modular Framework

The objective of a negotiation dialogue system is to

analyze a series of utterances 𝑥



,𝑥



,….,𝑥



associated with a dialogue scenario 𝑐 and produce a

distribution for the response utterance 𝑥



. In this

research, we use a framework (He et al ., 2018) that

incorporates dialogue acts into the strategies of

traditional goal-oriented dialogue systems (Young et

al ., 2013). Figure 1 shows this framework.

The framework shown in Figure 1 consists of the

following three types of modules.

1. A parser that transforms input utterances into

dialogue acts by leveraging the dialogue history

𝑥



, the dialogue act history 𝑧



, and a

negotiation scenario 𝑐.

2. A manager that predicts a dialogue act 𝑧



as a

response to input, using the dialogue act history

𝑧



and the negotiation scenario 𝑐.

3. A generator that takes the predicted dialogue act

𝑧



from the manager and converts it, along with

the dialogue history 𝑧



, into a natural language

response 𝑥



2.4.1 Parser

The Modular framework focuses on dialogue acts,

which are composed of intents that convey the

purpose of an utterance and arguments tailored to

specific scenarios. For instance, the utterance “I have

it listed for $220” is categorized as a dialogue act with

the intent init-price and the argument price=200.

Dialogue acts serve as structural models that offer a

high-level understanding of a sentence rather than

aiming to capture its entire meaning.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

136

Figure 1: Modular framework (He et al ., 2018).

Consequently, this framework uses a rule-based

parser that uses regular expressions and if–then rules.

Rule-based systems process information based on

artificially established rules, granting them the

advantage of high explainability and straightforward

implementation. The parser identifies keywords from

the utterance and aligns them with predefined

keyword patterns. Matching rules are organized as a

sequential list. When multiple patterns match, the first

identified intent is chosen. If no patterns apply, an

unknown intent is returned. The intents used and their

corresponding matching patterns from previous

studies (He et al ., 2018) are presented in Table 1.

2.4.2 Manager

The manager’s role is to identify the appropriate

dialogue act for an utterance that the agent should

choose at each time step 𝑡, considering the history of

dialogue acts 𝑧



and the dialogue scenario 𝑐. The

manager is trained through supervised learning,

enhanced by an attention mechanism, and also

incorporates reinforcement learning with three

distinct reward functions.

In supervised learning, we aim to maximize the

likelihood of the training data based on the provided

dialogue act history 𝑧



and dialogue scenario 𝑐 ,

while also learning the transition probability

𝑝



𝑧



|𝑧



,𝑐. During the agent’s listening turn, the

Long Short Term Memory (LSTM) encodes the

incoming dialogue acts. Conversely, during the

agent’s speaking turn, a different LSTM decodes the

tokens in the dialogue act. The hidden layer’s state is

maintained across conversations to ensure a

comprehensive dialogue history.

In reinforcement learning, three reward

functions—Utility, Fairness, and Length—are

optimized using the policy gradient method. Utility is

the agent’s self-interested objective, designed as a

linear function of the final interaction price. Buyers

achieve a utility of 1 at the target price, while sellers

reach a utility of 1 at the list price. Additionally, both

agents have zero utility at the midpoint between the

Table 1: Intent used in the previous study (He et al ., 2018).

Intent Matching Patterns

intro

Hi, hello, hey, hiya, howdy, how are

you, interested

inquiry

starts with an interrogative word (e.g.,

what, when, where) or particle (e.g.,

do, are)

inform previous dialogue act was inquire

init-price first price mention

vague-price

No price mention and comedown,

highest, lowest, go higher/lower, too

high/low

counter-price New price detected

insist

The same offer as the previous one is

detected

disagree No, not, n't, nothing, don't

agree

Not disagree and, ok, okay, great,

perfect, deal, that works, I can do that

unknown Does not match any rule

list price and the buyers’ target price. Fairness focuses

on equalizing the utilities for both buyers and sellers.

Length measures the number of utterances in a

dialogue, promoting the agent to maintain the

conversation for as long as possible. If no agreement

is made, the reward assigned is −1.

2.4.3 Generator

The generator’s primary function is to transform

dialogue act predictions (𝑧



) made by the manager

into natural language utterances ( 𝑥



). Previous

research has used a search-based approach to

implement the generator. The search-based approach

leverages a database of templates generated from the

training dataset’s utterances, which have been

analyzed by a parser. Each utterance is converted into

a template by replacing specific words with

placeholders based on the corresponding dialogue act.

For instance, the utterance “Would you take $705 for

it?” is transformed into the template “Would you take

[price] for it?” by substituting the numerical value

“$705” with the placeholder [price].

Natural language utterances are generated by

assessing the similarity between the template context

and the current context. We represent each context as

a BOW vector weighted by TF-IDF. Similarity is then

calculated by taking the dot product of the two

context vectors. To enhance the diversity of generated

utterances, we select one utterance from the top K

candidates guided by a distribution derived from a 3-

gram language model trained on the training data.

Negotiation Dialogue System Using a Deep Learning-Based Parser

137

3 PROPOSED DIALOGUE ACTS

AND DEEP-LEARNING BASED

PARSER

3.1 Proposed Dialogue Acts

Table 2 presents the results of classifying utterances

from the C

RAIGSLISTBARGAIN dataset using a rule-

based parser from a previous study (He et al ., 2018).

The table shows that the rule-based parser assigns

“unknown” intent to approximately 25% of the

utterances. This indicates that the parser is unable to

classify these utterances. In the negotiation dialogue

system framework, sentence understanding and

generation rely on dialogue acts. The high “unknown”

intent rate signifies that approximately 25% of the

generated responses may not accurately reflect the

intended dialogue act. This indicates a potential

weakness in the system’s ability to effectively

understand and respond to user input.

To enhance the classification accuracy of utterances

and assign them to appropriate intents, we introduce

two new intents: “supplemental“ and “thanks.” The

“supplemental” intent signifies the provision of

additional information that contributes to the

negotiation process. It may encompass detailed item

descriptions, personal stories, or other relevant details.

In negotiation dialogues, supplemental information

often plays a crucial role in achieving a favorable

outcome. For instance, highlighting an item’s

appealing features and associated benefits can

increase the price, while disclosing one’s financial

situation and reasons for wanting to purchase can lead

to a price reduction. While supplemental explanations

in response to partner inquiries are categorized as

“inform,” spontaneous information sharing is often

classified as “unknown.” Introducing a supplemental

dialogue act allows for utterances unrelated to the

negotiation but supportive of it. For example,

“Thanks” signifies gratitude toward the partner.

Because negotiation dialogues involve human

communication, they do not necessarily conclude

immediately after agreement. In many instances, after

expressing intent to reach an agreement, expressions

of gratitude toward the partner are observed. Existing

intent classifications lacked the ability to capture

these expressions of gratitude, and most of them were

classified as “unknown.” By adding “thanks,” we can

address communication aspects beyond negotiations.

Below are examples of utterances classified as “supp-

lemental” and “thanks”.

Table 2: Intent classification with a rule-based parser.

Intent

# of

utterances

% of total # of

utterances

unknown 9592 24.793

counter-price 7738 20.001

inquiry 5056 13.069

init-price 4629 11.965

intro 4611 11.918

inform 2321 5.999

disagree 2027 5.239

agree 1896 4.901

insist 432 1.117

vague-price 386 0.988

Total 38688 100

 Example of Supplemental

Utterance = I can afford to pay $72 for it.

Dialogue Act = init-price

Utterance = This is antique, so although it is

used, it is a very good bookcase.

Dialogue Act = supplemental

 Example of Thanks

Utterance = Ok, I can accept $12.

Dialogue Act = agree

Utterance = Great, thanks!

Dialogue Act = thanks

In this study, we use a total of 12 intents, including

those listed in Table 1, along with the additional

intents “supplemental“ and “thanks.”

3.2 Annotation

Our study requires training data to develop a deep

learning model capable of estimating intent. To

prepare these data, seven individuals, encompassing

both members of the general public and university

students, annotated each of the 5,987 dialogues in the

RAIGSLISTBARGAIN dataset. To maintain consistent

annotation quality across workers, we established

classification criteria for each intent. Utterances in the

dataset were then read according to these criteria to

determine their respective intents. To enhance

efficiency and minimize typographical errors, aliases

were assigned to intents. Annotators used these

aliases during the process, with subsequent

conversion of all aliases to their original intents after

annotation completion. The classification criteria for

intents in annotation are shown in Table 3.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

138

Table 3: Classification criteria for annotation.

Intent (Alias) Classification criteria

intro (g)

Utterances that don't include a price

indication and include words of greeting

(e.g., hi, hello, hey, good day) or

indicating a willingness to negotiate

(e.g., interested in, do you have any

question?, I want to buy)

Inquiry (q)

Utterances beginning with an

interrogative word (e.g., what, when,

where) or particle (e.g., do, are), or

question to a partner

Inform (f) Responses to questions

init-price (p)

Utterances that include a price indication

and are the first price offer in the

dialogue

vague-price (v)

Price negotiation utterances that ask for a

price increase or reduction without a

price listed

counter-price (c)

Price offer utterances that includes a

price indication, follows "init-price", and

don't fall under "insist", "agree", or

"disagree"

insist (i)

Utterances proposing the same price as

the previous one, or asserting the

legitimacy of the price proposed

disagree (d)

Utterances to decline a price offer from a

partner

agree (a)

Utterances to accept a price offer from a

partner

supplemental (s)

Utterances that provide supplementary

information to advance negotiations,

such as detailed information of an item,

and don't fall under "inform"

thanks (t)

Utterances expressing gratitude to a

partner (e.g., thanks, thank you)

unknown (u)

Utterances that cannot be classified into

any intent

Following the completion of annotation by each

worker, the content undergoes a review and

correction process. Each utterance is annotated by

one worker and subsequently checked and corrected

by a separate worker. Once all utterances have been

annotated and reviewed, they are consolidated to

form the new training data.

3.3 Learning Framework:

Deep Learning-Based Parser

In this study, we propose a deep learning-based parser

by fine-tuning BERT, a prominent Transformer

encoder model (Vaswani et al ., 2017) that is well-

suited for text classification tasks.

Figure 2: Inference process flow of a deep learning-based

parser.

For the implementation using Hugging Face

Transformers, Bert-base-uncased was chosen as the

pretrained model, with the proposed dataset serving

as the training data. The tokenizer associated with the

selected model was used to tokenize the training data.

Recognizing the interactive nature of dialogue,

including negotiation dialogue where utterances

frequently rely on preceding partner statements, the

tokenizer input was structured as a two-sentence input.

This approach considers not only the utterance to be

classified but also the immediately preceding

utterance. Utterance flow is segmented into

individual dialogues. For the initial utterance in a

dialogue, the [PAD] token is provided as the

preceding utterance. Fine-tuning is then performed

using the Trainer class. Figure 2 shows the inference

processing flow used by the parser in this study. To

facilitate performance comparison, parsers based on

ALBERT (Lan et al ., 2020), DistilBERT (Sanh et al .,

2019), and RoBERTa (Liu et al ., 2019), all

derivatives of BERT, were also developed. For these

implementations, albert-base-v1, distilbert-base-

uncased, and RoBERTa-base were selected as the

respective pretrained models.

4 PERFORMANCES OF DEEP

LERANING-BASED PARSERS

4.1 Inference Using Deep

Learning-Based Parsers

We fine-tuned four pretrained models (BERT,

ALBERT, DistilBERT, and RoBERTa) using

annotated training data and evaluated the resulting

models through cross-validation.

Tables 4 and 5 present the performance

comparison results of fine-tuning for each pretrained

model, highlighting the best results for each

evaluation metric in bold. Table 4 shows that all

pretrained models achieved a classification accuracy

rate of approximately 83%. RoBERTa exhibited the

Negotiation Dialogue System Using a Deep Learning-Based Parser

139

Table 4: Parser performance comparison.

Model Accuracy (%) Train runtimes (s)

BERT 83.456

1.451 × 10



ALBERT 83.196

0.898 × 10



DistilBERT 82.959

𝟎. 𝟖𝟏𝟎 × 𝟏𝟎

𝟒

RoBERTa 83.836

1.563 × 10



Rule-base 43.960 -

Table 5: Intent classification evaluation by F1 score.

BERT ALBERT DistilBERT ROBERTa

# of

Utterances

intro 0.886 0.885 0.881 0.888 1426

inquiry 0.906 0.905 0.902 0.906 1877

inform 0.900 0.895 0.896 0.900 1427

init-price 0.833 0.833 0.822 0.830 1139

vague-price 0.618 0.627 0.596 0.622 316

counter-price 0.863 0.849 0.853 0.864 1855

insist 0.174 0.115 0.131 0.187 111

disagree 0.535 0.505 0.502 0.540 152

agree 0.837 0.824 0.833 0.840 1147

supplemental 0.596 0.594 0.577 0.591 380

thanks 0.758 0.775 0.766 0.767 363

unknown 0.302 0.248 0.260 0.294 95

macro avg 0.684 0.671 0.668 0.686 10288

highest accuracy, surpassing DistilBERT, which had

the lowest accuracy, by approximately 0.9%.

Conversely, DistilBERT exhibited the shortest

training runtimes, approximately half that of

RoBERTa. ALBERT and DistilBERT, both

lightweight BERT variations, maintained accuracy

rates in 1% of the original BERT while considerably

reducing training runtimes. The BERT derivative

models had an accuracy variation of ±0.5%. However,

the lightweight versions, such as DistilBERT,

demonstrated significantly faster training times.

Therefore, for larger datasets, using ALBERT or

DistilBERT is recommended owing to their improved

efficiency.

Table 5 shows that RoBERTa achieved the

highest F1 score across all seven dialogue acts,

outperforming all other models. Notably, BERT

attained the highest F1 scores for supplemental and

unknown intents, ALBERT for vague-price and

thanks intents, and both BERT and ALBERT for init-

price. This indicates that the classification of easier

dialogue acts may be influenced by the specific

pretrained model used. Conversely, DistilBERT

consistently performed below all other models across

all evaluation metrics.

Examining individual dialogue acts, those with

over 1,000 data points consistently achieved F1

values exceeding 0.8 across all models. Conversely,

the remaining six dialogue acts with fewer data points

exhibited F1 values below 0.8, suggesting a decrease

in classification accuracy for less frequent dialogue

acts in negotiation dialogues. The F1 values for

“insist” and “unknown” were particularly low, at

0.187 and 0.302, respectively. “Insist” represents a

dialogue act aimed at reiterating a previous price offer.

Given the input method used in this study, which

involved inputting two sentences (the current

utterance and the previous utterance), it is possible

that this input method may have resulted in numerous

misclassifications owing to the inability to fully

capture contextual information. Because “unknown”

encompasses a collection of sentences that resist

classification, its features may not have been

adequately learned during training, potentially

contributing to the reduced classification accuracy.

Misclassifications of sentences that should be

classified as “unknown” into other dialogue acts may

occur because the model learns features beyond

human comprehension through deep learning,

thereby accurately classifying them into the correct

dialogue act. This presents a potential advantage of

deep learning-based parsers, particularly relevant to

negotiation dialogue systems. However, for the

“insist” intent, it is necessary to either increase the

number of training samples classified as “insist” or

merge it with other dialogue acts to increase

classification accuracy.

In conclusion, the pretrained models best suited

for deep learning-based parsers are ALBERT, which

demonstrates low computational complexity and

execution time alongside high classification accuracy

for the data-sparse dialogue acts of “vague-price” and

“thanks,” and RoBERTa, which exhibits the highest

overall classification accuracy.

4.2 Results of Inference Using Deep

Learning-Based Parsers

Table 5 presents the classification results of sentences

in the C

RAIGSLISTBARGAIN dataset using the deep

learning-based parser. For this parser, a model based

on RoBERTa, which achieved the highest accuracy

rate, was used. Comparing Table 5 with Table 1

reveals that our proposed dialogue acts successfully

reduced the proportion of sentences classified as

“unknown” from 24.793% to 0.772%. Both “counter-

price” and “inquiry” accounted for over 18% of the

total utterance count, representing a substantial

portion of all utterances. In the context of price

negotiation dialogues, “counter-price” is deemed

essential because price offers form the central theme

of the conversation. However, because all questions

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

140

Table 6: Intent classification using a deep learning-based

parser.

Intent # of utterances % of total # of utterances

unknown 9592 24.793

counter-price 7738 20.001

inquiry 5056 13.069

init-price 4629 11.965

intro 4611 11.918

inform 2321 5.999

disagree 2027 5.239

agree 1896 4.901

insist 432 1.117

vague-price 386 0.998

Total 38688 100

are currently classified under “inquiry,” there is

potential for further subdivision. “Disagree” and

“insist” constituted a relatively small proportion of

the total utterances, accounting for 1.068% and

0.677%, respectively. Owing to the tendency of deep

learning-based parsers to exhibit lower classification

accuracy for classes with limited training data,

integrating “insist” and “disagree” into other dialogue

acts would be advantageous. While integrating

“disagree“ poses a challenge owing to its frequent

occurrence in negotiation scenarios, “insist“ shares

considerable similarities with “counter-price” and

“vague-price,” suggesting a potential solution of

merging it into these two dialogue acts.

5 EVALUATIONS

5.1 Task

We evaluate our approach using the

CRAIGSLISTBARGAIN task (Section 2.2), where a

buyer and a seller negotiate the price of an item based

on the information listed on Craigslist. In this task,

the only argument in the dialogue act is the price.

5.2 Models

This study compares two models: a rule-based parser

commonly used in previous research and our

proposed deep learning-based parser. Both methods

are applied to parse the dataset incorporating the

newly annotated dialogue acts (Section 3.2).

The parser outputs are then used to generate

training and validation data for supervised learning,

along with n-gram models and utterance templates for

the generator. Subsequently, the parsed data are used

to perform supervised learning on the relationship

between utterances and dialogue acts, resulting in a

model denoted as SL. Finally, the pretrained SL

model is reinforced with the three reward functions

(Section 2.4.2), producing the models RL

utility

, RL

fair

and RL

length

. Our experimental setup encompasses a

total of eight models: SL and RL models using rule-

based parsers and SL and RL models using deep

learning-based parsers.

5.3 Evaluation Setup

Negotiation dialogue experiments were performed

using our model in a web application based on

previous work (He et al ., 2018). Figure 3 shows the

negotiation screen of this web application. Users are

presented with the scenario and item description in

the upper right corner and then interact with a

randomly selected model. Messages from the model

appear in the box located at the bottom left, and users

can input their reply messages in the chat box below.

Upon reaching an agreement, users input the final

offer price in the box labeled “Final Agreement” on

the right. If the partner submits a final offer, users can

either accept or reject it. Because the model does not

consistently generate perfectly context-appropriate

utterances, the dialogue may sound unnatural at times.

Should users encounter difficulty continuing the

dialogue, they can terminate the negotiation by

pressing the “Quit” button at the bottom left.

Ten subjects were recruited for the experiment.

Each subject participated in a total of 10 negotiation

dialogues, with all eight models being selected with

equal probability.

The experiment uses five evaluation indices. In

addition to the three indices used in reinforcement

learning—Utility, Fairness, and Length—we also

include Agreement Rate and Human-likeness.

Agreement Rate is the proportion of negotiations that

resulted in an agreement, which is the primary

objective in negotiation scenarios. It is calculated by

dividing the number of successful agreements by the

total number of dialogues performed for each model.

Human-likeness is an indicator of the model’s

human-like behavior during dialogue. After each

negotiation, users were asked, “Do you think your

partner demonstrated reasonable human behavior?”

They then rated the dialogue content on a 5-point

Likert scale.

Negotiation Dialogue System Using a Deep Learning-Based Parser

141

Figure 3: The negotiation screen on the web application.

5.4 Result

Table 7 presents the results of the human evaluation

experiment for the negotiation dialogue system,

grouped by optimization goal. The highest-

performing results in each group are highlighted in

bold. Table 8 presents an example of dialogue

between a human and the negotiation dialogue

system. SL(deep) demonstrated a more conscious

understanding of the dialogue flow and pursued

actions aligned with its own interests. As illustrated

in Table 8(b), SL(deep) not only responded

appropriately to the other party’s price offers but also

to other questions. In the context of price

negotiations, it did not simply present a price but also

attempted to persuade the user by providing

justifications for the offered price. This characteristic

is evident in the improvements observed in Utility

and Fairness, as presented in Table 7. However,

SL(rule) outperformed SL(deep) in terms of the

negotiation agreement rate across all three

optimization objectives. This is likely because Utility

and Fairness scores were lower for SL(rule), leading

to earlier compromise and agreement.

No considerable difference in human-likeness was

observed between SL(deep) and SL(rule). Both

models effectively generated dialogue involving price

offers or acting as questioners. However, when acting

as sellers, they struggled to respond appropriately to

questions about the item’s condition. Additionally,

with SL(deep), the accurate determination of dialogue

acts through deep learning sometimes leads to

inflexibility in utterance decisions. For instance, in the

initial exchange of Table 8(b), SL(deep) repeats “hello”

twice. This is because that deep learning has

established a pattern where an “intro” dialogue act

from the other party often elicits an “intro” response.

These limitations can be attributed to the search-based

generator. Search- based generators exhibit limited

utterance variation because all templates are derived

from a training dataset. Consequently, while the

dialogue act accurately captures the general structure

Table 7: Evaluation results of a human-assisted negotiation

dialogue system (highest scores in bold).

Ag↑ Ut↑ Fa↑ Len↑ Hu↑

# of

Exam

SL(rule) 0.85 -0.49 -0.98 12.31 3.4 13

SL(deep) 0.80 -0.44 -0.88 16.40 3.5 15

utility

(rule) 0.55 -0.35 -0.70 18.91 2.6 11

utility

(deep) 0.41 0.23 -0.46 13.25 2.4 12

fair

(rule) 0.86 -0.81 -1.62 13.21 3.1 14

fair

(deep) 0.73 -0.57 -1.14 13.27 2.8 11

length

(rule) 0.54 -0.16 -0.32 15.77 2.8 13

length

(deep) 0.75 -0.36 -0.72 11.92 3.4 12

Table 8: Example conversation between a human and a

negotiation dialogue system (model utterances in bold).

Title: New EC Technology Portable Bluetooth Speaker 6W

Listing Price: $15

Description: "EC Technology® 6W Dual-Driver Portable Bluetooth STEREO Speaker w/

Built in Cell Phone/Car Speakerphones -Black", "From the manufacturer:", "Incredible sound

quality surprising fullness realized through 2*3W driver, distinctive Cavity Design."

(a) A: Human (Buyer) B: SL(rule) (Seller)

B: hi, you're interested in the item? A: hi, i'm interested in it. how about $11?

B: not its too low A: hmm, how about your offer?

B: 15 A: ok, well, how about $13

B: i think i can do 12

A: OFFER($12) B: ACCEPT

Title: MODERN Contemporary White Queen Leather Bed Upholstery Tufted Italia

Listing Price: $349

Description: "This modern VERY BEAUTIFUL Queen bed is inspired by Italian design and

has an Asian rubberwood construction. A white, vinyl upholstery with a tall, square-tufted

headboard and clean lines highlight this elegant bed", "Included: One (1) headboard, one (1)

footboard, one (1) set of rails, three (3) slats", "Materials: Asian rubberwood", "Wood finish:

Black", "Upholstery materials: Vinyl, Faux L

(b) A: Human (Buyer) B: SL(deep) (Seller)

B: hello A: hello, i'm interested in this item.

B: hello A: do you have any information about this item?

B: very small stains mostly wear and tear A: how long have you used?

B: 3 months A: nice. well i want to buy it in $250. what do you say?

B: sorry there are 6 item even at 349 dollars that's a steel, less than 348 dollars a chair.

A: how about 265? B: lol

A: ok, how about 275? B: certification will be provided. 348 last price

A: how about 300? B: i can do 298.

A: OFFER($12) B: ACCEPT

of the utterance, the generator’s capacity for detailed

expression in natural language remains insufficient,

resulting in no observable improvement in human-

likeness.

Overall, the models were more inclined to make

compromises compared to humans, indicating that

they remain less effective at negotiation and

highlighting areas for further improvement.

6 CONCLUSIONS

In this study, we addressed three key tasks: creating a

dataset with new dialogue act proposals, developing

a deep learning-based parser, and performing

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

142

dialogue experiments with a negotiation dialogue

system using the deep learning-based parser.

Compared to traditional rule-based approaches,

the deep learning-based parser demonstrated impro-

ved accuracy in classifying utterances into their

correct dialogue acts while considerably reducing the

number of utterances classified as “unknown.”

Among the various pretrained models evaluated, Ro-

BERTa achieved the highest classification accuracy,

while ALBERT effectively minimized the decline in

accuracy while simultaneously reducing computati-

onal complexity and execution time. Moreover, in the

negotiation dialogue experiments, the system using

the deep learning-based parser exhibited enhanced

performance in terms of utility and fairness.

This paper primarily focused on enhancing the

parser component of the module framework using

dialogue acts. Consequently, the performance of the

dialogue act approach can be further optimized by

improving the remaining managers and generators. In

recent years, the LLM approach has emerged as the

dominant method for chatbots (Fu et al ., 2023) (Zhao

et al ., 2023). Therefore, we are currently exploring

the integration of LLMs as generators, incorporating

dialogue acts into the prompts (Wagner et al ., 2024).

Furthermore, we will perform dialogue experiments

comparing our proposed method with LLMs, aiming

to further demonstrate the value of the dialogue act

approach, which effectively captures the structural

outline of an utterance.

REFERENCES

Asher, N., Hunter, J., Morey, M., Benamara, F., &

Afantenos, S. (2016). Discourse Structure and Dialogue

Acts in Multiparty Dialogue: the STAC Corpus, In

Proceedings of the 10th International Conference on

Language Resources and Evaluation (LREC), pp.

2721-2727.

Cano-Basave, A. E., & He, Y. (2016). A study of the impact

of persuasive argumentation in political debates. In

Proceedings of the 2016 Conference of the North

American Chapter of the Association for

Computational Linguistics: Human Language

Technologies, pages pp.1405-1413.

Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019).

BERT: Pre-training of Deep Bidirectional

Transformers for Language Understanding, In

Proceedings of the 2019 Conference of the North

American Chapter of the Association for

Computational Linguistics: Human Language

Technologies (NAACL-HLT), p.4171-4186.

Fisher, R., Ury, W. L., & Patton, B. (2011). Getting to yes:

Negotiating agreement without giving in. Penguin.

Fu, Y., Peng, H., Khot, T., & Lapata, M. (2023). Improving

language model negotiation with self-play and in-

context learning from ai feedback. arXiv:2305.10142.

He, H., Chen, D., Balakrishnan, A., & Liang, P. (2018).

Decoupling Strategy and Generation in Negotiation

Dialogues, In Proceedings of the 2018 Conference on

Empirical Methods in Natural Language Processing

(EMNLP), pp. 2333-2343.

Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma, P.,

& Soricut, R. (2020). ALBERT: A Lite Bert for Self-

supervised Learning of Language Representations, In

Proceedings of the International Conference on

Learning Representations (ICLR).

Lewicki, R. J., Saunders, D. M., & Minton, J. M. (2011).

Essentials of negotiation. McGraw-Hill/Irwin Boston,

MA, USA.

Lewis, M., Yarats, D., Dauphin, Y. N., Parikh, D., & Batra

D. (2017). Deal or No Deal? End-to-End Learning for

Negotiation Dialogues, In Proceedings of the 2017

Conference on Empirical Methods in Natural

Language Processing (EMNLP), pp. 2443-2453.

Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D.,

Levy, O., Lewis, M., Zettlemoyer, L., & Stoyanov, V.

(2019) .Roberta: A Robustly Optimized BERT

Pretraining Approach, arXiv:1907.11692.

Sanh, V., Debut, L., Chaumond, J., & Wolf, T. (2019).

DistilBERT, a distilled version of BERT: smaller, faster,

cheaper and lighter, arXiv:1910.01108.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,

L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017).

Attention Is All You Need, In Proceedings of the 31st

International Conference on Neural Information

Processing Systems (NeurlPS), pp. 6000-6010.

Wagner, N., & Ultes, S. (2024). On the Controllability of

Large Language Models for Dialogue Interaction. In

Proceedings of the 25th Annual Meeting of the Special

Interest Group on Discourse and Dialogue, pp. 216-

221.

Young, S., Gašić, M., Thomson, B., & Williams, J. D.

(2013). POMDP-Based Statistical Spoken Dialog

Systems, In A review. In Proceedings of the IEEE,

Vol.101(5), pp. 1600-1179.

Żelasko, P., Pappagari, R., & Dehak, N. (2021). What Helps

Transformers Recognize Conversational Structure?

Importance of Context, Punctuation, and Labels in

Dialog Act Recognition, Transactions of the

Association for Computational Linguistics, Vol.9,

pp.1163-1179.

Zhan, H., Wang, Y., Feng, T., Hua, Y., Sharma, S., Li, Z.,

Qu, L., & Haﬀari, G. (2020). Let’s Negotiate! A Survey

of Negotiation Dialogue Systems. arXiv:2212.09072.

Zhao, W. X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y.,

et al. (2023). A survey of large language models.

arXiv:2303.18223.

Negotiation Dialogue System Using a Deep Learning-Based Parser

143