Agent AI with LangGraph: A Modular

Framework for Enhancing Machine Translation

Using Large Language Models

Jialin Wang1 and Zhihua Duan2

1
Executive Vice President,Ferret Relationship Intelligence
arXiv:2412.03801v1 [cs.CL] 5 Dec 2024

Burlingame, CA 94010, USA

jialinwangspace@gmail.com
https://www.linkedin.com/in/starspacenlp/
2
Intelligent Cloud Network Monitoring Department
China Telecom Shanghai Company,Shanghai, China
duanzh.sh@chinatelecom.cn

Abstract. This paper explores the transformative role of Agent AI

and LangGraph in advancing the automation and effectiveness of ma-
chine translation (MT). Agents are modular components designed to
perform specific tasks, such as translating between particular languages,
with specializations like TranslateEnAgent, TranslateFrenchAgent, and
TranslateJpAgent for English, French, and Japanese translations, respec-
tively. These agents leverage the powerful semantic capabilities of large
language models (LLMs), such as GPT-4o, to ensure accurate, contextu-
ally relevant translations while maintaining modularity, scalability, and
context retention.
LangGraph, a graph-based framework built on LangChain, simplifies the
creation and management of these agents and their workflows. It sup-
ports dynamic state management, enabling agents to maintain dialogue
context and automates complex workflows by linking agents and facili-
tating their collaboration. With flexibility, open-source community sup-
port, and seamless integration with LLMs, LangGraph empowers agents
to deliver high-quality translations.
Together, Agent AI and LangGraph create a cohesive system where
LangGraph orchestrates agent interactions, ensuring that user inputs are
analyzed, routed, and processed efficiently. Experimental results demon-
strate the potential of this system to enhance multilingual translation
accuracy and scalability. By highlighting modular design and automated
workflows, this paper sets the stage for further innovations in intelligent
machine translation services.

Keywords: Large Language Model · Agent · LangChain · LangGraph

· GPT-4o · GLM-4 · NLP · machine translation · sequence to sequence ·
PyTorch · NMT.
1 Introduction
In machine translation, it has always been one of the hotspots in deep learning
research how to choose models that are more efficient and more suitable for
translation. In recent years, professional researchers in this field of deep learning
and AI have been exploring and improving the relevant models for machine
translation, and a lot of experiments have been carried out repeatedly. Artificial
intelligence is developing, machine translation technology is also innovating and
developing, which makes machine translation move to a more advanced level.
Treisman and Gelade put forward the attention mechanism method [1], which
can simulate the attention model of the human brain, and the probability dis-
tribution of attention can be calculated to highlight the influence of one input
on the output. In short, when observing a picture, people first notice part of
the picture rather than browse the whole content. Then they adjust the focus
of attention sequentially in the process of observation. This method of attention
mechanism has a good optimization effect on the traditional model. Therefore,
in order to improve the system performance in a natural way, the attention
mechanism is applied to the sequence-to-sequence network model.
With the rapid development of large model technology, the field of Machine
Translation (MT) has made significant progress. Traditional machine transla-
tion methods relied on statistical models and rules, but in recent years, neural
network-based models, especially Large Language Models (LLMs) such as GPT-
4o, Llama 3.2, ERNIE-4, GLM-4, have become the focus of research due to their
powerful capabilities in understanding and generating natural language, demon-
strating excellent performance in machine translation tasks. LangGraph, as a
graph-based framework, offers a flexible and efficient way to construct agents
that can leverage the capabilities of large language models to perform complex
tasks, achieving high-quality machine translation. The research presented in this
paper not only provides a new technological path for the field of machine trans-
lation but also serves as a reference for other application areas that require
complex language understanding and generation capabilities.

2 Research Progress of Related Technologies

2.1 Machine Translation
Automatic translation or machine translation, its process is to realize the mu-
tual transformation of two natural languages. In other words, it researches on
how to realize the conversion of multiple natural languages through computer.
This is one of the most important research directions in the field of natural
language processing [3], and also one of the ultimate goals of artificial intelli-
gence. Machine translation technology is developing continuously, especially the
emergence of neural machine translation in recent years. The climax of the ap-
plication development of machine translation has arrived, which accelerates the
improvement of the quality of machine translation. In the field of Internet social
networking, cross-border e-commerce, tourism, and more vertical areas, machine
translation helps people overcome language barriers and meet the needs of big
data translation. Nowadays, with constant improvement, machine translation
technology is playing a vital role in political, economic, and cultural exchanges.

Phrase-based Statistical Machine Translation The core of statistical ma-

chine translation is as follows: after establishing a certain data of parallel corpus,
carries on the statistical analysis, so as to build a relevant model for machine
translation. Statistical machine translation steps are as follows: the first is the
language of constructing reasonable statistical models, the second definition to
estimate model parameters, the design parameter estimation algorithm.
Assuming that the letter y represents the source statement, the machine
translation model will look for a sentence m with the highest probability in the
target statement. m can be calculated by this equation:
m = arg maxe p(m | y) (1)

Network-based Neural Machine Translation Neural Machine Translation

(NMT) is one of the methods of machine translation. [5-7]. Compared with tradi-
tional phrase-based translation systems composed of many individually adjusted
components, NMT prepares to train and build a single large-scale neural net-
work, which can read sentences and output correct translation. [9].
At present, Machine translation is a hot research direction. It is a deep neural
network, which can be learned from corpus and composed of many neurons
through training, which is the process of translating between two languages, that
is, input the source language first, and then train the neural network to get the
target language. By simulating the human brain, the function of understanding
before translation can be achieved. The greatest advantage of this method is that
the translated sentences are more fluent, more grammatical, and easy for the
public to understand. Compared with other translation techniques, the quality
of this kind of translation has been significantly improved.

Recurrent Neural Network The most commonly used neural network is the
recurrent neural network (RNN), which can store the historical information of
the current word through learning. The probability of the next word’s occurrence
is calculated according to the whole context, so as to overcome the disadvantage
that the N-Gram language model cannot use the remote context information
in the sentence[10]. Recurrent neural network (RNN) is adopted by Socher in
syntactic parsing[11]. It is a typical three-layer deep learning model[12], which is
composed by Irsoy and Cardie. Recurrent neural network has been proved to be
effective in solving serialization problems, able to use context information and
has achieved good results in machine translation and other fields. However, the
problem of gradient explosion and disappearance exists in the process of solving
Recursive neural networks, which is not good enough for long text processing.
The long-term and short-term memory proposed later solves the long-sequence
problem well.
 The output layer, the input layer, and the hidden layer constitute the tradi-
tional neural network structure. There is a close connection between layers and
no connection between nodes. RNN means that the current output of a sequence
is related to the previous output. See Figure 1 for a detailed explanation.

Fig. 1. Typical machine learning process.

Attention Mechanism in Neural Networks Attention refers to the ability

of a person’s mental activity to focus on something. The attention mechanism
is based on the human visual attention mechanism. Simply speaking, the visual
attention mechanism means that when people observe a picture, they first notice
part of the picture, rather than the whole content, and then people will adjust
the focus of attention in the process of observation. Therefore, the attention
mechanism in the field of neural networks is derived from this principle.
In the ordinary encoder-decoder network structure, all the key information of
the input sequence is compressed into fixed-length vectors, but there is informa-
tion loss in the process of operation. When processing long sentences, especially
those over a fixed length, the performance of the encoder-decoder neural net-
work will be worse and worse, and the ability of the model will be gradually
reduced. As sentences get longer and longer, the performance of the encoder-
decoder neural network will also be gradually reduced. [7] To solve this problem,
Yoshua Bengio realized the translation from English to French by simulating the
attention mechanism in the neural network. Each output element in the out-
put sequence refers to the input sequence information by weight, so that the
alignment between the input sequence and the output sequence can be achieved.
Generate an output, then add a weight vector to the output as input to the
decoder. The attention mechanism is divided into stochastic attention mecha-
nisms and soft attention mechanisms[14]. Attention mechanism in deep learning
is more like memory access, in which all the details of the hidden state of the
input sequence need to be accessed before determining the weight size.
 In this experiment, a sequence-to-sequence network model consisting of two
layers of recurrent neural networks is used, and an attention mechanism is added
to the model.

2.2 Deep Learning and PyTorch

Deep learning is a part of machine learning, which is a representation-based

learning method in machine learning. It can imitate human brain structure,
efficiently process extremely complex input data, and extract abstract features
from data, so as to achieve the effect of intelligent learning[4]. The advantage of
in-depth learning is to use semi-supervised or unsupervised feature learning and
hierarchical features to extract effective algorithms instead of artificial feature
acquisition.
At present, in-depth learning frameworks that researchers are using include
TensorFlow, PyTorch, Caffe, Theano, Deeplearning4j, ConvNetJS, MXNet, Chainer,
and so on. In-depth learning frameworks are widely used in most areas, such as
natural language processing and speech recognition, and have achieved good re-
sults. The deep learning framework used in this experiment is PyTorch, which
is an artificial intelligence learning system launched by Facebook. Although the
underlying optimization is implemented on the basis of C, almost all the frame-
works are written in Python. Therefore, the source code looks concise and clear.
Compared with other frameworks, it can support GPU, create dynamic neural
networks, prioritize Python, etc. Compared with TensorFlow, PyTorch is more
concise and intuitive. Thereby, this framework is chosen in the experiment of
training the machine translation model.

2.3 LLMs in Machine Translation

In the field of natural language processing, large language models (LLMs) such
as GPT-4o, Llama 3.2, ERNIE-4, and GLM-4 have demonstrated exceptional
language processing capabilities due to their training on vast corpora and their
large number of parameters. These models are not only capable of understanding
complex human instructions and performing a variety of tasks, including machine
translation, but they also possess abilities such as in-context learning and chain-
of-thought, which allow them to optimize predictive results through contextual
information. Recent research indicates that advanced LLMs like ChatGPT have
surpassed traditional supervised learning models in certain language pairs. LLMs
show great potential in multilingual translation, bringing new opportunities and
challenges to the field of machine translation. Enhancing their capabilities in
multiple languages is crucial for establishing new translation paradigms.

2.4 LangGraph-Based Agent Translation Application

LangChain is an application development framework powered by Large Lan-

guage Models (LLMs), offering open-source building blocks, components, and
third-party integrations that enable developers to construct applications. Lang-
Graph is a framework built upon LangChain, simplifying the creation and man-
agement of agents and their runtime environments. In the application of machine
translation, LangGraph, through its state management system, allows agents to
maintain dialogue and remember contextual information when handling trans-
lation tasks. The translation agents can provide more coherent and contextually
relevant translation results, thereby enhancing the accuracy of translations.

3 Experimental design

3.1 Data Sources and Technical Routes

This experiment trains and tests the model with the help of English and French
datasets, whose form is English and French sentence pairs. After importing
the processed data into the language model package, the model is going to be
trained, and the machine translation model and evaluation model are estab-
lished. Then, the results are compared to continuously optimize the machine
translation model. Finally, the experimental results are obtained.

3.2 Data Text Algorithms

In the experiment, the data input sequence is defined as X, in which each word
is x, then an input sequence statement can be represented as (x1 , x2 , . . . , xT ).
The output statement sequence translated by the RNN model is defined as Y ,
then the output sequence statement can be represented as (y1 , y2 , . . . , yT ). The
sequence-to-sequence network used in this experiment is a neural network model
composed of 2 hidden layers, 256 hidden nodes, and 2 RNNs, which is used to
train processed data texts to achieve the effect of mutual translation between
English and French.

3.3 Model Processing

The Steps of the Experiment from Data Processing:

Step 1: Index the input and target statements in the network layer and store
the uncommon words in a dictionary separately.
Step 2: Convert the data file stored in Unicode character form into the file
stored in ASCII form, and make all content lowercase, while pruning most punc-
tuation symbols.
Step 3: Divide the files stored in the data into rows, then divide them into
pairs, which are in the form of English and French, and standardize the text
according to the length and content of the data.
Step 4: Import the data into the model for training.
This experiment uses Seq2seq model, which consists of two RN networks,
decoder and encoder. The function of the decoder is to read the vector and
produce the output sequence Y. The function of the encoder is to read the input
sequence X and output a single vector. The encoder is regarded as a network,
and the decoder as another network, moreover, the attention mechanism is added
into decoder. The model structure is shown in Figure 2.

Fig. 2. Sequence-to-Sequence Model Structure.

While training data, the source statement sequence (x1 , x2 , . . . , xT ) is first

input through the encoder. Then the decoder gives the first input statement a
start symbol < SOS > and takes the last The hidden layer of the encoder as its
first hidden layer, where the hidden layer ht behaves in time t:

ht = sigm(W hx xt + W hh ht−1 ) (2)

The output sequence generated by the RNN model is (y1 , y2 , . . . , yT ). The

process is accomplished by iteration of steps 2 and 3:

y = W yt ht (3)

3.4 Implementation of Agent Design Based on LangGraph

As shown in Figure 3, LangGraph is a framework for building agents, allowing

developers to create complex workflows by defining states, nodes, and edges.
This paper implements a multilingual translation system that can automatically
select the appropriate translation agent to translate the text based on the in-
put text that requires translation. The IntentAgent is responsible for parsing
the input text and intent, and selecting the suitable translation agent according
to the predefined mapping relationships. The translation agents include Trans-
lateEnAgent, TranslateFrenchAgent, and TranslateJpAgent, corresponding to
translation tasks in English, French, and Japanese, respectively. Each transla-
tion agent is an independent LangGraph agent that completes the translation
task by calling on the capabilities of the respective large language models to
perform translation services. For example, when TranslateFrenchAgent receives
a translation request, it calls the translation service to translate the text from
the source language into French. With a modular design and automated work-
flows, the LangGraph-based agent system can flexibly handle translation tasks in
different languages, providing a new solution for the field of machine translation.

Fig. 3. Multilingual Translation Agent.

4 Experimental results and analysis

4.1 Experimental results

The training results and translation results are obtained through the training
model. After processing the data, 75,000 words are finally selected for experi-
ment. By comparing the target sentences and the exact sentences, it can be seen
that the translation effect is still good.
The progress and average loss of training statements are shown in Figure 4.
 Fig. 4. Training Status Table.

Taking two simple sentences from Fig 5 to Fig 6 for example, the x and y
axes represent the words of the source and target sentences respectively. Each
pixel represents the weight size of the source statement to the target statement,
that is, the expression of attention mechanism. The right bar graph is weighted
from 0.0 to 1.0 (0.0 is black, 1.0 is white).

Fig. 5. Training Statement (1)

4.2 Experimental Analysis

The evaluation criteria of machine translation include BLEU [15], NIST [16],
METEOR [17] and TER [18]. The above criteria used to evaluate machine trans-
lation are all used to measure the quality of machine translation, the quality of
 Fig. 6. Training Statement (2)

quantitative model and the translation results by comparing machine transla-

tion with manual translation. BLEU was used as the evaluation standard in the
experiment to measure the accuracy and word error rate of translation.
The value range of BLEU is from 0 to 1. The closer the value is to 1, the
better the effect of machine translation and the lower the error rate of words
are. English-French data sets is used in the experiment to store the predicted
translated data and the reference translation data in different documents, and
analyzed the translation quality by calculating the values of Bleu1, Bleu2, Bleu3
and Bleu4. 300 statements are taken for calculation, and the result is shown in
table 1.
The GRAM in the table represents the number of words, and when the
GRAM is 4, the BLEU value is relatively low, but the accuracy is high when
the number of words is large, so BLEU4 is selected as the final evaluation value.
From the BLEU4 value, it can be seen that the evaluation value is relatively
low, indicating that the translation effect is not ideal. Through the analysis of
the experimental results, the main problems are as follows:

1. The model structure is relatively simple and the number of hidden layers in
RNN is not sufficient, and the training is imperfect.
2. The number of data sets is insufficient with single direction is single, which
cannot cause the model study well, resulting in poor translation effect.

4.3 Experiment Analysis of LangGraph-Based Agent

This experiment involves the implementation of a machine translation task using

an agent constructed on the LangGraph framework. As shown in Figure 7, this
paper has designed a flowchart that guides the translation process by analyzing
the language category and determining the intent, supporting the translation of
three languages: English (EN), French (FR), and Japanese (JP).
– Start Node (start): The beginning of the process.
– Analyze Language Category: Identify the language category of the input
text.
– Determine Intent: Based on the input text, ascertain the intent of the
text to select the appropriate translation process.
– Translation Node: Depending on the determined intent, the process calls
upon translateToEN, translateToFrench, or translateToJP to perform
the translation.
– End Node (end): The conclusion of the process, where the translation
result is outputted.

Fig. 7. Translation Agent Data Flowchart.

 As shown in Figure 8, this study implemented a machine translation agent
using the LangGraph framework and conducted translation experiments from
English to French and from French to English. Figure (a) illustrates the transla-
tion experiment from English to French, while Figure (b) shows the translation
experiment from French to English. The experimental results indicate that for
translation tasks between English and French, the agent based on the LangGraph
framework produced translation results that align with expectations, demon-
strating the agent’s ability to accurately comprehend the source language text
and generate equivalent text in the target language. The experiments not only
verified the agent’s translation capabilities but also highlighted its potential for
practical applications.

Fig. 8. English-French Translation Example.

4.4 Discussion

Although this study has achieved positive results, there are still several areas
that require further exploration and improvement:
1. Expanded Context Handling: Demonstrate LangGraph’s ability to retain
long-term context across paragraphs or entire documents rather than isolated
sentences.
2. Dynamic Agent Adaptation: Showcase how agents can dynamically adapt
to different domains or contexts using LangGraph, e.g., switching from casual
conversational translation to technical or formal translation styles.
3. Interactive Features: Implement interactive feedback loops where users can
provide corrections, and agents learn and adapt over time.
4. Human in the loop:Delve into the application of Human-in-the-loop within
the LangGraph framework, setting breakpoints during the graph execution pro-
cess to permit human intervention and adjustment. Engage in state editing and
human feedback, modify the state of the graph during execution and adjust
the workflow based on human input. Utilize the time travel feature, review the
execution state of the graph by retracing its steps to facilitate debugging and
optimization. This sets a course for future research in this domain.
5 Conclusion

This study underscores the significant potential of Agent AI and LangGraph

in advancing machine translation technology. By leveraging the modularity and
task-specific capabilities of agents alongside the robust workflow management
provided by LangGraph, the proposed system achieves a high degree of flexibil-
ity, scalability, and accuracy in multilingual translation tasks. The integration
of large language models (LLMs) like GPT-4o further enhances the semantic
understanding and contextual relevance of translations.
LangGraph’s graph-based framework plays a pivotal role in orchestrating
these agents, enabling seamless state management, dynamic task allocation, and
efficient multi-agent collaboration. This modular and automated design sim-
plifies the addition of new languages and workflows, ensuring adaptability to
evolving translation needs. Additionally, the system demonstrates the feasibility
of maintaining contextual coherence and accuracy, essential for practical appli-
cations in diverse domains.
While the study achieved promising results, limitations in data availability
and model complexity leave room for future exploration. Expanding the dataset,
incorporating additional languages, and introducing features like long-term con-
text retention and human-in-the-loop feedback will further refine the system’s
capabilities.
The application of Agent AI with LangGraph highlights a new paradigm for
machine translation, combining the strengths of advanced LLMs with modular,
scalable frameworks. This approach not only drives improvements in translation
quality but also paves the way for broader adoption of intelligent, adaptable
language-processing solutions in diverse fields such as e-commerce, education,
and international communication.

References

1. BANSAL M, MIWA M. End-to-end relation extraction using LSTMs on sequences

and tree structures[C]. New York: Association for Computational Linguistics, 2016:
1105-1116.
2. ZHANG J, ZONG C Q. Application of Neural Network Language Model in Sta-
tistical Machine Translation [J]. Technology Intelligence Engineering, 2017, 3(3):
21-28.
3. LIU Y. The forefront of Progress in Neural Machine Translation [J]. Journal of
Computer Research and Development, 2017, 54(6): 1144-1149.
4. SUN Z Y, LU C X, SHI Z Z, et al. Deep Learning Research and Progress [J].
Computer Science, 2016, 43(2): 1-8.
5. BLUNSOM P, KALCHBRENNER N. Recurrent continuous translation mod-
els[C]//Proceedings of the ACL conference on empirical methods in natural lan-
guage processing. [s.l.]: [s.n.], 2013: 1700-1709.
6. LE Q, STUSKEVER I, VINYALS O. Sequence to sequence learning with neural
networks[C]//Advances in neural information processing systems. [s.l.]: [s.n.], 2014:
1-9.
 7. CHO K, VAN MERRIEHBOER B, BAHDANAU D, et al. On the properties of
neural machine translation: encoder-decoder approaches[C]//English workshop on
syntax, semantics and structure in statistical translation. Doha: [s.n.], 2014: 154-
163.
8. KOEHN P, OCH F J, MARCU D. Statistical phrase based transla-
tion[C]//Proceedings of the 2003 conference of the North American chapter of
the association for computational linguistics on human language technology. Ed-
monton, Canada: Association for Computational Linguistics, 2003: 48-54.
9. BAHDANAU D, CHO K, BENGIO Y. Neural machine translation by jointly learn-
ing to align and translate[C]//Proceedings of ICLR. [s.l.]: [s.n.], 2015.
10. WANG L, YANG J, CHEN L, et al. Parallel Optimization Algorithm of Chinese
Language Model Based on Cyclic Neural Network [J]. Journal of Applied Sciences,
2015, 33(3): 253-261.
11. SOCHER R, LIN C, NG A, et al. Parsing natural scenes and natural language with
recursive neural networks[C]//Proceedings of the 28th international conference on
machine learning. Washington: [s.n.], 2011: 129-136.
12. IRSOY O, CARDIE C. Deep recursive neural networks for compositionality in lan-
guage[C]//Proceedings of the 27th international conference on neural information
processing systems. Montreal, Canada: MIT Press, 2014: 2096-2104.
13. HOCHREITER S, BENGIO Y, FRASCONI P. Gradient flow in recurrent nets:
the difficulty of learning long term dependencies[M]. [s.l.]: Wiley-IEEE Press, 2001:
237-243.
14. HUANG P. Character recognition in natural scene based on in-depth learning [D].
Hangzhou: Zhejiang University, 2016.
15. PAPINENI K, ROUKOS S, WARD T, et al. BLEU: a method for automatic evalu-
ation of machine translation[C]//Proceedings of the 40th annual meeting on asso-
ciation for computational linguistics. Philadelphia: Association for Computational
Linguistics, 2002: 311-318.
16. DODDINGTON G. Automatic evaluation of machine translation quality using
n-gram co-occurrence statistics[C]//Proceedings of the second international con-
ference on human language technology research. San Diego, California: Morgan
Kaufmann Publishers Inc., 2002: 138-145.
17. BANERJEE S, LAVIE A. METEOR: An Automatic Metric for MT evaluation
with improved correlation with human judgments[C]//Proceedings of the ACL
workshop on intrinsic and extrinsic evaluation measures for machine translation
and/or summarization. Ann Arbor: [s.n.], 2005: 65-72.
18. SNOVER M, DORR B, SCHWARTZ R, et al. A study of translation edit rate with
targeted human annotation[C]//Proceedings of association for machine translation
in the Americas. [s.l.]: [s.n.], 2006: 223-231.