A Mixed Neural Network and Support Vector Machine Model for Tender

Creation in the European Union TED Database

Sangramsing Kayte

and Peter Schneider-Kamp

Department of Mathematics and Computer Science (IMADA), University of Southern Denmark (SDU),

Odense M, 5230, Denmark

Keywords:

Natural Language Processing, Natural Language Understanding, Named-entity Recognition, Logistic

Regression, European Union, Tender Electronic Daily, Common Procurement Vocabulary.

Abstract:

This research article proposes a new method of automatized text generation and subsequent classiﬁcation

of the European Union (EU) Tender Electronic Daily (TED) text documents into predeﬁned technological

categories of the dataset. The TED dataset provides information about the respective tenders includes features

like Name of project, Title, Description, Types of contract, Common procurement vocabulary (CPV) code,

and Additional CPV codes. The dataset is obtained from the SIMAP-Information system for the European

public procurement website, which is comprised of tenders described in XML ﬁles. The dataset was pre-

processed using tokenization, removal of stop words, removal of punctuation marks etc. We implemented a

neural machine learning model based on Long Short-Term Memory (LSTM) nodes for text generation and

subsequent code classiﬁcation. Text generation means that given a single line or just two or three words of the

title, the model generates the sequence of a whole sentence. After generating the title, the model predicts the

main applicable CPV code for that title. The LSTM model reaches an accuracy of 97% for the text generation

and 95% for code classiﬁcation using Support Vector Machine(SVM). This experiment is a ﬁrst step towards

developing a system that based on TED data is able to auto-generate and code classify tender documents,

easing the process of creating and disseminating tender information to TED and ultimately relevant vendors.

The development and automation of this system will future vision and understand current undergoing projects

and the deliveries by a SIMAP-Information system for European public procurement tenders organisation

based on the tenders published by it.

1 INTRODUCTION

Natural language processing (NLP) is a wide area

concerned with the application of computers to an-

alyze, understand, and derive meaning from human

language in a smart and useful way. NLP can orga-

nize and structure knowledge to perform tasks such as

automatic summarization, translation, named entity

recognition, relationship extraction, sentiment analy-

sis, speech recognition, and topic segmentation (Man-

ning et al., 2014). The unswerving increase of cate-

gories for tender documents has resulted in the need

for methods that can automatically classify new doc-

uments within subcategories of the text. In this pa-

per we attempt to make the categorization process

automated using a machine learning approach, i.e.,

to learn from the European public procurement ten-

https://orcid.org/0000-0002-9156-3794

https://orcid.org/0000-0003-4000-5570

der data in Tender Electronic Daily (TED) database.

Given text describing the tender, the model predicts

codes using the common procurement vocabulary

(CPV) that are relevant for this text. We commence

this paper with an overview of the European Union

(EU) TED and then move to the introduction of a few

terminologies which are relevant to this paper. TED

works under the Policies of European Public procure-

ment as drawn by the Directorate General for Internal

Market, Industry, Entrepreneurship, etc. A tender is

a document that a purchasing agent publishes to an-

nounce its request for certain goods or services. The

whole electronic tendering system is similar to that of

a traditional one, i.e. a selling agent submits a bid

to the purchasing organisation against the tender en-

quiry, which evaluates the quoting vendors based on

the technical speciﬁcations as well as ﬁnancial terms

and conditions.

Then the purchasing agent announces the suc-

cessful bidder. We carried out the classiﬁcation

Kayte, S. and Schneider-Kamp, P.

A Mixed Neural Network and Support Vector Machine Model for Tender Creation in the European Union TED Database.

DOI: 10.5220/0008362701390145

In Proceedings of the 11th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2019), pages 139-145

ISBN: 978-989-758-382-7

139

on the SIMAP-Information system for European

public procurement tenders available on its web-

site (https://ted.europa.eu/TED/main/HomePage.do).

eNotices is the online tool for preparing public pro-

curement notices and publishing them in the Supple-

ment to the Ofﬁcial Journal of the EU.

There is a standard format for the calling of

tenders in public procurement in European nations,

which are referred online with the process, the Ten-

der European Daily (TED) database. In this, there are

different sections allotted to various categories like

construction, ﬁnancing, operation, and distribution of

water bottling facilities, etc. The syntactically-parsed

information is maintained in these TED tenders (Gel-

derman et al., 2006). The TED tenders are called

from various sectors associated with public develop-

ment and information is stored in the form of text and

numbers. The important features of the data are given

in Table 1.

The primary goal of this research is to classify ten-

ders according to (parts of) the title of the project. Ti-

tles are generated from the user-given text. The im-

plemented system achieves an accuracy of 97% when

generating titles from user-given text and accuracy

of 95% when predicting CPV codes from the gen-

erated titles. After a deep inspection of the dataset,

it becomes apparent that the CPV code describes the

uniqueness of each tender. Here, we tried to extract

information about certain tenders in TED based on

year-wise allotment along with their CPV codes. So

when the CPV code is given it retrieves the informa-

tion about the project.

In Section 2 of this paper, work related to the pro-

posed method is described. Section 3 reviews relevant

natural processing techniques. Section 4 describes the

corpus of tenders from TED used in this project. The

proposed framework is detailed in Section 5. Sec-

tion 6 covers a preliminary experimental evaluation

of the framework. We conclude in Section 7.

2 RELATED WORK

The neurons interconnected with each other simulate

the network structure of neurons in the human brain

(Le et al., 2011). The nodes are interconnected by

edges, and then nodes are organized into layers. Com-

putation and processing are done in a feed-forward

manner, and errors can be back-propagated to pre-

vious layers to adjust the weights of corresponding

edges (Hoskins et al., 1990). The hidden nodes are

randomly assigned to avoid the extra load pay-off in

the structure. In a simple network, the weights are

usually learned in one single step which results in

faster performance (Towell and Shavlik, 1993). For

more complex relations, deep learning methods with

multiple hidden layers are adopted. These methods

were made more feasible with the advances of com-

puting power in hardware and the Graphics Process-

ing Unit (GPU) (Cires¸an et al., 2012).

This allows them to deal with a time sequence

with temporal relations such as speech recognition

(Waibel et al., 1995). According to a previous com-

parative study of RNNs in NLP, RNNs are found to

be more effective in sentiment analysis (Wang et al.,

2016). Thus, we focus on RNNs with Long Short-

Term Memory (LSTM) in this paper. As the time

sequence grows in LSTM, it’s possible for weights

to grow beyond control or to vanish. To deal with

the vanishing gradient problem in training RNNs and

LSTM has been proposed to learn long-term depen-

dency among longer time period (Graves et al., 2005).

In addition to input and output gates, forget gates are

added in LSTM. They are often used for time se-

ries prediction and hand-writing recognition (Sunder-

meyer et al., 2012). The shortfall of LSTM is that

they are only able to make use of the previous con-

text. To avoid this, LSTM is designed for processing

the data in both directions with two separate hidden

layers, which are then feed forwards to the same out-

put layer. Using LSTM will run your inputs in two

ways, one from past to future and one from future to

past. This will help to improve the results and under-

stand the context in a better way (Wang et al., 2018).

For NLP, it is useful to analyze the distribu-

tional relations of word occurrences in documents

(Joachims, 2002). The simplest way is to use one-hot

encoding to represent the occurrence of each word in

documents as a binary vector (Turian et al., 2010). In

distributional semantics, word embedding models are

used to map from the one-hot vector space to a con-

tinuous vector space in a much lower dimension than

conventional bag-of-words (BoW) model (Cho et al.,

2014). Among various word embedding models, the

most popular ones are distributed representation of

words such as Word2Vec and GloVe, where neural

networks are used to train the occurrence relations be-

tween words and documents in the contexts of train-

ing data (Pennington et al., 2014). In this paper, the

Word2Vec word embedding model is used to repre-

sent words in short texts. Then, LSTM classiﬁers are

trained to capture the long-term dependency among

words in short texts. The sentiment of each text can

then be classiﬁed as positive or negative (Wang et al.,

2016). In this paper, LSTM is utilized in learning

sentiment classiﬁers of short texts, like title and CPV

code.

KMIS 2019 - 11th International Conference on Knowledge Management and Information Systems

140

Table 1: Project features of data.

Features Data type Values

Name Text singe line free text

Description Text multiple line free text

Types of contract Text one of WORKS, SUPPLIES, or SERVICES

CPV code Number from predeﬁned hierarchical catalog of codes

additional CPV codes List of numbers from predeﬁned hierarchical catalog of codes

3 NATURAL LANGUAGE

PROCESSING

NLP investigates the use of computers to process

or to understand human languages for the purpose

of performing useful tasks like information ﬁltering,

language identiﬁcation, readability assessment, and

sentiment analysis, etc. NLP is an interdisciplinary

ﬁeld that is a combination of the subﬁeld of com-

puter science, information engineering, and artiﬁcial

intelligence. From a scientiﬁc perspective, NLP aims

to model the cognitive mechanisms underlying the

understanding and production of human languages

(Deng and Liu, 2018). The following are the impor-

tant pipeline of NLP talks processing:

Text Segmentation.

In this method, the sentence is broken into a sequence

of contiguous parts called segments. There exist an-

other sequence segmentation like sentence segmenta-

tion breaking a piece of string into sentences which

is an important post-processing stage for speech tran-

scription and chunking also known as shallow pars-

ing to ﬁnd important phrases from sentences, such as

noun phrases (Pak and Teh, 2018).

Named Entity Recognition (NER).

NER is the process of ﬁnding and tagging named en-

tities existing in the given text into pre-deﬁned cat-

egories. The NER task is hugely dependent on the

knowledge base used to train the NER extraction al-

gorithm, so it may or may not work depending upon

the provided dataset it was trained on.

Word Embedding.

A word embedding learning method is used to cal-

culate these embeddings using distributional seman-

tics; that is, they approximate the meaning of a word

by considering how often it co-occurs with other

words in the corpus. These embedding-based met-

rics usually approximate sentence-level embeddings

using some heuristic to combine the vectors of the in-

dividual words in the sentence. The sentence-level

embedding between the generated and reference re-

sponse are compared using a measure such as a cosine

distance (Mikolov et al., 2013).

Word2vec.

Word2vec is a two-layer neural net used in the recur-

rent neural network that processes text. Its input is a

text corpus and its output is a set of vectors. It gives

the best results for any kind of sequence neural net-

work models to learn node and sequence representa-

tions based on node sequences (Liang et al., 2016).

Bag-of-Words (BOW).

The main idea of BOW is to predict the target word

with surrounding context words. For convenient,

the surrounding words are symmetric (so as in skip-

gram), i.e., a window with size m is predeﬁned and the

task is to predict the target word w

with a sequence of

words (wc − m, ..., wc − 1, wc + 1, ..., wc + m), where

denotes the word at position c (Le and Mikolov,

2014).

Text Cleaning.

The text cleaning is needed for converting the text into

a particular input format. This helps to remove the

tagged data, punctuation, stop words, abbreviations

from the given input text (Hardeniya et al., 2016).

Following are a few steps to clean the data:

• Removing the HTML tags

• Removing Punctuation and Stopwords

• Stemming

• Conversion of data text to lower case

4 CORPUS DESCRIPTION

The corpus consists of the information about the an-

nouncements and calls related to tenders in TED. The

data is stored in an open XML format. We extracted

A Mixed Neural Network and Support Vector Machine Model for Tender Creation in the European Union TED Database

141

it to a Comma-Separated Values (CSV) format, con-

taining the information in form of text and numbers

as indicated in Table 1. The CSV ﬁle contains infor-

mation like the title of the project and CPV code as

an important entity along with other sub-entities like

description, reference of the project. The other enti-

ties like types of contract and a detailed description

of a particular project are also part of the dataset. In

this way, we are having a dataset of approx. 40,000

texts for each year from 2015 to 2019. This data will

be helpful for training and implementing the machine

learning model for the required system.

5 PROPOSED FRAMEWORK

The proposed framework consists of NLP library for

pre-processing such using Natural Language Toolkit

(NLTK) library and unidirectional LSTM architecture

for sequence prediction and classiﬁcation. We have

also applied linear support vector machine (SVM)

classiﬁer to classify CPV main code category from

other relevant ﬁelds, which is shown in the architec-

ture in Figure 1.

The tender category of the title describe overall in-

formation about the particular product. Initially, ﬁrst

4-6 words from the title are given as input for the sys-

tem, the LSTM model generates the overall text in-

formation. At the ﬁrst stage, pre-processing and word

features are extracted. Secondly, the Word2Vec word

embedding model is used to learn word representa-

tions as vectors.

The external input gate unit g

(t)

is computed sim-

ilarly to the forget gate (with a sigmoid unit to ob-

tain a getting value between 0 and 1), but with its

own parameters of weights and model: (Hochreiter

and Schmidhuber, 1997)

(t) = σ

∑

i, j

(t)

∑

i, j

(t−1)

(1)

The output of the LSTM cell can also be shut off,

via the output gate q

(t)

, which also uses a sigmoid unit

for gating: (Hochreiter and Schmidhuber, 1997)

(t) = σ

∑

i, j

(t)

∑

i, j

(t−1)

(2)

which has parameters b

, U

, W

for its biases,

input weights and recurrent weights, respectively.

Among the variants, one can choose to use the cell

state q

(t)

as an extra input (with its weight) into the

three gates of the i − th unit, as shown in the above

equation. Thus the LSTM model takes the recurrent

input of weights and concurrently gives the output to

the cell state (Hochreiter and Schmidhuber, 1997).

5.1 Preprocessing

In this stage, we ﬁltered the text data into required

format needed for the training of the system. The pre-

processing is divided into four different steps:

1. Cleaning consists of getting rid of the less useful

parts of a text through stopword removal, dealing

with capitalization and characters, and other mi-

nor details.

2. Annotation This process involves the applica-

tion of a scheme to texts which include structural

markup and part-of-speech tagging.

3. Normalization consists of the translation (map-

ping) of terms in the scheme or linguistic re-

ductions through Stemming, Lemmatization, and

other forms of standardization.

4. Analysis consists of statistically probing, manip-

ulating, and generalizing from the dataset for fea-

ture analysis.

5.2 Tokenization

Tokenization is the process of splitting text ﬁelds into

meaningful segments by locating the word boundaries

like CPV code, title from the given text database.

The points where one word ends and another begins.

For computational linguistic purposes, the words thus

identiﬁed are frequently referred to as tokens. In writ-

ten languages where no word boundaries are explic-

itly marked in the writing system, tokenization is also

known as word segmentation, and this term is fre-

quently used synonymously with tokenization.

5.3 Word Embedding

Word embeddings is the representation of title, de-

scription of words in the form of vectors. The aim

is to represent words via vectors such that similar text

words or words used in a similar context are close to

each other while antonyms end up far apart in the vec-

tor space.

The conversion of word vectors into a high-

dimensional space can be envisioned as shown in Fig-

ure 2

5.4 Dense Layer

The dense layer contains a linear operation in which

every input is connected to every output by weight (so

there are n inputs * n outputs weights - which can be

a lot!). Generally, this linear operation is followed by

a non-linear activation function.

KMIS 2019 - 11th International Conference on Knowledge Management and Information Systems

142

Figure 1: The system architecture of the proposed approach.

Figure 2: Visualization of Word Embeddings.

5.5 Support Vector Machines

SVM is text classiﬁcation approach which can be ap-

plied in both types of classiﬁcation and regression

problems. The feature of SVM model ﬁnd a max-

imum marginal hyperplane(MMH) that best divides

the dataset into classes and minimization of error.

The classiﬁer identiﬁes the unique data points using

a hyperplane with the largest amount of margin using

training data.SVM ﬁnds an optimal hyperplane which

helps in classifying new data points and so also re-

ferred as discriminative classiﬁer (Sassano, 2003).

The decision function g in SVM framework is de-

ﬁned as: (Sassano, 2003)

g(x) = sgn( f (x))) (3)

f (x) =

∑

i=1

αK(x

, x) + b (4)

where K is a kernel function and σi are weights

5.6 Long Short-Term Memory (LSTM)

RNNs have a powerful sequence study capability,

which can ﬁnely describe the dependence relationship

of title and CPV code data (Zhang and Lu, 2018).

RNNs such as LSTM are powerful models that are

capable of learning effective feature representations

of sequences when given enough training data (Zhang

et al., 2016). RNNs are layered neural networks that

include recurring connections between layers. The re-

currence creates a temporal effect on the network, al-

lowing certain network connections (parameters) to

be used at different times. This allows RNNs to

capture temporal relationships from a data sequence,

which make them appropriate for predicting sequen-

tial data.

The LSTM deals especially in handling sequences

that have a large gap between relevant information for

learning or predictions of new features in the training

set. The attractive feature of LSTM is that it effec-

tively deals with the long-term dependence problem

in time sequences (Sutskever et al., 2014).

Figure 3: The LSTM cell. a-Input gate, c-Forget gate, b-

Output gate, d- output from the cell (Skovajsov

a, 2017).

Figure 3 shows the structure of the LSTM net-

work, where the repeated modules represent the hid-

den layer in each iteration. Each hidden layer contains

a number of neurons. Each neuron performs a linear

matrix calculation on the input vector and then out-

puts the corresponding results after the nonlinear ac-

A Mixed Neural Network and Support Vector Machine Model for Tender Creation in the European Union TED Database

143

tion of the activation function (Arjovsky et al., 2016).

In each iteration, the output of the previous iter-

ation interacts with the next word vector of the text

determines the preservation or abandonment of the in-

formation and the update of the current state. h

is the

input of the hidden layer in this iteration. According

to the current state information, the predicted output

value of hidden layer y is obtained, and the output

vector h

is provided for the next hidden layer at the

same time. Whenever there is a new word vector in-

put in the network, the output of the hidden layer at

the next moment is calculated in conjunction with the

output of the hidden layer at the last moment. The

hidden layer circulates and keeps the latest state (Po-

liak et al., 2017)

6 EXPERIMENTAL EVALUATION

We have calculated the accuracy of our NLP model

using text prediction and classiﬁcation. This involves

the concepts of true positives, true negatives, false

positives, and false negatives are also evaluated. False

positives are cases the model incorrectly labels as pos-

itive that are actually negative, or, in our example, the

text which is not identiﬁed is considered a false neg-

ative. True positives are data points classiﬁed as pos-

itive by the model that actually are positive (meaning

they are correct), and false negatives are data points

the model identiﬁes as negative that actually are pos-

itive (incorrect). The LSTM-based model we imple-

mented achieves an accuracy of 97% for the text gen-

eration and 95% for the code classiﬁcation.

Figure 4: ROC curve against the training accuracy and loss.

An Receiver operating characteristic (ROC) curve

plots the accuracy on the x-axis against the number

of epochs on y-axis. The true positive rate (TPR)

is the recall and the false positive rate (FPR) is the

probability of a false alarm. Figure 4 shows the ROC

curve of training accuracy and training loss of the sys-

tem against the number of epochs. We can read from

observations that the training accuracy gradually in-

creases and loss decreases as the number of epochs for

training of the system progresses. This generates and

classiﬁes the data accurately from the given dataset of

the model.

Figure 5: Confusion Matrix training accuracy.

Figure 5 confusion matrix is a summary of predic-

tion results on a classiﬁcation of text data from given

dataset. In this matrix, the number of correct and in-

correct predictions like title of project and CPV code

are summarized with count values and broken down

by each class. The confusion matrix helps to identify

the ways in which your classiﬁcation model is con-

fused when it makes predictions. It also clears the

errors being made by your classiﬁer but more impor-

tantly the types of errors that are being made. So in

this project, the confusion matrix is applied for the

classiﬁcation of text and the percentage accuracy is

95% where the title and CPV code is exactly identi-

ﬁed from the given text data.

7 CONCLUSIONS

This experimental paper reported a high-accuracy ap-

proach based on an LSTM model for predicting CPV

codes for TED tenders using a two step approach of

generating texts and subsequently classifying the gen-

erated texts. Future work is to further increase the

performance of the system as well as to automate the

remaining steps of TED tender creation to the fullest

extent possible.

REFERENCES

Arjovsky, M., Shah, A., and Bengio, Y. (2016). Unitary

evolution recurrent neural networks. In International

Conference on Machine Learning, pages 1120–1128.

KMIS 2019 - 11th International Conference on Knowledge Management and Information Systems

144

Cho, K., Van Merri

enboer, B., Gulcehre, C., Bahdanau, D.,

Bougares, F., Schwenk, H., and Bengio, Y. (2014).

Learning phrase representations using rnn encoder-

decoder for statistical machine translation. arXiv

preprint arXiv:1406.1078.

Cires¸an, D., Meier, U., and Schmidhuber, J. (2012). Multi-

column deep neural networks for image classiﬁcation.

arXiv preprint arXiv:1202.2745.

Deng, L. and Liu, Y. (2018). Deep Learning in Natural

Language Processing. Springer.

Gelderman, C. J., Ghijsen, P. W. T., and Brugman,

M. J. (2006). Public procurement and EU tendering

directives–explaining non-compliance. International

Journal of Public Sector Management, 19(7):702–

714.

Graves, A., Fernandez, S., and Schmidhuber, J. (2005).

Bidirectional LSTM networks for improved phoneme

classiﬁcation and recognition. In International Con-

ference on Artiﬁcial Neural Networks, pages 799–804.

Springer.

Hardeniya, N., Perkins, J., Chopra, D., Joshi, N., and

Mathur, I. (2016). Natural Language Processing:

Python and NLTK. Packt Publishing Ltd.

Hochreiter, S. and Schmidhuber, J. (1997). Long short-term

memory. Neural computation, 9(8):1735–1780.

Hoskins, J. C., Kaliyur, K. M., and Himmelblau, D. M.

(1990). Incipient fault detection and diagnosis us-

ing artiﬁcial neural networks. In IJCNN 1990, Inter-

national Joint Conference on Neural Networks, San

Diego, CA, USA, June 17-21, 1990, pages 81–86.

Joachims, T. (2002). Learning to classify text using sup-

port vector machines, volume 668. Springer Science

& Business Media.

Le, Q. and Mikolov, T. (2014). Distributed representations

of sentences and documents. In International confer-

ence on machine learning, pages 1188–1196.

Le, Q. V., Ranzato, M., Monga, R., Devin, M., Chen, K.,

Corrado, G. S., Dean, J., and Ng, A. Y. (2011). Build-

ing high-level features using large scale unsupervised

learning. arXiv preprint arXiv:1112.6209.

Liang, D., Altosaar, J., Charlin, L., and Blei, D. M. (2016).

Factorization meets the item embedding: Regulariz-

ing matrix factorization with item co-occurrence. In

Proceedings of the 10th ACM conference on recom-

mender systems, pages 59–66. ACM.

Manning, C., Surdeanu, M., Bauer, J., Finkel, J., Bethard,

S., and McClosky, D. (2014). The stanford corenlp

natural language processing toolkit. In Proceedings

of 52nd annual meeting of the association for com-

putational linguistics: system demonstrations, pages

55–60.

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

Dean, J. (2013). Distributed representations of words

and phrases and their compositionality. In Advances in

neural information processing systems, pages 3111–

3119.

Pak, I. and Teh, P. L. (2018). Text segmentation techniques:

a critical review. In Innovative Computing, Optimiza-

tion and Its Applications, pages 167–181. Springer.

Pennington, J., Socher, R., and Manning, C. (2014). Glove:

Global vectors for word representation. In Proceed-

ings of the 2014 conference on empirical methods in

natural language processing (EMNLP), pages 1532–

1543.

Poliak, A., Rastogi, P., Martin, M. P., and Van Durme, B.

(2017). Efﬁcient, compositional, order-sensitive n-

gram embeddings. In Proceedings of the 15th Confer-

ence of the European Chapter of the Association for

Computational Linguistics: Volume 2, Short Papers,

pages 503–508.

Sassano, M. (2003). Virtual examples for text classiﬁca-

tion with support vector machines. In Proceedings of

the 2003 conference on Empirical methods in natural

language processing, pages 208–215. Association for

Computational Linguistics.

Skovajsov

a, L. (2017). Long short-term memory descrip-

tion and its application in text processing. In 2017

Communication and Information Technologies (KIT),

pages 1–4. IEEE.

Sundermeyer, M., Schluter, R., and Ney, H. (2012). LSTM

neural networks for language modeling. In Thirteenth

annual conference of the international speech commu-

nication association, pages 194–197.

Sutskever, I., Vinyals, O., and Le, Q. V. (2014). Se-

quence to sequence learning with neural networks. In

Advances in neural information processing systems,

pages 3104–3112.

Towell, G. G. and Shavlik, J. W. (1993). Extracting reﬁned

rules from knowledge-based neural networks. Ma-

chine learning, 13(1):71–101.

Turian, J., Ratinov, L., and Bengio, Y. (2010). Word rep-

resentations: a simple and general method for semi-

supervised learning. In Proceedings of the 48th an-

nual meeting of the association for computational lin-

guistics, pages 384–394. Association for Computa-

tional Linguistics.

Waibel, A., Hanazawa, T., Hinton, G., Shikano, K., and

Lang, K. J. (1995). Phoneme recognition using time-

delay neural networks. Backpropagation: Theory, Ar-

chitectures and Applications, pages 35–61.

Wang, J., Yu, L.-C., Lai, K. R., and Zhang, X. (2016). Di-

mensional sentiment analysis using a regional CNN-

LSTM model. In Proceedings of the 54th Annual

Meeting of the Association for Computational Lin-

guistics (Volume 2: Short Papers), volume 2, pages

225–230.

Wang, J.-H., Liu, T.-W., Luo, X., and Wang, L. (2018). An

LSTM approach to short text sentiment classiﬁcation

with word embeddings. In Proceedings of the 30th

Conference on Computational Linguistics and Speech

Processing (ROCLING 2018), pages 214–223.

Zhang, K., Xu, J., Min, M. R., Jiang, G., Pelechrinis, K.,

and Zhang, H. (2016). Automated it system failure

prediction: A deep learning approach. In 2016 IEEE

International Conference on Big Data (Big Data),

pages 1291–1300. IEEE.

Zhang, Y. and Lu, X. (2018). A speech recognition acoustic

model based on lstm-ctc. In 2018 IEEE 18th Inter-

national Conference on Communication Technology

(ICCT), pages 1052–1055. IEEE.

A Mixed Neural Network and Support Vector Machine Model for Tender Creation in the European Union TED Database

145