2.1 Economic analysis of public procurement

The size and economic importance of PP have attracted the interest of many researchers, particularly in the field of economics. For instance, Iimi (Reference Iimi2006) examines procurement auctions for Japanese official development assistance projects and concludes that a 1% increase in the number of offers decreases the price of contracts by 0.2%. Similarly, Onur, Ozcan & Taş (Reference Onur, Ozcan and Tas2012b) show that procurement prices in Turkish PP auctions are negatively related to the number of offers. Taş (Reference Taş2022) provides empirical evidence on the significance of PP regulation quality on EU PP outcomes. Fazekas and Blum (Reference Fazekas and Blum2021) state that governments can achieve 5–10% price savings by improving PP practices. Gorgun et al. (Reference Gorgun, Kutlu and Tas2022) examine the impact of descriptions on PP notices. They find that competition is considerably higher when descriptions contain more information. Furthermore, they report that EU governments could save up to € 80 billion if PP descriptions were to contain detailed information. In our work, we focus on developing prediction models that may be beneficial for authorities and economists working on PP, rather than analyzing the economic impact of PP calls.

2.2 Prediction tasks for public procurement calls

Several researchers have worked on various prediction tasks about PP such as cost increase with respect to low bids (Williams and Gong Reference Williams and Gong2014), uncertainty risks of projects in bidding processes (Lee and Yi Reference Lee and Yi2017; Jang et al. Reference Jang, Yi, Son, Kang and Lee2019), contracts likely to result in corruption investigations (Gallego, Rivero, and Martínez Reference Gallego, Rivero and Martínez2021), the number of bidders (Demiray, Ozbayoglu, and Taş Reference Demiray, Ozbayoglu and Tas2015; Rabuzin and Modrusan Reference Rabuzin and Modrusan2019), and award price (García Rodríguez et al. Reference García Rodríguez, Rodríguez Montequín, Ortega Fernández and Villanueva Balsera2019). We now focus on studies on prediction tasks similar to ours.

Demiray et al. (Reference Demiray, Ozbayoglu and Tas2015) predict the number of offers in Turkish PP auctions using support vector machine (SVM) with nine features such as estimated cost, whether it involves goods or services, and whether it is open to foreign firms. Rabuzin and Modrusan (Reference Rabuzin and Modrusan2019) focus on Croatian PP calls and predict whether a PP call will receive a single offer or more, considering that a single offer might be an indicator of corruption. They use Naive Bayes, logistic regression (LR), and SVM with term frequency-inverse document frequency (TF-IDF) vectors. In order to predict the number of offers in the TED dataset, Gorgun, Kutlu, and Taş (Reference Gorgun, Kutlu and Taş2020) investigate several features including named entities, description lengths, TF-IDF vectors, business sector, country, and language codes. They report that KNN with n-gram vectors yields the highest performance. In their work, they use only 59,397 PP notices and translate all non-English descriptions to English using automatic translation tools. In our work, we use a much larger dataset and utilize multilingual language models, instead of machine translation.

García Rodríguez et al. (Reference García Rodríguez, Rodríguez Montequín, Ortega Fernández and Villanueva Balsera2019) predict the award price of Spanish PP calls using random forest (RF) regressor with 14 features such as date, duration, business sector, province, and municipality. Lima et al. (Reference Lima, Silva, de Souza Mendes, de Carvalho, Araujo and de Barros Vidal2020) extract the texts of 15 million Brazilian PP calls and annotate 1907 of them as “risky” in terms of corruption and fraud. They use a bidirectional long short-term memory model to detect risky calls. In our work, we use a dataset covering multiple countries and languages and work on four different prediction tasks.

Imhof and Wallimann (Reference Imhof and Wallimann2021) propose a supervised method that uses the distribution of bids to detect collusive groups of firms in procurement auctions. Aldana et al. (Reference Aldana, Falcón-Cortés and Larralde2022) propose an ensemble of RF to detect corrupt contracts in Mexico’s PP. They use a wide range of features such as descriptors of the buyer/supplier relation, type of contract, and others.

Modrusan et al. (Reference Modrusan, Rabuzin and Mrsic2020) focus on determining which parts of the PP calls are important to predict suspicious procurement procedures. They determine 11 terms such as “minimum,” “maximum,” “requirement,” “technical and expert” based on their interviews with experts in the field of the PP process. Subsequently, they show that using paragraphs that contain one of the selected terms and the following 10 paragraphs increases performance in predicting suspicious procurement procedures.

In our study, we work on four prediction tasks including the number of offers, whether a contract will receive a single offer, whether a foreign firm will be awarded, and effectiveness of contracts. In all our prediction tasks, we use only descriptions of calls to investigate their impact on the outcome of the procurement processes.

2.3 Studies using TED dataset

In our work, we use TED dataset to develop and test our models. Because it is a large, publicly available multilingual dataset, several researchers have utilized it as a resource for various NLP tasks. For instance, Ahmia et al. (Reference Ahmia, Béchet and Marteau2018) prepare a corpus of aligned sentences extracted from the TED dataset, which may be useful for machine translation studies. Simperl et al. (Reference Simperl, Corcho, Grobelnik, Roman, Soylu, Ruíz, Gatti, Taggart, Klima, Uliana, Makgill and Lech2018) focus on constructing a knowledge graph from TED data. They indicate multilinguality as a challenge in linking and comparing documents for analysis. They propose mapping word representations into a common semantic vector space which enables the comparison of document similarities across languages. Similarly, Zhang et al. (Reference Zhang, Jasaitis, Freeman, Alfrjani and Funk2023) focus on constructing a structured procurement contract database using TED dataset to ease the tendering process in healthcare. The authors mention the incompleteness of documents and differences in XML format as challenges they encountered. Mehrbod and Grilo (Reference Mehrbod and Grilo2018) develop a named entity recognizer to empower semantic search in the TED dataset.

2.4 Multilingual document classification

As we propose classification and regression models for multilingual documents, now we discuss studies on multilingual document classification. Pappas and Popescu-Belis (Reference Pappas and Popescu-Belis2017) propose multilingual hierarchical attention networks for document classification and evaluate their approach on a dataset covering news articles written in eight different languages. Van der Heijden et al. (Reference Van der Heijden, Yannakoudakis, Mishra and Shutova2021) propose a meta-learning approach for few-shot multilingual and cross-lingual adaptation.

A number of studies assess the performance of MBERT in various conditions. For instance, Wu and Dredze (Reference Wu and Dredze2019) investigate the performance of MBERT in a cross-lingual zero-shot transfer setting for various NLP tasks including document classification, part-of-speech tagging, named entity recognition, natural language inference, and dependency parsing. They show that MBERT yields high performance in all tasks. Pires et al. (Reference Pires, Schlinger and Garrette2019) show that MBERT’s performance decreases for the languages with different sentence structures from English. Karthikeyan et al. (Reference Karthikeyan, Wang, Mayhew and Roth2019) study the cross-lingual ability of MBERT, analyzing its architecture, linguistic properties of languages, and the learning objectives. They find that the lexical overlap between languages plays a negligible role in the cross-lingual performance of MBERT.

Many researchers also investigate whether we should use multilingual models trained with large datasets covering various languages or monolingual models trained with relatively smaller datasets focusing on a single language. For instance, Catelli et al. (Reference Catelli, Bevilacqua, Mariniello, di Carlo, Magaldi, Fujita, De Pietro and Esposito2022) explore the effects of cross-lingual learning for Italian sentiment analysis for hotel reviews. They show that MBERT outperforms the Italian-specific BERT model when trained with a mixed dataset composed of English-Italian reviews while Italian BERT outperforms MBERT when trained with only Italian reviews.

Piscitelli et al. (Reference Piscitelli, Arnaudo and Rossi2021) develop a CNN architecture for tweet classification using multilingual MUSE model. They evaluate its performance for zero-shot learning in English, Spanish, and Italian languages and show that multilingual training increases the performance for the languages with limited training data whereas it decreases the performance for English.

Aluru et al. (Reference Aluru, Mathew, Saha and Mukherjee2020) compare the performances of several approaches for multilingual hate speech detection for different data sizes and monolingual and multilingual settings. For the monolingual setting, they report that LASER with LR model outperforms others when the training data are limited. On the other hand, BERT outperforms others when there are more training data. They also compare the performances of LASER and MBERT in a cross-lingual setting and report that LASER achieves the highest score in six languages, whereas MBERT outperforms LASER in three languages. However, when these models are further trained with data in the target language, MBERT performs LASER in eight languages.

To our knowledge, our work is the first study working on these four prediction tasks on a dataset covering 22 European languages. We compare the performance of MBERT and XLMR for our prediction tasks. In addition, we compare cross-lingual performance of our models and explore which language pairs are suitable for cross-lingual training. Furthermore, we investigate whether we should use multilingual or monolingual models for our prediction tasks.

3.1 Topics

PP notices are categorized hierarchically such that notices are first divided into divisions, then groups, then classes, then categories, and finally subcategories. The dataset we use has 46 divisions, 322 groups, 1304 classes, 3403 categories, and 7841 subcategories in total, covering a wide range of topics, such as the delivery of different food products for armed forces in Berlin, Germany (PP notice ID: 2012385044), and the purchase of construction materials for the University Hospital in Vilnius, Lithuania (PP notice ID: 201286594).

3.2 Data extraction

The TED dataset is publicly available in complex but structured XML files. However, there are still many challenges in extracting the data due to the multilingual nature of the data, changes in the format of XML files throughout the years, and human errors. For instance, we observed that there is no common XML tag name for descriptions. While the majority of the descriptions are provided in three description-specific XML tags (short_descr, short_description_contract, and short_contract_description), a portion of the contracts provides description under general XML tags which might also contain other types of information. Therefore, regarding the general XML tags, we used the ones which contain the word “description” or its translation in the respective document’s original language (e.g., “Beskrivelse” for Danish and “Opis” for Polish) in the XML tag description. Furthermore, in contracts with multiple original languages, we select English descriptions if available. Otherwise, we randomly choose one of the respective original languages. In total, we created a dataset that consists of 333.832 notices. However, the estimated price value was provided in only 106.122 notices.

Figure 1. The number of all notices for each language in our dataset.

Figure 2. Number of notices that provide the estimated price information for each language in our dataset.

3.3 Language

Contract documents are written in 24 different European languages: Bulgarian (BG), Czech (CS), Danish (DA), German (DE), Greek (EL), English (EN), Spanish (ES), Estonian (ET), Finnish (FI), French (FR), Irish (GA), Croatian (HR), Hungarian (HU), Italian (IT), Lithuanian (LT), Latvian (LV), Maltese (MT), Dutch (NL), Polish (PL), Portuguese (PT), Romanian (RO), Slovak (SK), Slovenian (SL), and Swedish (SV). Figures 1 and 2 show the distribution of contracts’ original language in our dataset for all notices we extracted and for notices in which the estimated price information is provided, respectively. The distribution of languages is unbalanced, such that notices in German, French, Polish, English, and Dutch (i.e., the five most frequent languages) constitute 70.58% of all notices. Considering only notices with estimated price information, the distribution of languages is again unbalanced, such that notices written in Polish, German, Romanian, Italian, and Spanish (i.e., the five most frequent languages) constitute 66% of the data. In our work, we do not use Maltese and Irish due to the excessively limited number of calls written in these languages. Overall, our dataset covers descriptions written in 22 different languages with various alphabets (e.g., Latin, Cyrillic, and Greek).

4.1 Predicting number of offers ( ${\textbf{P}}_{\textbf{NumOffers}}$ )

EU authorities require a competitive open market to effectively utilize public resources. The procurement price is likely to decrease if the awarding process is competitive. Therefore, the number of offers submitted for PP notices is an effective metric to assess the level of competition (Onur, Özcan, and Taş Reference Onur, Özcan and Taş2012a). In this task, we predict the number of offers for a given PP notice using its description.

4.2 Predicting whether a contract will receive a single offer ( ${\textbf{P}}_{\textbf{SingleOffer}}$ )

PP calls with a single offer are used as an indicator for potential corruption (Rabuzin and Modrusan Reference Rabuzin and Modrusan2019). Wachs, Fazekas, and Kertész (Reference Wachs, Fazekas and Kertész2021) investigate corruption risk in EU PP using single-bidder contracts since there is no competition. Charron et al. (Reference Charron, Dahlström, Fazekas and Lapuente2017) show that single-bidding is a significant red flag for corruption in PP. In contrast to the ${\textrm{P}}_{\textrm{NumOffers}}$ task, we focus on finding PP notices that received exactly one offer. In this task, we consider calls with a single offer as positive class and apply a binary classification of PP calls using their descriptions.

4.3 Predicting whether a foreign firm will be awarded ( ${\textbf{P}}_{\textbf{FFAward}}$ )

Firms in any EU member country can apply for PP calls of authorities in any member country. However, it is still challenging for a firm to be awarded by a foreign authority due to being less familiar with the respective country than local firms. Thus, if a foreign firm can win a contract, it indicates market transparency and highlights the competitiveness of the bidding process (Kutlina-Dimitrova and Lakatos Reference Kutlina-Dimitrova and Lakatos2016). In this task, we consider PP calls in which a foreign firm is awarded as positive class and predict whether a foreign firm wins a contract for a given PP notice using its description.

4.4 Predicting effectiveness of contracts ( ${\textbf{P}}_{\textbf{Effectiveness}}$ )

The EU authorities calculate the estimated prices for PP calls by conducting extensive market analyses. The calculation of the estimated price is an essential part of the EU PP process. On average, actual contract prices are 90–95% of estimated prices, according to studies that analyze PP markets in Japan (Ishii Reference Ishii2009), Italy (Coviello, Guglielmo, and Spagnolo Reference Coviello, Guglielmo and Spagnolo2018), Turkiye (Onur, Ozcan, and Taş Reference Onur, Ozcan and Tas2012b), and the EU (Fazekas and Tóth Reference Fazekas and Tóth2018). Accordingly, awarded prices higher than estimated prices suggest an ineffective bidding process (Taş et al. Reference Taş, Dawar, Holmes and Togan2019). These contracts have poor value-for-money performance. In this task, we predict whether a given PP notice will cause an ineffective contract (i.e., the awarded price is higher than the estimated price), using its description. We consider ineffective contracts as the positive class.

5.1 Fine-tuning multilingual language models

The dataset we use covers 22 different languages and requires multilingual language models for our prediction tasks. Therefore, we fine-tune a transformer model for each prediction task. We investigate MBERT (Devlin et al. Reference Devlin, Chang, Lee and Toutanova2019) and XLMR (Conneau et al. Reference Conneau, Khandelwal, Goyal, Chaudhary, Wenzek, Guzmán, Grave, Ott, Zettlemoyer and Stoyanov2020) separately to compare their performances. We only use extracted description parts of the PP notices for fine-tuning. We use truncation and padding for the overflowing and the underflowing cases, respectively. In the fine-tuning process, an additional linear layer is appended to the transformer model with random weights. The additional layer is task-specific; therefore, a classification layer is added for ${\textrm{P}}_{\textrm{SingleOffer}}$ , ${\textrm{P}}_{\textrm{FFAward}}$ , and ${\textrm{P}}_{\textrm{Effectiveness}}$ tasks, and a regression layer is added for ${\textrm{P}}_{\textrm{NumOffers}}$ . Subsequently, these models are trained with cross-entropy loss and mean square error loss for the classification and regression tasks, respectively.

5.2 Filtering the training data based on CPV codes

As mentioned in Section 3, the TED dataset covers a wide range of procurement services and products in different countries. Therefore, the values we predict are also likely to be affected by the sector of PP calls and the country of the procuring authorities. For instance, the number of offers for the purchase of construction materials might be significantly smaller than the number of offers for the delivery of food products. In addition, if a country has no local firm eligible to provide a procured item (e.g., a particular construction material), then a foreign company will be awarded. In order to consider the differences across sectors, we divide the training dataset into multiple subsets, where each subset has PP notices for a separate business division (i.e., the first two digits of CPV codes). Subsequently, we train a separate model for each business division and language using the respective data. Additionally, economic conditions play a significant role in PP outcomes. Gugler, Weichselbaumer, and Zulehner (Reference Gugler, Weichselbaumer and Zulehner2015) show that competition in PP increases during economic crises. The contract prices in the Austrian PP decreased by 3.3% during the financial crisis of 2009. We use sentence embeddings returned by pre-trained multilingual language models to represent the notice descriptions. We investigated various machine learning algorithms such as KNN, RF, MLP, and logistic/linear regression.

5.3 Fine-tuning using Triplet Networks

We apply TNs (Hoffer and Ailon Reference Hoffer and Ailon2015) for our prediction tasks. Our goal is to make the embeddings of PP notices that have the same label closer (i.e., lower cosine distance) compared to an arbitrary PP notice. In particular, in TNs, the input consists of three examples: anchor a, positive p, and negative n. The loss function is as follows:

(1) \begin{equation} \textrm{max}(\textrm{CosineDistance}({{v}}_{{a}},{{v}}_{{p}}) - \textrm{CosineDistance}({{v}}_{{a}},{v}_{{n}})+\epsilon,0) \end{equation}

where $\epsilon$ is the margin that ensures that $v_p$ is at least $\epsilon$ closer to $v_a$ than $v_n$ and $v_a$ , $v_p$ , and $v_n$ are the embeddings for a, p, and n, respectively.

We select triplets for each task using the training dataset as follows. For each anchor PP a, we select a positive PP p such that (i) a and p are written in the same language, (ii) they share the same CPV code, and (iii) their labels are equal. Selecting negatives is more difficult because if the selected negative PP is not related to the positive PP, the model can easily distinguish it from the positive PP, which would make this triplet fruitless for training. Therefore, instead of selecting negatives uniformly at random, we select three different negatives for each pair (a,p) based on their CPV codes. In particular, we randomly select (i) one negative sample among those with the same CPV code, (ii) another negative sample among those sharing the first five digits of the CPV, and (iii) another one with any CPV code. All selected negatives are written in the same language as a.

Because the label distributions of tasks and the sizes of the dataset differ from each other, we select a similar number of triplets for each task by adjusting the number of positive examples selected for each anchor notice a. We train a single model for both ${\textrm{P}}_{\textrm{NumOffers}}$ and ${\textrm{P}}_{\textrm{SingleOffer}}$ tasks based on the number of offers, because using the actual number of offers is more informative than using binary labels in this kind of training.

The number of triplets for the tasks ${\textrm{P}}_{\textrm{NumOffers}},{\textrm{P}}_{\textrm{SingleOffer}},{\textrm{P}}_{\textrm{FFAward}}$ , and ${\textrm{P}}_{\textrm{Effectiveness}}$ is 1,268,827, 1,268,827, 1,048,603, and 1,127,083, respectively. We first fine-tune MBERT model using the loss function described above. Subsequently, we use the fine-tuned model to obtain an embedding for each description. Then, these embeddings are fed into a KNN model that returns the median label value of the $k$ -closest neighbor for each task. We refer to this model as ${\textrm{TN}}_{\textrm{KNN-MBERT}}$ . As in our filtering method in Section 5.2, we also apply filtering methods to our KNN model. In particular, we filter based on CPV codes ( ${\textrm{TN}}_{\textrm{KNN-MBERT-CPV}}$ ) and language ( ${\textrm{TN}}_{\textrm{KNN-MBERT-LANG}}$ ). In the ${\textrm{TN}}_{\textrm{KNN-MBERT-CPV}}$ and ${\textrm{TN}}_{\textrm{KNN-MBERT-LANG}}$ methods, we ensure that all the k neighbors belong to the same CPV code and are written in the same language, respectively.

6.1 Experimental setup

Train and test datasets. The details of our dataset are provided in Table 2. We observe that a portion of PP documents does not contain information about the predicted outcome, yielding different dataset sizes for each task. For all four tasks, we use 80% of the datasets for training and the rest for testing. The label distributions for both the ${\textrm{P}}_{\textrm{FFAward}}$ and ${\textrm{P}}_{\textrm{Effectiveness}}$ tasks are highly imbalanced. Only 14.5% of notices are found to be ineffective, and in 3.3% of calls a foreign firm is awarded. The label distribution for the ${\textrm{P}}_{\textrm{SingleOffer}}$ tasks is slightly more balanced compared to the other classification tasks. 21.0% of the notices received a single offer. For the regression task, that is, ${\textrm{P}}_{\textrm{NumOffers}}$ , the standard deviation and the average number of offers are 3.079 and 3.918, respectively.

Table 2. Statistics of our dataset

Evaluation metrics. In order to evaluate the performances of models, we report mean absolute error (MAE) for ${\textrm{P}}_{\textrm{NumOffers}}$ task and $F_1$ score for ${\textrm{P}}_{\textrm{SingleOffer}}$ , ${\textrm{P}}_{\textrm{FFAward}}$ , and ${\textrm{P}}_{\textrm{Effectiveness}}$ tasks. When calculating the $F_1$ score, we consider contracts with a single offer, awarded foreign firms, and ineffective contracts (i.e., minority classes) as positive classes.

Baseline models. We compare our models against the five baselines listed below.

• Median of the Sector ( ${\textrm{Med}}_{\textrm{Sec}}$ ). As discussed before, the sector might have a significant impact on the results of PP calls. For instance, the number of eligible firms in a sector places an upper bound on the number of offers a PP call can receive. This method, ${\textrm{Med}}_{\textrm{Sec}}$ , returns the median value of all contracts in the respective sector. If that sector has no contract, we use the median value of all contracts.

• Median of the Sector in the Procuring Country ( ${\textrm{Med}}_{\textrm{SecCnt}}$ ). The economic status of the procuring country may also affect the outcome of procurement calls. Therefore, ${\textrm{Med}}_{\textrm{SecCnt}}$ returns the median value of the contracts in the respective sectors of the procuring country. If no contract in the train set is suitable, we use the median value of all contracts.

• KNN w/uni-grams ( ${\textrm{KNN}}_{\textrm{unigram}}$ ). Gorgun et al. (Reference Gorgun, Kutlu and Taş2020) report that KNN with n-grams is effective in predicting the number of offers on the TED dataset. Therefore, we extract 10K uni-grams based on their TF-IDF scores and use a KNN model.

• Linear/Logistic Regression w/uni-grams ( ${\textrm{LR}}{\textrm{unigram}}$ ). Similar to ${\textrm{KNN}}_{\textrm{unigram}}$ , we use linear regression for the ${\textrm{P}}_{\textrm{NumOffers}}$ task and LR for the others with 10K uni-grams.

• Linear/Logistic Regression with Macroeconomic Variables ( ${\textrm{LR}}_{\textrm{MEV}}$ ). To assess the determinants of the number of offers, Taş et al. (Reference Taş, Dawar, Holmes and Togan2019) use a generalized linear model with procurement-specific variables such as the type of contracting authority, country of authority, the first two digits of CPV codes (i.e., division of the sector), whether the contract was covered by the Government Procurement Agreement, and the type of procedure. We implement their model for our prediction tasks and use linear and LR models for our regression and classification tasks, respectively.

Implementation. We use NLTKFootnote ^f for tokenization and stemming. In stemming, we use SnowBallStemmer for the languages it supports. Otherwise, we use the stemmer by selecting the language attribute as “porter.” We use scikit libraryFootnote ^g for TF-IDF vectors. We set the minimum document frequency to 0.8. For the baseline methods, we use scikit library with default parameters.

Hyper-parameter tuning. In order to set the hyper-parameters, we further split the initial training set into training and validation sets. We use 25% of the initial training set as the validation set, that is, 60% and 20% of all data are used as the training and validation sets, respectively. Once parameters are tuned, we use the whole training data, that is, 80% of all data, to train models. The hyper-parameters we tune are as follows.

For fine-tuned models, MBERT and XLMR models, we test the learning rates: 1e-5, 2e-5, 3e-5 for each task and choose the best-performing parameters for the remaining experiments. Accordingly, we set the learning rate for fine-tuning models to 1e-5 for ${\textrm{P}}_{\textrm{NumOffers}}$ , ${\textrm{P}}_{\textrm{SingleOffer}}$ , and ${\textrm{P}}_{\textrm{FFAward}}$ tasks, and 2e-5 for the ${\textrm{P}}_{\textrm{Effectiveness}}$ task. We fine-tune MBERT and XLMR models for 3 epochs with a batch size of 32.

For the KNN models, we examine various values for k in $\{1,3,5,\ldots, 51\}$ for each task in the validation set. Accordingly, we set k to 39 for the ${\textrm{P}}_{\textrm{NumOffers}}$ task and 3 for the ${\textrm{P}}_{\textrm{FFAward}}$ , ${\textrm{P}}_{\textrm{SingleOffer}}$ , and ${\textrm{P}}_{\textrm{Effectiveness}}$ tasks. For the TN-based training procedure, we set the learning rate to 8e-6 and $\epsilon$ to 0.1 in all tasks and used a batch size of 16.

6.2 Experimental results

In this section, we first compare our proposed models and the baselines (RQ-1 and RQ-2). Next, we report the impact of multilingual training (RQ-3). Finally, we present our analysis for cross-lingual training (RQ-4).

6.2.1 Impact of filtering, machine learning algorithms, and text representation

In this experiment, we train various machine learning algorithms with different multilingual embeddings with and without filtering of the train data based on CPV codes. In particular, we use MBERT and XLMR to represent descriptions and train separate models using RF, MLP, KNN, and logistic/linear regression (LR). Table 3 shows the results.

Table 3. Experimental results for various machine learning algorithms using embedding of MBERT and XLMR models. The values in the second column indicate whether filtering based on CPV code is applied or not. The best result for each task is written in bold.

For the ${\textrm{P}}_{\textrm{NumOffers}}$ task, filtering has a slightly positive impact in most of the cases. In addition, for the ${\textrm{P}}_{\textrm{SingleOffer}}$ task, filtering improved the performance in all cases. However, in the ${\textrm{P}}_{\textrm{Effectiveness}}$ task, filtering has a negative impact in general, and the best-performing model is ${\textrm{MLP}}_{\textrm{MBERT}}$ without filtering. Regarding the results for ${\textrm{P}}_{\textrm{FFAward}}$ task, the impact of filtering is mixed without any clear positive or negative impact.

However, considering the impact of embeddings, we do not have a clear conclusion. In the ${\textrm{P}}_{\textrm{NumOffers}}$ and ${\textrm{P}}_{\textrm{SingleOffer}}$ tasks, XLMR yields a higher performance in most of the cases whereas MBERT is preferable for the ${\textrm{P}}_{\textrm{FFAward}}$ task. In the ${\textrm{P}}_{\textrm{Effectiveness}}$ task, the average performances of MBERT-based methods are higher than the average performances of XLMR-based methods. It is also noteworthy that LR with XLMR yields an F1 score of 0 in the ${\textrm{P}}_{\textrm{FFAward}}$ and ${\textrm{P}}_{\textrm{Effectiveness}}$ tasks. This may be owing to the highly limited number of positive cases in these tasks. From the perspective of the machine learning algorithms, we observe that MLP yields the best results in all four tasks.

Overall, MLP with BERT embeddings yields the best results in classification tasks (i.e., ${\textrm{P}}_{\textrm{SingleOffer}}$ , ${\textrm{P}}_{\textrm{FFAward}}$ , and ${\textrm{P}}_{\textrm{Effectiveness}}$ ), whereas MLP with XLMR embeddings outperforms other models in the ${\textrm{P}}_{\textrm{NumOffers}}$ task. In addition, the best-performing models for the three tasks use our filtering approach.

In our next experiment, we evaluate the impact of filtering in our TN-based approach. For all tasks, we compare the performances of ${\textrm{TN}}_{\textrm{KNN-MBERT}}$ (i.e., no filtering), ${\textrm{TN}}_{\textrm{KNN-MBERT-LANG}}$ (i.e., language based filtering), and ${\textrm{TN}}_{\textrm{KNN-MBERT-CPV}}$ (i.e., CPV based filtering). The results are shown in Table 4.

Table 4. Experimental results for the fine-tuning procedure based on triplet networks for each task. The best results are written in bold.

Table 5. Experimental results for baselines, fine-tuned MBERT and XLMR models, and the best-performing models based on filtering and TN for each task. The best results are written in bold. The standard deviations across multiple runs of XLMR and MBERT models are shown. We calculated p-values between the best-performing models (shown in bold) and all baselines for each task and found that all p-values are less than 0.01.

We observe that filtering based on CPV slightly improves the performance in all tasks except for the ${\textrm{P}}_{\textrm{Effectiveness}}$ task. In addition, filtering based on language has a positive impact on all tasks, yielding the best results for all tasks. This might be because almost all European countries have different official languages. Therefore, filtering based on language also reduces the negative impact of macroeconomic differences across countries in our models.

6.2.2 Comparison against baselines

Now, we compare our models against the baseline models. In particular, we report the performance of the baseline methods explained in Section 6.1, fine-tuned MBERT and XLMR models, and the best-performing models in our previous experiments (i.e., ${\textrm{TN}}_{\textrm{KNN-MBERT-LANG}}$ and MLP configurations yielding the highest score for each task).

We run our transformer-based methods three times (due to random seed value) and report the average score here. Table 5 presents the results.

${\textrm{Med}}_{\textrm{SecCnt}}$ seems to be a highly strong baseline method, outperforming fine-tuned XLMR model, our MLP-based methods, and all other baselines in ${\textrm{P}}_{\textrm{NumOffers}}$ task. This suggests that the number of offers highly depends on the sector and country, without a high variance across the notices.

However, the fine-tuned MBERT model and ${\textrm{TN}}_{\textrm{KNN-MBERT-LANG}}$ outperform all baseline models in the ${\textrm{P}}_{\textrm{NumOffers}}$ task. All our methods outperform ${\textrm{LR}}{\textrm{MEV}}$ , which utilizes macroeconomic variables, providing empirical evidence that the description of notices impacts the number of offers in PP.

In the ${\textrm{P}}_{\textrm{SingleOffer}}$ task, ${\textrm{Med}}_{\textrm{SecCnt}}$ achieves the best results compared with the other baselines. However, all of our models, except ${\textrm{MLP}}_{\textrm{MBERT}}$ , outperform ${\textrm{Med}}_{\textrm{SecCnt}}$ .

The fine-tuned MBERT model yields the highest classification accuracy. In the ${\textrm{P}}_{\textrm{FFAward}}$ task, all our methods outperform all baselines. ${\textrm{Med}}_{\textrm{SecCnt}}$ is again a strong baseline and outperforms all other baseline models. While the awarding of a foreign firm can be affected by many macroeconomic issues, it is noteworthy that our models are able to catch this macroeconomic situation based on the descriptions.

For ${\textrm{P}}_{\textrm{Effectiveness}}$ , interestingly, ${\textrm{KNN}}_{\textrm{unigram}}$ outperforms all other baselines and our fine-tuned MBERT and XLMR models. This suggests that notices with particular phrases have similar effectiveness results. MLP with MBERT yields the highest score, and ${\textrm{TN}}_{\textrm{KNN-MBERT-LANG}}$ outperforms all baselines.

Regarding MBERT versus XLMR, we observe that fine-tuned MBERT outperforms fine-tuned XLMR in all tasks. In comparison with our models, our findings are mixed as no method consistently outperforms others. However, ${\textrm{TN}}_{\textrm{KNN-MBERT-LANG}}$ consistently outperforms all baselines in all tasks, whereas the fine-tuned models and MLP-based models outperform the baselines in three tasks.

In order to better understand the results, we calculate the precision and recall scores for each classification task. The precision score of the best-performing model for ${\textrm{P}}_{\textrm{SingleOffer}}$ , ${\textrm{P}}_{\textrm{FFAward}}$ , and ${\textrm{P}}_{\textrm{Effectiveness}}$ is 0.606, 0.511, and 0.434, respectively. Similarly, the recall score of the best-performing model for ${\textrm{P}}_{\textrm{SingleOffer}}$ , ${\textrm{P}}_{\textrm{FFAward}}$ , and ${\textrm{P}}_{\textrm{Effectiveness}}$ is 0.340, 0.321, and 0.159, respectively. We note that precision scores are higher than recall scores in all cases.

While our proposed models outperform baselines, their performance might be considered low. There might be many reasons for these scores. Firstly, we focus on predicting the outcome of PP processes by only using notice descriptions in order to explore their economic impact. However, obviously there are other factors affecting PP outcomes. For instance, if a country does not have any firm working on nuclear energy, a procurement call for building a nuclear energy plant cannot be won by a local firm regardless of the PP description. Therefore, we believe that the performance of our models can be improved by incorporating macroeconomic features of the countries. However, utilizing this kind of features is out of scope of this study. Furthermore, the label distribution is highly imbalanced with very few examples for positive classes. For instance, only 3.5% of the contracts in the train set are won by foreign firms. This kind of problems might be potentially solved by applying data augmentation methods. We leave this issue as a future work.

Figure 3. Performance of the fine-tuned MBERT in ${\textrm{P}}_{\textrm{NumOffers}}$ task for each language. In the normalization of the MAE scores, we divide the MAE scores by the average number of offers of the respective language.

6.2.3 Prediction performance for each language

In order to further analyze multilingual aspect of our dataset, we calculate the performance of fine-tuned MBERT for each language and task. Figure 3 shows the results for the ${\textrm{P}}_{\textrm{NumOffers}}$ task. The lowest MAE scores of the fine-tuned MBERT are for Croatian, Estonian, Slovak, Danish, and Czech, whereas the highest MAE scores are for Spanish, German, Portuguese, Italian, and Greek. Generally, we observe that the MAE score is higher for the languages of economically developed countries than for those of economically developing countries. This might be because the number of offers can be generally high in economically developed countries due to having more firms suitable to participate in bidding processes. Therefore, we divide the MAE score for each language by the average number of offers for each respective language. The results are shown as a line graph in Figure 3. Normalized MAE scores for languages are usually similar and have a lower variance than MAE scores. We achieve the lowest normalized MAE scores for Danish, Dutch, and German, which are Germanic languages. Their linguistic similarity to English, which is also a Germanic language, might be one of the factors that resulted in a higher performance than others.

Figure 4 presents $F_1$ score of the fine-tuned MBERT model and ratio of procurement notices with a single offer in the train set for each language. Generally, we observe a correlation between single offer ratio and $F_1$ score. For instance, we achieve the highest $F_1$ scores for Polish, Hungarian, and Romanian, which are also languages with the highest single offer ratio in the train set. In addition, we have the lowest $F_1$ scores for English, Danish, and Dutch, which are the languages with lowest single offer ratio. Overall, our results suggest that the performance of the model is affected by the number of labels in the train set instead of the linguistic features of languages. Therefore, the performance of the models might be also improved using data augmentation methods.

Figure 4. Performance of the fine-tuned MBERT in ${\textrm{P}}_{\textrm{SingleOffer}}$ task and the ratio of procurement notices with a single offer in the train set for each language.

We present the $F_1$ score of the fine-tuned MBERT model and the ratio of procurement notices in which the award price exceeds the estimated price (i.e., ineffective procurement calls) in the train set for each language in Figure 5. We observe that while the fine-tuned MBERT model achieves a high prediction accuracy for particular languages (e.g., Croatian and Finnish), it cannot detect any ineffective PP call correctly for seven languages. Among these seven languages, most of them have very low ineffective procurement ratio (e.g., Spanish and Bulgarian). Furthermore, in our dataset, there is no cost-ineffective Estonian call. However, while the ineffective procurement ratio for Latvian PP calls is not highly limited, the fine-tuned MBERT model again achieves a $F_1$ score of zero. This might be because the total number of calls in Latvian is also highly limited (see Figure 2). Interestingly, the model performance is low for all Romance languages (i.e., Italian, Spanish, Portuguese, and French, and Romanian), while we do not see any clear pattern for other language families.

Figure 5. Performance of the fine-tuned MBERT in ${\textrm{P}}_{\textrm{Effectiveness}}$ task and the ratio of effective procurement notices in the train set for each language.

Figure 6. Performance of the fine-tuned MBERT in ${\textrm{P}}_{\textrm{FFAward}}$ task and the ratio of procurement notices in which a foreign company is awarded.

Figure 7. Performance of the fine-tuned MBERT model using various train data for all four tasks. “Or” stands for “Original”. The columns that correspond to the models trained using all data (All) show the average of the three runs.

Figure 8. Comparison of cross-lingual MAE scores for the ${\textrm{P}}_{\textrm{NumOffers}}$ task. Lighter colors indicate better MAE scores. The overall MAE score of each language is given in the right-most column (Test). Languages are sorted in ascending order based on their overall performance.

Finally, we present the results for the ${\textrm{P}}_{\textrm{FFAward}}$ task in Figure 6. Our observations are as follows. First, we do not observe any pattern regarding any language family. Second, the performance of the model seems mostly affected by the positive class ratio. Finally, the fine-tuned MBERT model interestingly cannot detect any contract in which a foreign firm is awarded for the Czech and Dutch languages.