Log In Sign Up

Cross-Language Learning for Entity Matching

Transformer-based matching methods have significantly moved the state-of-the-art for less-structured matching tasks involving textual entity descriptions. In order to excel on these tasks, Transformer-based matching methods require a decent amount of training pairs. Providing enough training data can be challenging, especially if a matcher for non-English entity descriptions should be learned. This paper explores along the use case of matching product offers from different e-shops to which extent it is possible to improve the performance of Transformer-based entity matchers by complementing a small set of training pairs in the target language, German in our case, with a larger set of English-language training pairs. Our experiments using different Transformers show that extending the German set with English pairs is always beneficial. The impact of adding the English pairs is especially high in low-resource settings in which only a rather small number of non-English pairs is available. As it is often possible to automatically gather English training pairs from the Web by using annotations, our results could proof relevant for many product matching scenarios targeting low-resource languages.


page 1

page 2

page 3

page 4


Optimizing Transformer for Low-Resource Neural Machine Translation

Language pairs with limited amounts of parallel data, also known as low-...

Anchor-based Bilingual Word Embeddings for Low-Resource Languages

Bilingual word embeddings (BWEs) are useful for many cross-lingual appli...

WDC Products: A Multi-Dimensional Entity Matching Benchmark

The difficulty of an entity matching task depends on a combination of mu...

Harnessing Multilinguality in Unsupervised Machine Translation for Rare Languages

Unsupervised translation has reached impressive performance on resource-...

HumanAL: Calibrating Human Matching Beyond a Single Task

This work offers a novel view on the use of human input as labels, ackno...

Transformers without Tears: Improving the Normalization of Self-Attention

We evaluate three simple, normalization-centric changes to improve Trans...

Deep Indexed Active Learning for Matching Heterogeneous Entity Representations

Given two large lists of records, the task in entity resolution (ER) is ...

1 Introduction

Identifying offers for the same product is one of the central challenges for e-commerce applications such as price comparison portals and electronic market places. Training Transformer-based matchers using offers from different e-shops that share the same product identifier has proven to be a successful solution for product matching reaching F1 scores above 0.9 in many cases [4, 5, 3]. The bottleneck of this approach is that it requires a decent amount of pairs of offers for the products to be matched as training data. For widely-used languages such as English, the required training data can be extracted from large web crawls by relying on annotations which identify product titles, product descriptions, and product identifiers such as GTIN or MPN numbers within web pages111 [5]. For less widely used languages and less commonly sold products, it can be hard to find enough offers in the respective target language on the Web.

To deal with this problem, this paper explores the utility of English language offers for training product matchers for less widely spoken target languages such as German. For this, we experiment with training sets combining a large amount of English language offer pairs with smaller amounts of training pairs in the target language. We experiment with matchers that internally rely on BERT [2], a German version of BERT222, multilingual BERT333, XLM-R [1]

, as well as a SVM classifier. The multi-lingual Transformers were originally pre-trained on multi-lingual data and are fine-tuned without supervision for the alignment of cross-language vector spaces 

[6, 7].

Our experiments show that extending the German training set with English pairs is always beneficial. The impact of adding the English pairs is especially high in low-resource settings in which only a small set of German pairs is available.

The contributions of the paper are twofold: 1. Up to our knowledge, we are the first to experiment with cross-language learning for matching textual entity descriptions using Transformers. 2. We demonstrate that by combining English language training pairs and a rather small amount of training pairs in the target language, it is possible to learn product matchers reaching F1 scores above 0.9 and clearly outperforming matchers which were trained using only product offers in the target language.

Section 2 introduces the datasets that we use for the experiments. Section 3 gives an overview of the different matchers as well as the training process, while Section 4 presents and discusses the results of the experiments.

2 Datasets

We experiment with English- and German-language offers for mobile phones which have been crawled from 66 different e-shops, auction platforms, and electronic market places. Each offer contains a title, a description, and some product identifier such as a GTIN or MPN number. The data was collected using 150 mobile phones as seeds of the crawling process. These seeds contained widely-sold head products but also less-sold long-tail phones. In addition to offers for the seed phones, the dataset also contains offers for further phones which have been discovered during the crawling process. We group the offers into pairs by using shared GTIN, EAN, and MPN numbers as distant supervision [5]

. Afterwards, we remove the identifiers from the offers in order to prevent the matching from being trivial. Non-matching pairs were created by combining an offer for a seed product with an offer for a similar seed product or a similar offer from a phone which has been discovered during crawling. We arrange the pairs into language-specific training sets of different sizes. The training sets range from 450 to 7200 pairs and contain 50% matches and 50% non-matches. We use a German-language test set containing 1800 pairs (25% matches, 75% non-matches). Non of the pairs in the test set is also contained in the training sets. Half of the pairs in all sets were chosen randomly, while the other half contains hard corner-case pairs (measured by cosine similarity). This webpage

444 provides more information about the dataset creation process, offers for the same products in additional languages (Spanish and French), as well as the code for replicating the experiments.

3 Models and Baselines

We utilize three different pre-trained Transformer-based models from the HuggingFace555 library. As a monolingual English option, we deploy the BERT base model (’bert-base-uncased’). The model was pre-trained on the English Wikipedia and BookCorpus. We further use a German BERT model pre-trained on a diverse set of German data (’bert-base-german-dbmdz-uncased’), including the German Wikipedia, parts of the Common Crawl, and EU Bookshop corpus, which is about the same amount of pre-training data that was used for the English BERT. As multilingual models, we use multi-lingual BERT (’bert-base-multilingual-uncased’), which is trained on the top 100 largest Wikipedias, as well as XLM-RoBERTa (’xlm-roberta-base’), which is trained on a CommonCrawl corpus consisting of 100 different languages. This allows us to examine the performance increase afforded by models pre-trained in multiple languages. Furthermore, we calculate a simple word co-occurrence baseline using an SVM classifier for predictions.

We create the input sequences for the Transformer-based models by concatenating the title and description attributes of a product offer into one string and applying the respective tokenizer to both product offers in a pair to represent them in the standard input representation for Sequence Classification, i.e. [CLS] Product 1 [SEP] Product 2 [SEP] for BERT-based models. As input for the SVM baseline, we generate a bag-of-words word vector representation indicating co-occurring words in a product pair, which serve as input for the classifier.

For every experimental run, the learning rate was optimized in the range between 5e-6 and 1e-4 using a validation set and early stopping. If a given model did not improve for three consecutive epochs during hyper-parameter tuning, the run was stopped. During training, we fine-tuned the models for 25 epochs. We utilized a fixed batch size of 16 and a weight decay of 0.01. All other hyperparameters were set to their default values. The scores reported are averages over three runs that were individually trained using the same hyper-parameter setup.

4 Results and Discussion

F1 without EN 71.00 65.27 73.43 87.69 73.40
F1 with EN 71.05 74.29 89.83 91.44 86.98
Difference 0.05 9.02 16.40 3.75 13.58
Table 1: Evaluation results on the German test set w/ and w/o additional English training data. Training data sizes: EN - 7200, DE - 1800.

In the first set of experiments, we compare the performance of the different mono- and multi-lingual models on the German test set while training on the one hand with only 1800 German training pairs and on the other hand with the same 1800 German pairs and an additional 7200 English pairs. Table 1 shows the results of this experiment. When fine-tuned with only 1800 German pairs, the English BERT model scores the lowest overall, falling 6% F1 behind the SVM baseline at 65% F1, meaning that the language misfit between English pre-training and German fine-tuning has a severe negative impact. The German version of BERT can improve on the SVM by 2.5% F1, showing the importance of German language pre-training when compared to English BERT. The multilingual XLM-R achieves a comparable result to German BERT. Multilingual BERT achieves the best result with 87% F1, outperforming German BERT by 14% F1. The very strong performance of multilingual BERT in this case likely stems from the larger amount of pre-training data the model was trained on (Wikipedia dumps in 100 languages) in comparison to most other models. Even though only a part of its training data is in German, the large and varied pre-training data across different languages leads to a model that is highly training data efficient when being fine-tuned in German. After extending the German training set with the additional 7200 English training pairs (Row F1 with EN in Table 1), all models apart from the SVM improve significantly. German BERT sees the highest improvement with a gain of 16% F1 putting it 1.5% points behind multilingual BERT, which achieves the overall highest score of 91.5% F1. These results clearly show that adding training data in a high-resource language like English to a smaller amount of training data in the target language leads to improvements for all transformer models and thus is a promising course of action for practitioners.

DEEN 0 450 900 1800 3600 7200 0-7200
450 67.11 72.79 75.44 80.83 86.82 87.97 20.86
900 75.76 75.10 74.00 87.67 88.92 88.19 12.43
1800 87.69 88.43 88.38 90.17 90.72 91.44 3.75
3600 93.63 92.98 92.46 93.97 93.25 94.46 0.83
Table 2: Evaluation results on the German test set when varying the amount of training data for both languages.

In our second set of experiments, we train mBERT with different combinations of training data sizes for both languages in an effort to understand how much training data in the target language is mandatory and how much English training data needs to be added in order to reach a high performance level. Table 2 shows the results of these experiments. For the German training sets of size 900 and less, any combination of training data sets resulting in an overall amount of less than 2000 pairs lead to an F1 around 75% F1 or less regardless of the composition among the languages. Only when the training data consists of more than 2000 pairs, an F1 of more than 80% is achievable in all scenarios. If training data in the target language is scarce (450 pairs) adding English training pairs has a significant effect, leading to a large improvement in every step up until 3600 additional English pairs (86.82% F1). After this point, doubling the amount of English training data only yields an improvement of 1% F1. The beneficial effect of additional English training data is visible across all settings though it diminishes with increasing size of the German training set. Once the German training set reaches a size of 1800 pairs, the effect of adding English training data is no longer as strong as before but still leads to an overall better model, reaching a maximum of 94.5% F1 when trained with 3600 German and 7200 English pairs.

As an initial conclusion from the experiments, we can state that for successful cross-language learning during fine-tuning, a Transformer model that has also been pre-trained on large amounts of text in different languages should be chosen as starting point. Given that training pairs in head-languages such as English can automatically be extracted from the Web by relying on annotations [5], we believe that cross-language learning can contribute in many low-resource language scenarios to reduce labeling costs.

Acknowledgement: The datasets used in this paper were assembled by a student team consisting of Andreas Küpfer, Benedikt Ebing, Daniel Schweimer, Fritz Niesel, and Jakob Gutmann.


  • [1] Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., et al.: Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116 (2019)
  • [2] Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1. pp. 4171–4186 (Jun 2019)
  • [3]

    Li, Y., Li, J., Suhara, Y., Doan, A., et al.: Deep entity matching with pre-trained language models. Proceedings of the VLDB Endowment

    14(1), 50–60 (2020)
  • [4] Peeters, R., Bizer, C.: Dual-objective fine-tuning of bert for entity matching. PVLDB 14(10), 1913–1921 (2021)
  • [5] Primpeli, A., Peeters, R., Bizer, C.: The WDC training dataset and gold standard for large-scale product matching. In: Companion Proc. of The 2019 World Wide Web Conference. pp. 381–386 (2019)
  • [6]

    Ruder, S., Vulić, I., Søgaard, A.: A Survey of Cross-lingual Word Embedding Models. Journal of Artificial Intelligence Research

    65, 569–631 (Aug 2019)
  • [7]

    Wang, Z., Lv, Q., Lan, X., Zhang, Y.: Cross-lingual Knowledge Graph Alignment via Graph Convolutional Networks. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. pp. 349–357. Association for Computational Linguistics, Brussels, Belgium (Oct 2018)