Amazon Researchers Train Multilingual AI Model to Enhance Product Search

2020-01-10

Amazon operates in 14 countries worldwide, with 9 of them eligible for its Prime annual subscription service. With this in mind, the company has a grand ambition to ensure that its shopping experience is not compromised by language barriers. In order to find an efficient method for translating multiple languages, researchers at Amazon have developed a shopping model called "Multitask". In this model, the functionalities of different tasks overlap and often reinforce each other. This artificial intelligence is trained on data from several different languages, allowing for better results regardless of the language used. As explained by Amazon applied scientist Nikhil Rao in a blog post, this improvement is based on the idea that one language can fill in the gaps of another. For example, phrases that may be confusing in French may be clearer in German, so multilingual training can help enhance the distinctions between various product queries. The team's system maps queries related to products and product descriptions to the same region in the representation space, regardless of language. This is primarily done to help the model generalize the knowledge it has learned in one language to other languages, ensuring that search results in the same region of the space are closely related to each other. The system takes two inputs, a query and a product title, and outputs a single bit indicating whether the product matches the query. An encoder component utilizes Google's Transformer architecture, which the researchers claim has better scalability compared to other architectures, and the model's classifier combines the query and product encodings. The team trains the system by randomly selecting an input language and "teaching" it to classify query-product pairs using that language. It then undergoes a series of epoch training using annotated sample queries for each input language. The alignment phase minimizes the distance between the product name encoding and the query, ensuring that the output, tailored for different languages, shares a representation space. Amazon states that they have achieved "strong results" in experiments involving 10 different bilingual patterns, 10 trilingual patterns, and 1 pentalingual pattern. The F1 score, a common performance metric in artificial intelligence that takes into account the ratio of false positives and false negatives, shows that a multilingual model trained in both French and German outperforms a model trained only in French by 11% and a model trained only in German by 5%.

RECENT AI NEWS

RECENT AI TOOLS