Intent Detection in Search Queries with SEBERT Embeddings

In this paper, I will introduce a method of intent detection in search queries with the help of SBERT and cosine similarity… Search engines are important in a digital world today. Users desire prompt, precise and pertinent responses. The problem with traditional software is that it may not comprehend semantic meaning due to the use of the traditional systems that relied on a keyword system. The contemporary approaches to NLP demonstrate the potential to fill this gap with intent detection. Although this has been improved, the existing systems are still lacking a solution to the is- sues of short, vague queries, field-specific variations, and high computational cost of deep learning models. Bag-of-words and Standard TF-IDF approaches lack semantic relationships. We propose an SBERT-based intent detection algorithm, which utilizes even simple BERT to conduct large-scale similarity searches, which are not efficient. This starts with the preprocessing stage that involves tokenization, deletion of stop words and lemmatization. We then encode the queries, which are then dense vectors obtained through a siamese network of SBERT. The first step would be to enter the query in the SBERT embedding generator. Step 2 will be to store embeddings in fixed-length vector space. In Step 3, intent classification: to apply a supervised classifier such as Logistic Regression, SVM, or shallow neural networks to the embeddings or (b) apply unsupervised grouping algorithms such as K-Means or Agglomerative to form automatic groups of similar intent. Step 4 aims at calculating the semantic similarity between queries, and the cosine similarity is used to expand the intent categorization. Lastly,Step 5 assesses the results in reference to ground-truth labels in accuracy, precision, recall, and F1-score.Our method is much better than baselines. TF-IDF+SVM has an approximate 78 percent accuracy on benchmark datasets like ATIS and SNIPS. The accuracy of vanilla BERT fine-tuning is 86 and our SBERT-based system has an accuracy of 92 to 94 and F1-score is as high as 20. Moreover, SBERT removes the time of pair-wise comparison by almost 60 percent, and it is very scalable to execute queries in real time. Overall, contrary to current models, the SBERT framework is more accurate and efficient in intent detection in search engines and con- versational AI systems, and is a practical and scalable solution.