Social Media Analysis
Task: Huge amounts of documents are generated every day in Twitter, Facebook. blogs, online reviews etc. Customers and consumers provide their opinion on products, brands, or marketing campaigns. Evaluating these opinions automatically can provide valuable insights for product development, marketing and sales, and there are various companies that provide solutions for social media monitoring.
Solution: We developed core technologies for text understanding tasks such as Sentiment Analysis, Entity Extraction or Topic Detection, using a machine learning algorithm that was trained on more than 10’000 documents. Our software can be integrated easily in existing Social Media Monitoring Tools, allowing to process thousands of documents in parallel.
Task: Given the image of a page in a newspaper, identify and cut out all articles in this page. This is used in print media monitoring to increase their production performance for media clippings.
Solution: Use a combined approach based on image processing and text analysis that incorporates information such as font size, spacings between text blocks, topic of a text block, pixel density etc. All these attributes are fed into a meta-classifier that assigns the title, subtitles, text blocks and images to each single article.
Task: Researchers in sociology, education or humanities often conduct structured interviews to assess their data. These interviews need to be transformed into written text, which can be evaluated and analyzed more easily. Typically, this involves two steps: transcribing the spoken words into text (transciption), and deciding who uttered the text if there was more than one speaker (speaker detection). There exist several good tools for the first task, but speaker detection is still an open problem that we tackle in a research project.
Solution: For a target speaker, we are given an audio sample in advance, which you use for training the algorithms.This allows us to distinguish between the target speaker and all other speakers in an interview or discussion. To accomplish this, we use a new approach with temporal characteristics of a speaker to improve existing algorithms for speaker detection.