In February of this year, IBM’s new machine called Watson had beaten human Champions in the game of Jeopardy. Considering the fact that the questions in Jeopardy are asked in natural human language, beating human champions for a computer requires innovative techniques.
Lets try to understand how Watson works:
First of all, we should keep in mind that, the questions in Jeopardy are asked by giving clues. The real language is full of ambiguities and complexities. This makes it a very special game for a computer, because the data or keywords are not readily available for searching. As you know, computers don’t understand verbs, nouns or people, but only ones or zeros.
There are mainly four steps as far as how Watson comes up with its answers:
Step 1: Watson first parses the question into parts of speech. This way it can come up with different options for what is being asked and the type of question.
Step 2: Hypothesis Making: In this step, Watson makes hypotheses, literally thousands of them, for each possible option of what is being asked. At this stage, Watson analyses hundreds of millions of documents, to come up with as many answers as possible. The quantity of answers is more important than the quality here, as it needs to make sure that the correct answer is not being missed.
Step 3: Hypothesis Scoring: Here Watson begins to score the possible answers form step 1, first by eliminating the obvious wrong answers. After that, Watson continues its search through the documents to collect negative or positive evidence for the remaining options. Here is one of the most critical steps: Watson understands whether it is a negative or positive evidence because, it knows or rather, “understands” the relationships between words! It can then make decisions on evidences according to this knowledge. Thousands of algorithms work in parallel at this stage, to grade different options.
Step 4: Final Ranking: Watson also uses its previous experiences based on “similar” questions, in order to finalize the grade of answers, and usefulness of evidence. In the end, Watson comes up with possible ratings with different confidence levels percentages, for all possible answers.
See video here on IBM website.