Both versions of the patent tell us how a query dependent score and a query independent score might be calculated for an answer. In the first version of the patent, the statements only said that an answer score uses the query dependent score, and this newer version tells us that both the query dependent and query independent scores are combined to calculate an answer score (to decide which answer is the best choice for a featured snippet).
Before the patent discusses how query dependent and query bangladesh phone number data independent signals can be used to create a response score, the patent tells us this about the response score:
The answer passage scorer receives candidate answer passages from the answer passage generator and scores each passage by combining scoring signals that predict how likely the passage is to answer the question.
In some implementations, the answer passage scorer includes a query dependent scorer and a query independent scorer that respectively generate a query dependent score and a query independent score. In some implementations, the query dependent scorer generates the query dependent score based on an answer term match score and a query term match score.
Search query-dependent evaluation for the evaluation of featured snippet snippet candidates
Query-dependent ratings are based on the term properties of a text passage.
The answer-dependent questions in searches do not describe what a searcher is looking for, because the answer is unknown to the searcher at the time of the search.
The query-dependent scoring process begins by finding a set of likely answer terms and compares that set to an answer passage of the possible featured snippet candidate to generate a score for the answer term. The set of likely terms is likely to be taken from the first N ranked results returned for a query.
The process creates a list of terms from terms contained in the top-ranked subset of results for a query. The patent tells us that each result is parsed and each term is contained in a term vector (see the glossary post on vector space analysis ). Stop words can be omitted from the term vector.
For each term in the list, a term weight can be generated for the term. The term weight for each term can be based on many results in the highest-ranking subset of results in which the term appears, multiplied by an inverse document frequency (IDF) value for the term. The IDF value can be derived from a large corpus of documents and provided to the query-dependent evaluator. Or the IDF value can be derived from the first N documents in the returned results. (more on TF-IDF analyses from colleague Kai Spriestersbach ). The patent tells us that other suitable term weighting techniques can also be used.
The scoring process for each candidate answer passage term determines how many times the term appears in the passage of a potential featured snippet candidate. So, if the term "apogee" appears twice in one of the candidate's passages, the term score for "apogee" for that candidate's answer passage is 2; however, if the same term appears three times in another candidate's answer passage, the term score for "apogee" for the other candidate's answer passage is 3.
For each term in the candidate's answer passage, the scoring process multiplies its term weight by the number of times the term appears in the answer passage. So, suppose the term weight for "apogee" is 0.04. For the candidate's first answer passage, the score based on "apogee" is 0.08 (0.08.times.2); for the candidate's second answer passage, the score based on "apogee" is 0.12 (0.04.times.3).
Other features can also be used to determine an evaluation score. For example, the query-dependent evaluation process can determine an entity type for an answer to the search query. The entity type can be determined by identifying terms, such as people, places, or things, and selecting the terms with the highest rating. The entity can also be determined from the search query (e.g., for the search query "who is the fastest man," the entity type for an answer is "man"). For each answer candidate, the evaluation process then identifies entities that are described in the possible candidate. If the entities do not contain a match with the identified entity type, the evaluation score for the candidate answer passage is reduced. (More on the topic of entity types in the article Everything you should know about entity types, classes, and attributes ).
Suppose the following candidate answer passage is provided for evaluation in response to the question [who is the fastest man]: Olympic sprinters have often set world records for sprint events during the Olympic Games. The most popular sprint event is the 100-meter sprint.
The query-dependent score manager will identify several entities - Olympics, Sprinter, etc. - but none of them are of type "man." The term "sprinter" is gender-neutral. Accordingly, the answer term score is collapsed. The score may be a binary score, e.g. 1 for the presence of the term of the entity type and 0 for the absence of the term of the correct type; alternatively, it may be a measure of the probability that the correct term occurs in the answer candidate's passage.