Knowledge base construction
In various examples there is a knowledge base construction and/or maintenance system for use with a probabilistic knowledge base. The system has a probabilistic generative model comprising a process for generating text or other formatted data from the knowledge base. The system has an inference component configured to generate inference results, by carrying out inference using inference algorithms, run on the probabilistic generative model, in either a forward direction whereby text or other formatted data is generated, or a reverse direction whereby text or other formatted data is observed and at least one unobserved variable of the probabilistic generative model is inferred. The inference component is configured to update the knowledge base using at least some of the inference results.
1 . A knowledge base construction system for updating a probabilistic knowledge base with a probabilistic generative model, the system comprising:
a processor; and
a memory storing the probabilistic generative model, an inference component, and a source tracker,
the probabilistic knowledge base storing prior probability distributions of variables of the probabilistic generative model, the prior probability distributions representing likelihoods of particular values of the variables, and
the inference component comprising computer executable instructions that upon execution by the processor:
generate inference results using an inference algorithm that executes on the probabilistic generative model with unlabeled training data as input to the inference algorithm:
in a forward direction resulting in observed formatted data and a posterior distribution being generated, the posterior distribution representing a likelihood of possible inference results across a document corpus given a particular query, being approximated by applying heuristics, and converted into a plurality of search queries executed across the document corpus resulting in an observed variable of the probabilistic generative model being populated, and
in a reverse direction whereby the formatted data is observed and an unobserved variable of the probabilistic generative model is inferred; and
update the prior probability distributions stored in the probabilistic knowledge base using a portion of the inference results, and
the source tracker configured to determine source tracking information of the observed formatted data, the source tracking information comprising, for individual observations from the observed formatted data, security information of the observation,
wherein the source tracker is configured to store the source tracking information in the probabilistic knowledge base such that other applications are able to retrieve the source tracking information by querying the probabilistic knowledge base,
wherein the source tracker is configured to determine that an inferred fact has a single restricted source based on the source tracking information, and apply the same restriction to a query to the probabilistic knowledge base retrieving that inferred fact based on the determining.
2 . The knowledge base construction system of claim 1 , wherein the probabilistic knowledge base comprises an entity type having an entity, wherein the probabilistic generative model comprises a schema comprising a property type for the entity type, the property type comprising a format to convert a property value to a formatted property value.
3 . The knowledge base construction system of claim 2 wherein the formatted property value is part of an instance of the formatted data, wherein the probabilistic generative model comprises a template that combines the formatted property value into the formatted data.
4 . The knowledge base construction system of claim 3 wherein the template is selected based on template learning, wherein a template variable is inferred in template learning and another variable of the probabilistic generative model is observed.
5 . The knowledge base construction system of claim 1 wherein the source tracker is configured to store, for the individual observations from the observed formatted data, conflicting values of the source tracking information.
6 . The knowledge base construction system of claim 5 wherein the source tracker is configured to track an influence of the individual observations on unobserved or observed variables of the probabilistic generative model.
7 . The knowledge base construction system of claim 5 having user settings comprising source tracking rules, thresholds or criteria set according to the user settings, and whereby the inference component takes into account the source tracking rules, thresholds or criteria when observing a variable of the probabilistic generative model and/or inferring the variable of the probabilistic generative model.
8 . The knowledge base construction system of claim 5 comprising a type system comprising a library of built-in property types, where a property is a characteristic of an entity type of the probabilistic knowledge base, and wherein each built-in property type has a format to convert a property value to a formatted property value, and wherein the built-in property types are parameterized.
9 . The knowledge base construction system of claim 8 wherein the probabilistic generative model comprises probability distributions representing belief about the type of one or more properties of entities of the probabilistic knowledge base, and wherein the probability distributions are updated during the forward and the reverse directions.
10 . The knowledge base construction system of claim 8 wherein the library of property types comprises one or more of: a set type representing a plurality of values of the same type, a quantity type having a parameter determining a kind of quantity, a code type specifying a type of a regular expression, and an object type representing compound values of one or more types.
11 . The knowledge base construction system of claim 1 ,
wherein the probabilistic generative model comprises a process linking, for a given entity type of the probabilistic knowledge base, a schema of the given entity type with an instance of the given entity type, a template, and formatted data referring to the instance of the given entity type; and
wherein the process has a probability distribution representing uncertainty of the instance of the given entity type, schema, template, and formatted data referring to the instance of the given entity type.
12 . The knowledge base construction system of claim 1 wherein the probabilistic generative model comprises one or more of a property list model, a page model, and a value noise model, the property list model comprising a property list template representing name and value pairs, the page model comprising an encoding of an assumption in the probabilistic generative model that the number of entities referred to on a single page is smaller than the total number of entities referred to in a corpus of pages, the value noise model comprising a process which adds type-specific noise to a property value.
13 . A computer-implemented method for updating a probabilistic knowledge base, the method comprising:
storing, at a memory, a probabilistic generative model, the probabilistic knowledge base storing prior probabilistic distributions of variables of the probabilistic generative model, the prior probabilistic distributions representing likelihoods of particular values of the variables;
generating inference results using an inference algorithm that executes on the probabilistic generative model with unlabeled training data as input to the inference algorithm:
in a forward direction resulting in observed formatted data and a posterior distribution being generated, the posterior distribution representing a likelihood of possible inference results across a document corpus given a particular query, being approximated by applying heuristics and converted into a plurality of search queries executed across the document corpus resulting in an observed variable of the probabilistic generative model being populated, and
in a reverse direction whereby the formatted data is observed and an unobserved variable of the probabilistic generative model is inferred;
updating the prior probability distributions stored in the probabilistic knowledge base using a portion of the inference results;
determining source tracking information of the observed formatted data, the source tracking information comprising, for individual observations from the observed formatted data, security information of the observation;
storing the source tracking information in the probabilistic knowledge base such that other applications are able to retrieve the source tracking information by querying the probabilistic knowledge base; and
determining that an inferred fact has a single restricted source based on the source tracking information, and apply the same restriction to a query to the probabilistic knowledge base retrieving that inferred fact based on the determining.
14 . The method of claim 13 wherein for individual observations from the observed formatted data, storing the source tracking information includes storing conflicting values of source tracking information.
15 . The method of claim 14 comprising learning facts comprising values of properties of entities of the probabilistic knowledge base, by observing variables of the probabilistic generative model which relate to the observed formatted data, and carrying out the inference in the reverse direction to update a probability distribution related to properties of the entities,
wherein the source tracking information comprises, for individual observations from the observed formatted data, privacy information of the observation.
16 . The method of claim 14 comprising learning templates, by observing variables of the probabilistic generative model which relate to the observed formatted data, and carrying out the inference in the reverse direction to update a probability distribution related to the templates,
wherein, the source tracking information comprises, for individual observations from the observed formatted data an identifier of a source of the observation.
17 . The method of claim 14 ,
wherein, the source tracking information comprises, for individual observations from the observed formatted data, security information of the observation, and
wherein the probabilistic knowledge base initially comprises a seed example of an entity with at least two property values, and the method comprises bootstrapping construction of a schema and a plurality of templates by carrying of inference in the reverse direction over the probabilistic generative model.
18 . The method of claim 17 further comprising:
iteratively carrying out fact retrieval and template learning using a plurality of names of entities, wherein fact retrieval comprises inferring values of properties of the entities of the probabilistic knowledge base given at least observations of the observed formatted data; and
generating a first alternative conflicting value distribution and a second alternative conflicting value distribution for a particular entity name and property,
wherein the first alternative conflicting value distribution comprises a first alternative with the most number of observations in the observed formatted data, and
wherein the second alternative comprises, if present, a second alternative with the second most number of observations in the observed formatted data, otherwise the second alternative equaling the first alternative.
19 . The method of claim 18 wherein a search engine or a conversational assistant:
receives a query from a user computing device,
retrieves knowledge about the unobserved variable from the updated prior probability distributions stored in the probabilistic knowledge base in response to the query,
includes the retrieved knowledge about the unobserved variable in a response, and
transmits the response to the user computing device.