HackMD
Brain Knowledge Repository 2.0
Like a Knowledge Graph. Unlike any Knowledge Graph.
- Reasoning first knowledge graph
- Knowledge = Enumerate all the options, example Rules of the Game. Constraints my forward path
- Learning = How do I move with the knowledge, which is a better option. Goodness of State, given the action, state & goal. where do I go?
- Reasoning = Using Learning + Knowledge in recursive path
- Factoid question are not answerable via KG 1.0 APIs
- You shall know an Entity by the company it keeps in the Knowledge Graph!
Objective : The Reasoning Brain
E2E application of the Brain Reasoner.
- English Question –-> Brain Natural Language Engine (BNLE) –-> KnowledgeQuestion(state = ReasoningState)
- KnowledgeQuestion –-> Brain Knowledge Repository 2.0 –-> KnowledgeAnswer
- KnowledgeAnswer –-> Brain NLP Synthesis Engine (BNSE) –-> English Answer
Example
-
English Question –-> Brain Natural Language Engine (BNLE) –-> KnowledgeQuestion(state = ReasoningState)
- Finding the BrainToken
- Converting english question to BrainToken langugauge format
What is the population of India?
What is BrainToken("/common/attribute/country/population") of BrainToken("common/entity/country/1")common/entity/country/1 represents India node in the KG
- Remove stop words
[What BrainToken("/common/attribute/country/population")] BrainToken("common/entity/country/1")
- Create Knowledge Question/Reasoning State
-
KnowledgeQuestion –-> Brain Knowledge Repository 2.0 –-> KnowledgeAnswer
1.Action Given a state, where what all we can do?
Generating Path/Action/QueryAs the State.GIVEN token as EntityInstance and there is only one instance hence we can reach to country collection's doc/node '1' in the KG. As we are this node and no other GIVEN token is there hence the only option is to read attributes of node 1. This new information is passed to Update State.
One optimization possible is if we can utalized any token with same entityType as a filter, ex in this case ASK.BrainToken if of same entityType hence applicable to be applied as filter
Action = Query(EntityID(1), Attribute(/common/attribute/country/population))
1.1 State.GIVEN.BrainToken == EntityInstance
1.1.1 You can read all the types of attributes (schema) ex. cou ntry as entity type will have population, gdp, area etc 1.1.2 You can read all the attributes values ex. area = BrainQuantity(1234), pop = BrainQuantity(345) 1.1.3 You can read all predicate types ex. country can have [has_president, belongs_to] 1.1.4 You can move to take a path using a predicate. ex. county has_president and reach to person node 1.1.5 You can read all qualifiers1.2 State.GIVEN.BrainToken == EntityType
1.2.1 You can read all the types of attributes (schema) ex. country as entity type will have population, gdp, area etc 1.2.2 You can fetch all subclasses 1.2.3 You can list all instances of the entity type. ex, all country instances 1.2.4 You can read all predicate types ex. country can have [has_president, belongs_to] 1.2.5 You can read all qualifier schema1.3 State.GIVEN.BrainToken == PredicateType
2.Update State Generate a reward state
Input:
Action = Query(EntityID(1), Attribute(/common/attribute/country/population))
Query Response:
/common/attribute/country/
/common/attribute/country/1
/common/attribute/country/population
BrainQuantity("1234")Reward State = {State + Action}
3.Stop Mechanims: Is the given State is a GOAL State ?
Below state will become the Goal State, as the no action to take and ask is full_filled.
- Generate Knowledge Answer
For now we will consider we are just passing the goal state as KnowledgeAnswer but this section will evolve any logic regarwds to this.
KnowledgeAnswer(state= GoalState)
- Generate Knowledge Answer
-
KnowledgeAnswer –-> Brain NLP Synthesis Engine (BNSE) –-> English Answer
3.1 Template Classification apporach
Question : What is the BrainToken(X) of BrainToken(Y)?
Answer Template: BrainToken(X) of BrainToken(Y) is BrainToken(X.Value)3.2 Template value replacement
BrainToken("/common/attribute/country/population") of BrainToken("common/entity/country/1") is BrainToken(BQ("/common/attribute/country/population/1234"))
3.3 Converting to natual english (DL+Bert)
Population of India is 138 Crores
Scope
In this document scope, we will focus on factoid question first.
Factoid Question
-
What is BrainToken(Attribute) of BrainToken(EntityInstance)?
- What is the length of Ganga?
- Height of Virat Kohali?
- No of centures of MS Dhoni?
Can we answer above with State Thinking? Yes
- Height of Virat Kohali?
-
Step:1
- Start from ASK, mark the "/common/attribute/person/height" VISITED=TRUE
- get_next_context = CurrentContext("Height") + LinkedContext("Height", STATE)
- // LinkedContext is coming from NLP layer
- get_next_context ={"Virat Kohli", "Height of Virat Kohli"}
- // State would not change because of the next context information.
- query.add(getAttributeQuery(/common/attribute/person/height, /common/entity/person/1))
- execute(query)
Approach: Search problem that grows the graph through staged state-actions
Query Graph morphology
A query graph should have exactly one lambda variable to denote the answer, at least one grounded entity, and zero or more existential variables and aggregation functions.
Query graph parsing
Query graph generation, formulated as a search problem with staged states and actions. Each state is a candidate parse in the query graph representation and each action defines a way to grow the graph.
Who was U.S. president after the Civil War started
Topic entity / Entity Linking
- Topic Entity The score returned by the entity linking system is directly used as a feature.
- We use the entity linking approach proposed by (Yang and Chang, 2015) to detect entities mentioned by the given question.
Core Inferential Chain
-
For each detected entity s, we treat it as a subject constant vertex. Based on the KB, for each unique KB ‘path’ from s, where a KB ‘path’ means one hop predicate p0 or two hop predicates p1-p2, we construct a basic query graph hs, p0, xi or hs, p1-ycvt-p2, xi. ycvt and x are variable vertices, and x denotes the answer. For example, the basic query graph B in Figure 2 can be represented as hUnited States, officials-y0-holder, xi.
-
We use similarity scores of different CNN models described in Sec. 3.2.1 to measure the quality of the core inferential chain
- PatChain compares the pattern (replacing the topic entity with an entity symbol) and the predicate sequence.
- QuesEP concatenates the canonical name of the topic entity and the predicate sequence, and compares it with the question.
These two CNN models are learned using pairs of the question and the inferential chain of the parse in the training data
Constraint detection and binding
- When a constraint node is present in the graph, we use some simple features to check whether there are words in the question that can be associated with the constraint entity or property. Examples of such features include whether a mention in the question can be linked to this entity, and the percentage of the words in the name of the constraint entity appear in the question.
- To measure the similarity between the path p (e.g., director) and the ‘context pattern’ (e.g., “directed by e1”) of constraint Ci , we adopt a convolutional neural network (CNN) model
References
Who was U.S. president after the Civil War started?
BrainTokenization
Rough Work
Steps:
-
Tokenize
-
remove stop words
-
Identify BrainTokens
entityType
predicateType
AttributeType
entityInstance
predicateInstance
AttributeInstance -
Temporal Identification
standard lib
Bag = []
What/When/Which/Where/Who & How [W5H1] - film = Y^ = Looking for instance of Y^
-
Basic Dependeny Graph Generation
/common/actor/1 ==> Amitabh Bachhan
(star_in)(context : starred by Amitabh bachhan) - ngram - AnnotationService
[] –> star_in –> (Y1) –-> directed_by –> X1
–> produced_by –> X2
–> genre –> X3
Y2 Y3[/common/director/1] –> instanceOf (Y1^, X1^, X2^, X3^) – > X1
[/common/director/1] –> valueCheck (X1[1…n]) – > X1[k]
[/common/entity/flim] –> typeCheck [Y1^, X1^, …] –> Y1^
-
Add Constraints
Temporal
1. Implicit // TODO : 2. Explicit [https://stanfordnlp.github.io/CoreNLP/ner.html] after year 2000 [Before/After/Since/From] film -- year attributeType AttributeType[/common/attribute/film/year] BrainFilter /common/attribute/film/year > 2000 screened after 2000 screened >> flim.year_of_release result : /common/attribute/film/year > 2000 brainNameService : screen 3. Ordinal richest, poorest, fasted, maximum , first maximum, richest - maxAtN poor, cheepest - minAtN max[/common/attribute/film/revenue] min[/common/attribute/film/revenue][A1] –> star_in [Y1, Y2] –> directed_by [X1]
Answer : Y1, Y2
Dependency Graph >> Query >> ArangoDB
for f in films
f.outbound actor
actor.id = actor/1
&& f.outbound director
director.id = director/1
BrainExpression where
BrainToken 'op' LITERAL
/common/attribute/film/year > 2000
max[BrainToken]
max[/common/attribute/film/revenue]
DSL
DSL
Looking for an instance of Film Entity which is connect to Director and Actor via relationship and that also director/1 and actor/1
message State {
repreated entity_type
repreated entity
repreated predicate_type
}
message BaseGraph {
subject // actor/1 [Actor], context
string relation // star_in , context
repeated string possible_two_hop_options/path/query, [director. .....]
map<type, string> context
}
message Constraint {
// /common/attribute/film/year > 2000
repeated BrainExpression = 1;
//
}
message AnswerState {
//KNOWN vs UNKNOW
repeated brainToken // instance of flim, flim,
??
}
message DependecyGraph {
AnswerState
BaseGraph base_grap;
State
Constraint
}
for f in films
f.outbound actor
actor.id = actor/1
&& f.outbound director
director.id = director/1
FILTER
f.year > 2000
for ....
...
