1. Answering System
Definition: According to question (query) directly to search out answers instead of documents
2. The question answering system based Knowledge Mapping
classification:
1) Semantic Analysis (Semantic Parsing): Questions query into a formal, structured query to get an answer
2) semantic search (Answer Retrieval): Simple Search get candidate answers by question and answer candidates do similarity match
3. The main method of quiz
1) Analytical Semantics
2) knowledge-based semantic retrieval of map quiz
3) Neural symbolic computation (that were using the neural network, it generates symbols do, do match ah)
4. semantic representation
1) lambda- calculus logical expression
如:\lambda x.city(x), next_to(x, y)
2) DCS tree
The semantic parser
1) retrieves phrase -> Resource mapping -> Semantic composition
2) two issues: the phrase to resource mapping, text uprising
Question 1 + combination dictionary mapping rules
Dictionary filtering rule generating Dictionary -> train Parser -> filter Dictionary
Align the use of statistical machine translation
Open domain relation extraction learned was born in == PlaceOfBirth
Question 2 Ambiguity
PCCG (essentially MLE, and generation rules)
Model learning no- logical expression
1. back stamp generating any entity with variable substitution -> generates questions -> Learning Dictionary
2. repeat generated using semantic synonymously
3. The answer generated using the answer generation logic expressions, logical expressions as potential variables
6. The semantic retrieval (Answer feature matching question)
1) based on the display characteristics
Based on structural semantic question (lambda calculus) knowledge of the pattern matching sub-optimal FIG.
Pros Links -> determine the subject (Identify Core Inferential Chain)
Single body
2) end
Three questions
a. question shows how learning
How b. Knowledge Graph entities, relationships learning
c. calculating semantic matching
The basic steps
a. entity candidate generating link candidates are peripheral entities
b. candidate ranking
Consider multidimensional similarity
a plurality of angle calculation semantic matching question and knowledge, such as answer path, answer type, answer entity
b. each (answer questions) different parts of focus
The method of cross attention A and Q different parts Attention
Cons: training corpus global information OOV
Solve integration into the global information: Pre-training embedding and multi-task learning
7. Based on Neural symbol
Semantics analytical advantages: a display, high accuracy
Deep learning advantage: large-scale computing, you can learn
Three categories
1) sequence learning
The answer output logical expression
seq2Tree symbol string, a hierarchical tree (multilayer decode)
2) Operation Sequence
Book ahead entity, generating a corresponding output anonymously tree action
Nerve symbolic computation based remote supervision (reinforcement learning)
trick: pre-training courses
3) Stack neural network
Internal neural networks can explain
Reader Annotator, R is selected from column or table, A rank row or table
8. Prospects
1) small corpus database
2) can only handle simple issues (simple Q & A format)
3) Multi-Knowledge
4) interpretability