Predicting lexical answer types in open domain question and answering (QA) systems
In an automated Question Answer (QA) system architecture for automatic open-domain Question Answering, a system, method and computer program product for predicting the Lexical Answer Type (LAT) of a question. The approach is completely unsupervised and is based on a large-scale lexical knowledge base automatically extracted from a Web corpus. This approach for predicting the LAT can be implemented as a specific subtask of a QA process, and/or used for general purpose knowledge acquisition tasks such as frame induction from text.
1. A computer-implemented method of inferring a lexical answer type from a question, said method comprising:
extracting at least one syntactic frame from a question string;
querying a lexical knowledge database to obtain at least one replacement term for a focus of said at least one syntactic frame, wherein said focus is a part of the question indicating a lexical answer type to the question,
substituting at least one of said replacement terms with a generalized type information term using said database;
ranking said replacement terms; and
selecting a top-ranked replacement term as an inferred lexical answer type to said question.
2. The method according to claim 1 , wherein said syntactic frames comprise terms selected from the group consisting of: a subject term; a verb term; an object term; and an indirect object term.
3. The method according to claim 1 , wherein extracting a set of syntactic frames from a question string comprises:
using a dependency parser to identify terms in a grammatical structure from said question string to be used as said extracted syntactic frame.
4. The method according to claim 1 , wherein said lexical knowledge database comprises syntactic frames representing knowledge extracted from a corpus of data.
5. The method according to claim 1 , wherein said entity type knowledge comprises frames having terms with is_a relationships.
6. The method according to claim 1 , wherein querying said lexical knowledge database to obtain said at least one replacement term comprises:
applying a Generalized Frame Model to obtain said at least one replacement term.
7. The method according to claim 1 , further comprising:
filtering out said replacement terms if said replacement term is not related to the context of said question.
8. The method according to claim 6 , wherein said lexical knowledge database includes frame structures identified from a corpus of text, a frame structure having one or more slots-value pairs, a slot representing a syntactic role identified by a dependency relation, wherein said applying a Generalized Frame Model comprises:
specifying a frame cut comprising a sub-set of frames having non-empty slot-values for a given subset of slot-values pairs;
specifying a frame abstraction to determine a desired relationship among selected selected slot values from said given subset of slot-values pairs; and,
generating, from said frame cut, plural frame vectors defining a multi-dimensional vector space from which relationships among selected selected slot values is determined,; and,
processing said frame vectors to determine said desired relationship among selected selected slot values as defined by said abstraction.
9. The method according to claim 7 , wherein said desired relationship includes an inferred type of a slot representing one of: a syntactic subject role or syntactic object role in a corresponding verb and noun phrase.
10. A method for predicting a lexical answer types (LAT) in a question, the method comprising:
applying a frame-extraction utility to a question text to identify all frames involving a question focus, each frame having one or more slots-value pairs with a slot representing a syntactic role identified by a dependency relation, and including a question focus slot;
for each identified frame, creating a query frame structure having a focus slot variable, for each query frame structure, finding in a data corpus, a slot filler for the focus slot variable, said slot filler being part of a question focus from which said LAT is determined,
obtaining a score associated with each said slot filler found for each identified frame
ranking said slot fillers according to said scores; and
selecting a top-ranked slot filler as a predicted LAT to said question,
wherein a programmed processor device performs one or more of said applyuing, creating, finding, obtaining, ranking and selecting.
11. The method as in claim 10 , wherein said data corpus includes frame structures identified from a corpus of text, a frame structure having one or more slots-value pairs, a slot representing a syntactic role identified by a dependency relation, said finding a slot filler comprising:
specifying a frame cut comprising a sub-set of frames having non-empty slot-values for a given subset of slot-values pairs;
specifying a frame abstraction to determine a desired relationship among selected selected slot values from said given subset of slot-values pairs; and,
generating, from said frame cut, plural frame vectors defining a multi-dimensional vector space from which relationships among selected selected slot values is determined; and,
processing said frame vectors to determine said desired relationship among selected selected slot values as defined by said abstraction.
12. The method as claimed in claim 11 , wherein said desired relationship includes an inferred type for a slot filler of said focus slot representing one of: a syntactic subject role or syntactic object role in corresponding verb and noun phrases.
13. The method as claimed in claim 11 , wherein said finding slot fillers for the focus slot variable comprises: tabulating raw-frequency counts of slot-fillers found from said frame cut.
14. The method as claimed in claim 11 , wherein a specified frame cut includes an is_a frame cut including a sub-set of frames having is_a relationships representing associations between slot values and their lexical types, wherein, for each slot filler, said method comprising generating, a predicted LAT from said is_a frame cut.
15. The method as claimed in claim 11 , wherein said finding slot fillers for the focus slot variable further comprises: filtering the slot filler results of predicted LATs by Latent Semantic Analysis (LSA)-based topic similarity with a context of said question.
16. The method as claimed in claim 15 , wherein said filtering comprises:
computing a similarity between said slot filler and the question text; and
discarding any slot-filler whose similarity is below a threshold value.
17. The method as claimed in claim 14 , further comprising:
filtering the slot filler results of predicted LATs by Latent Semantic Analysis (LSA)-based topic similarity with a context of said question; and,
for each slot filler, obtaining a predicted LAT from said is_a frame cut.