Natural Language Questions for the Web of Data

transcript

Mohamed Yahya1, Klaus Berberich1, Shady Elbassuoni2

Maya Ramanath3, Volker Tresp4, Gerhard Weikum1

1 Max Planck Institute for Informatics, Germany2 Qatar Computing Research Institute

3 Dept. of CSE, IIT-Delhi, India 4 Siemens AG, Corporate Technology, Munich, Germany

EMNLP 2012

Example of question

• “Which female actor played in Casablanca and is married to a writer who was born in Rome?”.

• Translation to SPARQL :– ?x hasGender female – ?x isa actor– ?x actedIn Casablanca_(film) – ?x marriedTo ?w– ?w isa writer– ?w bornIn Rome

• Characteristics of SPARQL :– Complex query– good results– Difficult for the user

• Author wants: automatically create such structured queries by mapping the user’s question into this representation

Translate qNL to qFL

• qNL → qFL

• qNL : natural language question

• qFL : formal language query– (SPARQL 1.0)

• YAGO2s is a huge semantic knowledge base, derived from Wikipedia, WordNet and GeoNames.

• http://www.mpi-inf.mpg.de/yago-naga/yago/

sample facts from Yago2

• Examples of relations:– type, subclassOf, and actedIn.

• Examples of class:– person and film.

• Examples of Entities :– Entities are represented in canonical form such as ‘Ingrid_Bergman’ and ‘Casablanca_(film)’.– special type of entities : strings, numbers, and dates.

DEANNA

• DEANNA (DEep Answers for maNy Naturally Asked questions)

question sentence

• qNL = (t0, t1, ..., tn).

• Phrase = (ti, ti+1, ..., ti+l) q⊆ NL, 0 ≤ i, 0 ≤ l ≤ n• Phrase focus on entities, classes, and relations

Phrase detection

Phrases are detected that potentially correspond to semantic items such as‘Who’, ‘played in’, ‘movie’ and ‘Casablanca’.

Phrase detection

• A detected phrase p is a pair < Toks, l >– Toks : phrase– l : label (l {concept, relation})∈

• Pr : {<*, relation >}

• Pc : {<*, concept >}

concept detection

• works against a phrase-concept dictionary– phrase-concept dictionary : instances of the

means relation in Yago2

relation detection

• rely on a relation detector based on ReVerb (Fader et al., 2011) with additional POS tag patterns, in addition to our own which looks for patterns in dependency parses.

Phrase Mapping

• each phrase is mapped to a set of semantic items.

• To map concept phrases:– also relies on the phrase-concept dictionary.

• To map relation phrases:– rely on a corpus of textual patterns to relation

mappings of the form– {‘play’, ‘star in’, ‘act’, ‘leading role’} → actedIn– {‘married’, ‘spouse’, ‘wife’} → marriedTo

Example of Phrase Mapping

• ‘played in’ can either refer to the semantic relation actedIn or to playedForTeam and

• ‘Casablanca’ can potentially refer to Casablanca_(film) or Casablanca,_Morocco.

Dependency Parsing & Q-Unit Generation

Dependency parsing

• Dependency parsing identifies triples of tokens,or triploids– <trel, targ1, targ2>, where trel, targ1, targ2 q∈ NL

– trel : the seed for the relation phrase

– targ1 , targ2 : seeds for the concept phrase.– there is no attempt to assign subject/object roles

to the arguments.

Which female actor played in Casablanca?who is married to a writer who was born in Rome?

Q-Unit Generation

• By combining triploids with detected phrases, we obtain q-units.

• q-unit is a triple of sets of phrases, – <{prel P∈ r}, {parg1 P∈ c}, {parg2 P∈ c}>– trel p∈ rel , targ1 p∈ arg1 , and targ2 p∈ arg2 .

Joint Disambiguation

goal of the disambiguation step

• each phrase is assigned to at most one semantic item

• resolves the phrase boundary ambiguity– (only nonoverlapping phrases are mapped)

resulting subgraph for the disambiguation graph of Figure 3

Disambiguation Graph

• Joint disambiguation takes place over a disambiguation graph DG = (V, E), – V = Vs V∪ p V∪ q

– E = Esim E∪ coh E∪ q

Type of vertices

• V = Vs V∪ p V∪ q – Vs : the set of s-node • s-node is semantic items

– Vp : the set of p-node • p-node is phrases• Vrp : the set of relation phrases

• Vrc : the set of concept phrases

– Vq : a set of placeholder nodes for q–units

Type of edges

• Esim V⊆ p × Vs

– a set of weighted similarity edges

• Ecoh V⊆ s × Vs

– a set of weighted coherence edges

• Eq V⊆ q × Vp × d, d {rel, arg1, ∈arg2}– Called q-edge

Cohsem (Semantic Coherence)

• define the semantic coherence (Cohsem) – between two semantic items s1 and s2 as the

Jaccard coefficient of their sets of inlinks.• Three kinds of inlink– InLinks(e)– InLinks(c)– InLinks(r)

InLinks(e)• InLinks(e): the set of

Yago2 entities whose corresponding Wikipedia pages link to the entity.

• e.q. – Let e = Casablanca– InLinks(Casablanca) =

{Marwan_al-Shehhi , Ingrid_Bergman, …, Morocco…}

InLinks(c)• InLinks(c) = ∪e c ∈ Inlinks(e)• e.q. let c = wikicategory_Metropolitan_areas_of_Morocco

– InLinks(wikicategory_Metropolitan_areas_of_Morocco) = InLinks(Casablanca) InLinks(Marrakech) InLinks(Fes) ∪ ∪ ∪InLinks(Agadir) InLinks(Safi,_Morocco) InLinks(Oujda) ∪ ∪ ∪InLinks(Tangier) InLinks(Rabat)∪

InLinks(r)

• InLinks(r) = ∪(e1, e2) r ∈ (InLinks(e1) ∩ InLinks(e2))

Similarity Weights

• For entities– how often a phrase refers to a certain entity in

Wikipedia.• For classes– reflects the number of members in a class

• For relations– reflects the maximum n-gram similarity between

the phrase and any of the relation’s surface forms

Disambiguation Graph Processing

• The result of disambiguation is a subgraph of the disambiguation graph, yielding the most coherent mappings.

• We employ an ILP to this end.

Definitions (part1)

Definitions (part2)

objective function

Constraints(1~3)

Constraints(4~7)

Constraints(8~9)

This is not invoked for existential questions

resulting subgraph for the disambiguation graph of Figure 3

Query Generation

• not assign subject/object roles in triploids and q-units

• Example:– “Which singer is married to a singer?”• ?x type singer , ?x marriedTo ?y , and ?y type singer

5 Evaluation

• Datasets• Evaluation Metrics• Results & Discussion

Datasets• author's experiments are based on two collections of questions:

– QALD-1• 1st Workshop on Question Answering over Linked Data (QALD-1)• the context of the NAGA project

– NAGA collection• The NAGA collection is based on linking data from the Yago2 knowledge base

• Training set– 23 QALD-1 questions – 43 NAGA questions

• Test set– 27 QALD-1 questions – 44 NAGA questions

• Get hyperparameters (α, β, γ) in the ILP objective function.• 19 QALD-1 questions in Test set

Evaluation Metrics

• author evaluated the output of DEANNA at three stages– 1. after the disambiguation of phrases– 2. after the generation of the SPARQL query– 3. after obtaining answers from the underlying linked-data

sources• Judgement– two human assessors who judged whether an output item

was good or not– If the two were in disagreement , then a third person

resolved the judgment.

disambiguation stage

• The task of judges – looked at each q-node/s-node pair, in the context

of the question and the underlying data schemas, – determined whether the mapping was correct or

not – determined whether any expected mappings were

missing.

query-generation stage

• The task of judges– Looked at each triple pattern– determined whether the pattern was meaningful

for the question or not– whether any expected triple pattern was missing.

query-answering stage

• the judges were asked to identify if the result sets for the generated queries are satisfactory.

•Micro-averaging • aggregates over all assessed items

regardless of the questions to which they belong.

•Macro-averaging • first aggregates the items for the same

question, and then averages the quality measure over all questions.

•For a question q and item set s in one of the stages of evaluation

•correct(q, s) : the number of correct items in s•ideal(q) : the size of the ideal item set•retrieved(q, s) : the number of retrieved items

•define coverage and precision as follows:cov(q, s) = correct(q, s) / ideal(q)

prec(q, s) = correct(q, s) / retrieved(q, s).

Conclusions

• Author presented a method for translating natural language questions into structured queries.

• Although author’s model, in principle, leads to high combinatorial complexity, they observed that the Gurobi solver could handle they judiciously designed ILP very efficiently.

• Author’s experimental studies showed very high precision and good coverage of the query translation, and good results in the actual question answers.

Natural Language Questions for the Web of Data

Documents