Presentation is loading. Please wait.

Presentation is loading. Please wait.

Relational Inference for Wikification

Similar presentations


Presentation on theme: "Relational Inference for Wikification"— Presentation transcript:

1 Relational Inference for Wikification
Xiao Cheng and Dan Roth Task Definition Annotate input text with disambiguated Wikipedia titles: Motivation Current system already achieves the above performance, although the performance levels off at around 75%~80% F1 Limitation of Bag-of-words representation Relational Inference Identify Textual Relations Retrieve Relational Knowledge Formulate the inference problem Rerank via constraints Promote candidate pair: e 3 1 = Slobodan_Milošević e 4 21 = Socialist_Party_of_Serbia Textual Relations Demo: Evaluations Achieves significant improvement over the previous state-of-the-art systems Run the Relational Inference Wikifier (RI) “as-is” without retraining on the target domain, still obtains significant performance gain over our previous submitted Entity Linking system (Cogcomp). Discussion We are interested in extracting high-precision textual relations that help with disambiguation. Specifically, we focus on the following types of relations: Syntactico-semantic relations (Chan & Roth ‘10) Coreference relations Acronyms, partial names, nominal mentions ...ousted long time Yugoslav President Slobodan Milošević in October. Mr. Milošević's Socialist Party… Blumenthal (D) is a candidate for the U.S. Senate seat now held by Christopher Dodd (D), and he has held a commanding lead in the race since he entered it. But the Times report has the potential to fundamentally reshape the contest in the Nutmeg State. Chris Dodd From Wikipedia, the free encyclopedia The New York Times Connecticut Democratic Party (United States) United States Senate Richard Blumenthal Type Example Premodifier Iranian Ministry of Defense Possessive NYC’s stock exchange Formulaic Chicago, Illinois Preposition President of the US Argument 1 Relation Type Argument 2 Yugoslav President apposition Slobodan Milošević coreference Milošević possessive Socialist Party Slobodan Milošević From Wikipedia, the free encyclopedia Socialist Party of Serbia From Wikipedia, the free encyclopedia Relation Retrieval founded Uses DBPedia and Wikipedia page link relations as our knowledge base Retrieve lexically similar candidates and filter q1=(Socialist Party of France,?, *Milošević*) q2=(Slobodan Milošević,?,*Socialist Party*) Mubarak, the wife of deposed Egyptian President Hosni Mubarak, … Bag-of-words loses important relational information Modeling constraining interaction between concepts Need to link Mubarak to Suzanne Mubarak Identify relation (Mubarak, wife, Hosni Mubarak) Promote the pair of candidates coherent with text Mubarak wife Egyptian President Hosni Mubarak Relation Inference For each pair of entity candidates 𝑒 𝑖 𝑘 and 𝑒 𝑗 𝑙 , 𝑟 𝑖𝑗 (𝑘,𝑙) represents whether we found a relation in the text between their mentions AND a relation in our knowledge base 𝑤 𝑖𝑗 (𝑘,𝑙) either rewards or penalizes a relation 𝑟 𝑖𝑗 (𝑘,𝑙) for its coherency with the text. Relation scoring Relaxes constraint when ambiguity exists Scores each retrieved relation 𝜎 for each query 𝛼 𝜎 is the relation weight for different knowledge source 𝑆𝑖 𝑚 𝜎,𝑞 is the lexical similarity between the query 𝑞 and the retrieved relation 𝜎 𝑍 is the normalization factor so that the weights for each pair of mentions is always between 0 and 1 Special handling of “local knowledge” Creates NIL entity candidate for inference propagation, so that locally extracted high precision knowledge can be considered across long-range textual relations Uses off-the-shelf Integer Linear Programming (ILP) packages to optimize the objective function High-level algorithm description ek3=Milošević sk3 Slobodan_Milošević .7 Milošević_(surname) .1 Boki_Milošević Alexander_Milošević .05 ek4=Socialist Party sk4 Socialist_Party_(France) .23 Socialist_Party_(Portugal) .16 Socialist_Party_of_America .07 Socialist_Party_(Argentina) .06 Socialist_Party_of_Serbia .0 We showed that both linguistic and world knowledge, specifically the ability to use relational information, are crucial in the task of Wikification. We also introduced an extensible and efficient inference framework that leverages better language understanding. It will be very interesting to accumulate our knowledge about NIL entities so that our system will be able to fully address the Entity Linking task and potentially move to encylopedia resources other than Wikipedia. The performance gains and error analysis also calls for joint entity typing, coreference and disambiguation. 𝑟 34 (1,21) =1 This research is sponsored by DARPA under agreement number FA , and partly supported by the IARPA under contract number D11PC20155, by the ARL under agreement W911NF , and by the Multimodal Information Access & Synthesis Center at UIUC.


Download ppt "Relational Inference for Wikification"

Similar presentations


Ads by Google