Docstoc

The Semantic Web

Document Sample
The Semantic Web Powered By Docstoc
					RECOGNIZING TEXTUAL ENTAILMENT RTE-2 VENICE, ITALY April 10, 2006

University of Texas at Dallas

Computational Textual Inference
Sanda Harabagiu
1

University of Texas at Dallas

Outline
1.

2.
3. 4.

5.
6.

Forms of textual entailment Paraphrases and Entailment Contradictions and entailments A Special Case: Temporal Inference Applications Strategies

2

University of Texas at Dallas

Textual Entailment
 

The problem did not originate in the vacuum !!! There are many possible way of characterizing the forms of textual entailment
1. Lauri Karttunen and Annie Zaenen (PARC) have proposed a categorization in terms of Logical entailment vs. Conversational inference vs. Plausible inference vs. Presuppositions 2. Lauri Karttunen and Annie Zaenen have also raised the issue of Strong implicatives and Semi-implicatives 3. Textual entailments could also be characterized by: a) the form of knowledge they employ b) the inference framework that justifies them
3

University of Texas at Dallas

Entailments and implicative inferences



Logical entailment  Many terrorists were killed.[ Some terrorists were killed.



Conversational inference
 Bush

was able to convince McCain to campaign for him. terrorists were killed. Some terrorists were not killed.

… but Bush chose not to do it.
 Many

… in fact all of them.


Plausible inference
 Many

terrorists were shot. Many terrorist were killed. realized that the campaign was in trouble. did not realize that the campaign was in trouble.
4



Presupposition
 Kerry  Kerry

University of Texas at Dallas

Strong Implicatives
In affirmative sentences, strong positive implicatives such as manage entail that the embedded proposition is true, while strong negative implicatives such as fail entail that the embedded proposition is false. In negative sentences, the polarity of the entailment is reversed. Strong implicatives also carry presuppositions. (Otherwise they would be devoid of any meaning.) Kerry managed to hold on to his seat.
Entails: Presupposes: Kerry held on to his seat. It was difficult for Kerry to hold on to his seat.

Bush didn’t manage to find any oil.
Entails: Presupposes: Bush didn't find any oil. It was difficult for Bush to find oil.

The administration failed to track down the perpetrators.
Entails: The administration didn't track down the perpetrators. Presupposes:The administration tried or should have tried to track down the perpetrators.

Bush didn’t fail to read a report warning of al-Qaida attacks.
Entails: Bush read a report warning of al-Qaida attacks. Presupposes: Bush tried or should have tried to read the report.

Other strong implicative constructions:
Positive: bother to, happen to, get around to, succeed, take the trouble… Negative: forget to, avoid (-ing), neglect to, …
5

University of Texas at Dallas

Semi-Implicatives
In negative sentences, positive semi-implicatives entail that the embedded proposition is false; in affirmative sentences there is no entailment but there may be a "conversational implication" that the embedded proposition is true. Kerry wasn't able to convince McCain to run with him. Entails: Kerry didn't convince McCain to run with him. Kerry was able to convince McCain to run with him. Doesn't entail, strictly speaking, that Kerry convinced McCain to run with him. It is not a

Kerry would have been able to convince McCain to run with him.
In the actual world he wasn't able.

contradiction to say "Kerry was able to convince McCain to run with him but chose not to do it." However, in the absence of any contradictory information, the sentence is misleading if McCain was not convinced by Kerry.

More semi-implicative constructions:  She didn’t have a chance / time / money / courage… to follow your advice.  He wasn’t bold / clever / strong … enough to meet the challenge.


yield a negative entailment under negation, a positive conversational implicature in affirmative sentences if there is no counterindication. yield a negative entailment in affirmative sentences, a positive conversational implication in negative sentences

I was too scared / timid / stupid / distracted … to do what I promised.

6

University of Texas at Dallas

Knowledge used for Textual Entailment



Textual inference is difficult because of the various forms of knowledge that a system needs to have available for deriving inference:
     

Temporal knowledge Spatial knowledge Causal knowledge Commonsense knowledge Newsworthy knowledge Domain knowledge

7

University of Texas at Dallas

Examples


Temporal Knowledge:
T1: Sardar Patel faced imprisonment for the first time when he was assisting Gandhiji in the Salt Satyagraha. H1: Sardar Patel has never been in prison before. (PLAUSABLE) H2: Sardar Patel was convicted of a crime at least once in his life. (TRUE) T2: Herbicide use in some areas of the U.S. was delayed earlier in the year by heavy rains. H: Herbicides were used this year in the U.S. (TRUE)
8

University of Texas at Dallas

Examples


Spatial knowledge:
T1: The unemployed took to the streets of the German capital, Berlin, mirroring protests around the country. H1: The protests took place only in Berlin (FALSE). T2: John left Venice in the morning, taking the 10 am flight to London. H2: John was in Venice at 8 am. (TRUE) H3: John will be in London in the evening (PLAUSIBLE).

9

University of Texas at Dallas

Examples


Causal Inference
T: Darryl Strawberry recently avoided imprisonment when a judge sentenced him to a drug treatment center for violating his probation. H: Darryl Strawberry was sentenced because he violated his probation (TRUE).
Interactions with temporal inference H’: Darryl Strawberry has never been in prison. (UNKOWN)
10

University of Texas at Dallas

Examples


World Knowledge:
T1: Many cell phones have built-in digital cameras. H1: Some cell phones can be used to take pictures.(PLAUSIBLE: the cameras in the cell phones must work)
T2: Mr. Radley ordered a 16 ounce slab of slowly roasted Black Angus Prime Rib. H2: Radley is a vegetarian. (FALSE: Vegetarians don't eat meat; and people usually intend to eat what they order.)
11

University of Texas at Dallas

Trustworthiness
Reuters reports that Congress has passed the use of force resolution. Statement: Congress has passed the use of force resolution. Source: Reuters Author: uncommitted (reports)
Although the author is noncommittal, the reader may choose to take the statement as true if the source is trustworthy. trustworthy = well-informed and honest Reuters reports that the UN said on Monday that the Iraqis claim that Iraq has fully cooperated with the inspectors.

12

University of Texas at Dallas

Domain Knowledge


Examples:
T: In January-February 1997, China supplied Iran with 40,000 barrels of calcium hypochlorite. H: China provided Iran with decontamination materials. (TRUE, calcium hypochlorite is a chemical-biologicalradiological decontamination agent)
Interaction with world knowledge: Providing any amount of decontamination materials entails that some decontamination materials were provided.
13

University of Texas at Dallas

Our approach to textual entailment in October 2005



Our approach to textual/lexical entailment seeks to benefit from :  The availability of several forms of linguistic knowledge  Good alignment between a question and a text that entails it  Recognition of paraphrases  Identification of temporal information  Identification of intentions  Processing of semantic information (semantic frames, named entities)

EXAMPLE: Passage: Tehran continues to seek considerable production technology, training, expertise, equipment and chemicals from entities in Russia and China that could be used to help Iran reach its goal of an indigenous nerve agent production capability. Question: Does Iran want to be a self-sufficient producer of CW? Answer #1: Yes.; Polarity: true; Force: plausible; Source: world Because: An entity that has a goal of achieving a certain state wants to bring about that state.

14

University of Texas at Dallas

Knowledge Forms


Features characteristic to eighteen forms of knowledge were developed
Negative implication Positive implication Numeric Information
Date/Time Event Sequence Paraphrase

Intention KW Alternation Paraphrasing
Possibility Uncertainty Pragmatic

RW Knowledge Tense/Aspect Quantification
Antonymy Belief Statements Speech Act



Two independent efforts, students at UTD and the researchers at LCC have identified eighteen different knowledge forms. The two teams shares 12 common classes of knowledge forms. Each team had at least 6 classes that were unique to this effort. Therefore we could not agree on 33% of the classes with the other team!
15

CLASS
University of Texas at Dallas

COUNT

ACCURACY

Negative implication Positive implication Keyword Alternation

32
20 17

34.38%

90.00%
52.94%

Possibility
Paraphrase Speech Act Quantification Intention Event Sequence Uncertainty Belief World Knowledge Tense/Aspect Implicit Possession Pragmatic Antonymy

17
14 13 12 10 10 9 8 7 7 7 2 4

41.18%
85.71% 69.23% 66.67% 60.0% 60.0% 66.67% 50.0% 57.14% 42.86% 100% 0% 0%

The starting point:

Numeric
Date Total

3
2 194

66.67%
0% 57.73%
16

University of Texas at Dallas

Entailments and Paraphrases



Techniques used in discovering textual paraphrases can be used in determining textual entailment.



These techniques have limitations
We have used them with promising results for recognizing textual contradictions as well
17



University of Texas at Dallas

Applications




We have used TE for QA We have also used Textual Entailment for a dialog system that incorporates the QA:
  

Entailment at the dialog analysis level Entailment at the indexing and retrieval level Entailment at the answer justification level

18

University of Texas at Dallas

Startegies




Bag-of-words vs. Bag-of-entailments It is time to generate some strategies, roadmaps.



Thank you!

19


				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:91
posted:11/3/2009
language:English
pages:19