QA SYSTEMMaria Alexandropoulou

Max Kaufmann

Alena Hrynkevich

Bug fixes / improvements

• To actually take into account weights from web-results produced by exact and inexact queries

• Correctly post process Right and Left answer placeholder for web results of exact queries

• Return web results for inexact queries and rank them based on bigram overlap with search string

• Reformulations now use corrected past forms of verbs (including irregular) and 3rd person

• Caching of web search results to facilitate parameter tuning

Improved Question ProcessingReformulation Improvements

Not so hard questions we had no chance answering correctly for D2:

• Who is she married to? - Judi Dench• When was it founded? - American Enterprise Institute• When was the division formed? - 82nd Airborne Division• Who was the victim of the murder? - John William King

convicted of murder• For what crime was the deposed leader found guilty? - Pakistani

government overthrown in 1999• At what institute was this procedure done? - cloning of mammals

(from adult cells)

Improved Question ProcessingQuestion series

More complex scenario we did not attempt to address (we process each question in isolation):

Question series1• Who became Tufts University President in 1992? - Tufts University• Over which other university did he preside? - Tufts University

Question series2• What was her husband's title when she married him? - 165.6 - the

Queen Mum's 100th Birthday• What was his title when he died? - 165.7 - the Queen Mum's 100th

Birthday

Improved Question ProcessingA multi-step way of consulting the topic of the question to make it as unambiguous as possible:• Resolved personal, possessive and demonstrative pronouns (ex. "How

long was it used as a defense?" became "How long was Great Wall of China used as a defense?", but "How long did it take to build the Tower of Pisa?" remained intact; “At what institute was this procedure done? became “At what institute was cloning of mammals (from adult cells) procedure done? )

• Expanded notion about some concepts based on target (ex. "What is the division's motto?" became "What is the 82nd Airborne Division motto?" )

• Questions like "How old was Thabo Mbeki when he was elected president?" given topic "Thabo Mbeki elected president of South Africa" remained as is

• When we arrived at conclusion that question does not relate to topic well enough ex. "How old was the dam?" under the topic "Johnstown flood" we simply constructed a reformulations of original question and appended topic's text to each of them before searching

Improved Question ProcessingClassification• Software: Mallet• Classification Algorithm: MaxEnt• Features (inspired Li and Roth 2005)

• Named Entity Unigrams• POS tags for every word• Head of first NP after question word • Head of first VP after question word • Bigrams

• Ex. 171.1 NE_PERSON_Stephen NE_PERSON_Wynn WORDPOS_When_WRB WORDPOS_was_VBD WORDPOS_Stephen_NNP WORDPOS_Wynn_NNP WORDPOS_born_VBN WORDPOS_?_. NPHEAD_Stephen VPHEAD_born BG_When_was BG_was_Stephen BG_Stephen_Wynn BG_Wynn_born?

Question Classification Resutlts• Classified all questions using Li&Roth taxonomy

• Used training and testing published by Li and Roth (http://cogcomp.cs.illinois.edu/Data/QA/QC/)

• Accuracy: .808• Confusion matrix showed errors were fairly evenly distributed over

all classes

• Do we really care about all types of question?• Limited to the following ----->• Everything else was “OTHER”• Accuracy: .884• Errors were fairly evenly

distributed among selected

classes

ABBR:abbABBR:exp HUM:ind HUM:gr HUM:title HUM:indNUM:dateNUM:count

NUM:codeNUM:moneyNUM:percLOC:state LOC:cityLOC:otherLOC:countryLOC:mountENTY:color

Future question classification experiments

• which features were most helpful?• does pre-processing like tokenization, toLowercase, stemming help?

We got state-of-the art accuracy on Li and Roth test set, that leads us to believe the perf will be similar on our TRAC set for classes we are mostly interested in

We can easily re-train classifier model to incorporate more classes like ENTITY:food, ENTITY:body and ENTITY:animal to use this data during answer extraction stage

Improved Web SearchWeb results filtering based on question class

Filtering of web-passages: NUM:date, HUM:ind, HUM:title,

NUM:money, HUM:gr, ABBR:abb, NUM:perc, LOC:other, LOC:state, LOC:country, LOC:city

• if (className.equalsIgnoreCase("NUM:date"))• {• List<String> ll = stanfordNLP.extractNamedEntities(passage);//check if string contains named entities DATE_• for (String s : ll)• {• if (s.startsWith("DATE_"))• return true;

• ******• else if (className.startsWith("LOC:"))• {• List<String> ll = stanfordNLP.extractNamedEntities(passage);• for (String s : ll)• {• if (s.startsWith("LOCATION_") || s.startsWith("ORGANIZATION_")) • return true;

• etc

Answer Extraction Improvements• During document projection whenever no document is

returned we repeat the search by sending a more relaxed version of the query.• Second query is treated by Indri as a bag of words• This method leads to slight improvement of strict MRR

• Fixed encoding issue in the strings we get from Bing that led to non-sensical appearing in some ngrams

• Previously, we excluded ngrams containing words found in the question. Stop words are now excluded from that check since they are expected to be found in both the question and the answer without that meaning that they contain the answer.

TODO: Filtering of Answer Candidatesbased on question class

• Use NER results (ex. person)• Use rules (Date -> just Year or actual Date?)• Use lookup lists (animal, food, country. Profession, etc)

• Speedup • Even with limited runtime performance

ResultsSystem Lenient Strict

Baseline (D2) 0.0493692039364 0.0109649916451

D3 without classification 0.0786516699652 0.0138420106203

D3 with classification 0.085536244115 0.0165110229679

Tuned_D3 without classification 0.106098090365 0.0176606676724

Tuned_D3 with classification 0.112564337592 0.0199406397312

Issues• Performance issue when we incorporated more NLP features (tokenization, stemming, NER)

• Too little value so far from running classifier and doing some web passage filtering based on class label => need use NER data for more aggressive filtering at answer extraction stage

• Bing throwing cert exceptions at us, so we work with cached data

• A bug, so that combined answer is ranked lowers than its unigrams

Observations• Different view on data regarding some events between Internet and TRAC corpus, ex.

• How many events are part of the LPGA tour? 43• . ... 82 events on the LPGA Tour• the new tournaments means a net gain of one from 2012 to 28 official

events for the LPGA tour …

• How many people died from the massacre (tourists massacred at Luxor in 1997)? 68

• The Luxor Massacre refers to the killing of 62 people, mostly tourists• In 1997, fifty-eight innocent tourists were massacred in ...• the luxor massacre. ... massacred 62 people in 1997. Since then, Egypt's

tourism industry

Observations• TRAC corpus is frozen it time while Internet not:

• Who is the Secretary-General for political affairs? Danilo Turk• Jeffrey Feltman

• What is Tufts' current endowment“ $600 million, but that was for 2006• The endowment was valued at $1.45 billion on June 30 ...• Tufts' own endowment increased by $110 million in fiscal year 2010,

evidence…

Observations• Litowski patterns do not account for variation in

answers, ex.How many times was Moon a Pro Bowler? 9• nine

Where is the IMF headquartered? Washington• Washington D.C.• Washington DC

Who was the victim of the murder? (John William King convicted of murder) James Byrd Jr.?• James Byrd

Who did the Prince marry? Sophie Rhys.*Jones• Rhys-Jones• Sophie Rhys-Jones

Who is the senior vice president of the American Enterprise Institute? John Bolton• BOLTON John• JOHN R. BOLTON

Questions?