The article presents the experiments carried out as part of the participation in the main task of QA4MRE@CLEF 2012. In the developed system, we first combine the question and each answer option to form the Hypothesis (H). Stop words are removed from each H and query words are identified to retrieve the most relevant sentences from the associated document using Lucene. Relevant sentences are retrieved from the associated document based on the TF-IDF of the matching query words along with n-gram overlap of the sentence with the H. Each retrieved sentence defines the Text T. Each T-H pair is assigned a ranking score that works on textual entailment principle. A validate weight is automatically assigned to each answer options based on their ranking. A parallel procedure also generates the possible answer patterns from given questions and answer options. Each sentence in the associated document is assigned an inference score with respect to each answer pattern. Evaluated inference score for each answer option is multiplied by the validate weight based on their ranking. The answer option that receives the highest selection score is identified as the most relevant option and selected as the answer to the given question.
The main objective of QA4MRE [
Answer Validation (AV) is the task of deciding for given a question and an answer from a QA system, whether
the answer is correct or not and it was defined as a problem of RTE in order to promote a deeper analysis in
Question Answering [
Section 2 describes the corpus statistics. Section 3 describes the system architecture. The experiments carried out on test data sets are discussed in Section 4 along with the results. The conclusions are drawn in Section 5. As in the previous campaign, the task focuses on the reading of single documents and the identification of the answers to a set of questions about information that is stated or implied in the text. Questions are in the form of multiple choices, each having five options, and only one correct answer. The detection of correct answers is specifically designed to require various kinds of inference and the consideration of previously acquired background knowledge from reference document collections provided by the organization. Although the additional knowledge obtained through the background collection may be used to assist with answering the questions, the principal answer is to be found among the facts contained in the test documents given.
The 2012 test set will be composed of 4 topics, namely “Aids”, “Climate change” and “Music and Society” – the same topics adopted last year – plus the addition of a new topic, i.e. “Alzheimer”. Each topic will include 4 reading tests. Each reading test will consist of one single document, with 10 questions and a set of five choices per question. So, there will be in total: • • - 16 test documents (4 documents for each of the four topics) - 160 questions (10 questions for each document) with • - 800 choices/options (5 for each question) Participating systems will be required to answer these 160 questions by choosing in each case one answer from the five alternatives. There will always be one and only one correct option. Systems will also have the chance to leave some questions unanswered if they are not confident about the correctness of their response.
Topics, documents and questions were made available in English, German, Italian, Romanian, Spanish and two new languages added this year -Arabic and Bulgarian. We worked only with English language data. The Background Collections (one for each topic) are comparable (but not identical) topic-related collections created in all the different languages. 3
The architecture of machine reading system is described in Figure 1. Proposed architecture is made up of four main modules along with knowledgebase. Each of these modules is now being described in subsequent subsections.
Document processing module consists of three sub-modules: XML Parser, Named Entity (NE) Identification and Anaphora Resolution.
The given XML corpus has been parsed using XML parser. The XML parser extracts the document and associated questions. After parsing, the documents and the associated questions are extracted from the given XML documents and stored in the system.
For each question, system must identify the correct answer among the proposed alternative answer options. Each generated answer pattern corresponding to a question is compared with each sentence in the document to assign an inference score. The score assignment module requires that the named entities in each sentence and in each answer pattern are identified. The CRF-based Stanford Named Entity Tagger1 (NE Tagger) has been used to identify and
It has been observed that resolving the anaphors in the sentences in the documents improves the inference score of the sentence with respect to each associated answer pattern. The following basic anaphora resolution techniques have been applied in the present task. (a) Each first person personal pronoun in the set PNI = {‘I’, ‘me’, ‘my’, ‘myself’} generally refers the author of the document as describer. For example, the anaphors in the following sentence can be resolved in the following steps: I am going to share with you the story as to how I have become an HIV/AIDS campaigner.
Step1: < PNI > am going to share with you the story as to how < PNI > have become an HIV/AIDS campaigner. Step2: < PNI= “author” value= “Annie Lennox”> am going to share with you the story as to how < PNI= “author” value= “Annie Lennox” > have become an HIV/AIDS campaigner.
Step3: <NE=Person value= Annie Lennox> am going to share with you the story as to how <NE=Person value=
Annie Lennox> have become an HIV/AIDS campaigner.
In direct speech sentences PNI refers to the first named entity (speaker) of that sentence. For example, the anaphors in the following sentence can be resolved in the following steps: Frankie said, "I am number 22 in line, and I can see the needle coming down towards me, and there is blood all over the place.
Step 1: <NE=Person value=Frankie> said, " < PNI > am number 22 in line, and < PNI > can see the needle coming down towards < PNI >, and there is blood all over the place.
Step 2: <NE=Person value=Frankie> said, "< PNI= “NE” value= “Frankie” >” am number 22 in line, and < PNI= “NE” value= “Frankie” > can see the needle coming down towards < PNI= “NE” value= “Frankie” >, and there is blood all over the place.
Step 3: <NE= “Person” value= “Frankie”> said, " <NE= “Person” value= “Frankie”> am number 22 in line, and <NE=Person value=Frankie> can see the needle coming down towards <NE=Person value=Frankie>, and there is blood all over the place. (b) Each second person personal pronoun in the set PNHe/She = {‘he’, ‘his’, ‘him’, ‘her’, ‘she’} generally refers the last NE of the previous sentence. For example, the anaphors in the following sentence can be resolved in the following steps: I was invited to take part in the launch of Nelson Mandela's 46664 Foundation. That is his HIV/AIDS foundation. Step 1: < PNI > was invited to take part in the launch of <NE= “Person” value= “Nelson Mandela”>'s 46664
Foundation. That is <PNHe/She > HIV/AIDS foundation.
Step 2: < PNI= “author” value= “Annie Lennox”>was invited to take part in the launch of <NE=Person value= “Nelson Mandela”>'s 46664 Foundation. That is <PNHe/She = “PREV_NE” value= “Nelson Mandela”> HIV/AIDS foundation.
Step 3: < PN = “author” value= “Annie Lennox”>was invited to take part in the launch of <NE= “Person”
I value= “Nelson Mandela”>'s 46664 Foundation. That is <NE= “Person” value= “Nelson Mandela”> HIV/AIDS foundation.
But, in indirect speech sentences, PNHe/She refers to the first named entity (speaker) of that sentence. For example, the anaphors in the following sentence can be resolved in the following steps: Alexander Graham Bell famously said that on his first successful telephone Call.
Step 1: <NE= “Person” value= “Alexander Graham Bell”> famously said that on < PNHe/She > first successful telephone Call.
Step 2: <NE= “Person” value= “Alexander Graham Bell”> famously said that on < PNHe/She = “SEN_NE” value= “Alexander Graham Bell”> first successful telephone Call.
Step 3: < NE = “Person” value = “Alexander Graham Bell” > famously said that on <NE=“Person” value=“Alexander Graham Bell”> first successful telephone Call.
At first we convert each question into an affirmative sentence that denotes the answer pattern and place the </answer> template in place of the appropriate answer. The pattern generation module is rule based.
For example, let us consider the question id 7 in doc id 2 of the QA4MRE train set, Question: Where is the U.S. nuclear waste repository located? The generated pattern is The U.S. nuclear waste repository is located </answer>.
After Pattern generation the </answer> template is replaced by each answer option string forming the generated Hypothesis. The generated hypothesis is termed as the query. For example, for question id 7 (QA4MRE Train set), the following hypotheses (or queries) are generated for each of the answer options: H_1: The U.S. nuclear waste repository is located at Oklo.
H_2: The U.S. nuclear waste repository is located in Morsleben.
H_3: The U.S. nuclear waste repository is located in New Mexico.
H_4: The U.S. nuclear waste repository is located in a suitable geological formation.
H_5: The U.S. nuclear waste repository is located in the U.S. State of Nevada.
The corpus is in XML format. All the XML test data has been parsed before indexing using our XML Parser. The XML Parser extracts the sentences from the document. After parsing the documents, they are indexed using Lucene, an open source full text search tool.
After indexing has been done, the queries have to be processed to retrieve relevant sentences from the associated documents. Each answer pattern or query is processed to identify the query words for submission to Lucene. Each hypothesis has been submitted to Lucene after removing stop words (using the stop word list2 ). The remaining words are identified as the query words. Query words may appear in inflected forms in the question. For English, standard Porter Stemming algorithm3 has been used to stem the query words. After searching using Lucene, a set of sentences in ranked order are retrieved.
First of all, all query words are fired with AND operator. If at least one sentence is retrieved using the query with AND operator then the query is removed from the query list and need not be searched again. The rest of the queries are fired again with OR operator. OR searching retrieves at least one sentence for each query. Now, the top ranked relevant ten sentences for each query are considered for further processing In case of AND search only the top ranked sentence is considered. Sentence retrieval is the most crucial part of this system. We take only the top ranked relevant sentences assuming that these are the most relevant sentences in the associated document for the question from which the query has been generated.
Each retrieved sentence is considered as the Text (T) and is paired with each generated hypothesis (H). Each T-H
pair identified for each answer option corresponding to a question is now assigned a score based on the NER
module, Textual Entailment module, Chunking module, Syntactic Similarity module and Question Type module.
3.2.3.1 NER Module
It is based on the detection and matching of Named Entities (NEs) [
If the value of NE_Match is 1, i.e., 100% of the NEs in the hypothesis match in the text, then the T-H pair is considered as an entailment. The T-H pair is assigned the value “1”, otherwise, the pair is assigned the value “0”.
This TE module [
The original questions are pre-processed using Stanford Dependency parser [
The question sentences are pre-processed using Stanford dependency parser. The words along with their part of speech (POS) information are passed through a Conditional Random Field (CRF) based chunker [11] to extract phrase level chunks of the questions. A rule-based module is developed to identify the chunk boundaries. The question-retrieved text pairs that achieve the maximum weight are identified and the corresponding answers are tagged as “1”. The question-retrieved text pair that receives a zero weight is tagged as “0”.
This module is based on the Stanford dependency parser [
After dependency relations are identified for both the retrieved sentence and the hypothesis in each pair, the hypothesis relations are compared with the retrieved text relations. The different features that are compared are noted below. In all the comparisons, a matching score of 1 is considered when the complete dependency relations along with all of its arguments match in both the retrieved sentence and the hypothesis. In case of a partial match for a dependency relation, a matching score of 0.5 is assumed. a. Subject-Verb Comparison. The system compares hypothesis subject and verb with retrieved sentence subject and verb that are identified through the nsubj and nsubjpass dependency relations. A matching score of 1 is assigned in case of a complete match. Otherwise, the system considers the following matching process. b. WordNet Based Subject-Verb Comparison. If the corresponding hypothesis and sentence subjects do match in the subject-verb comparison, but the verbs do not match, then the WordNet distance between the hypothesis and the sentence is compared. If the value of the WordNet distance is less than 0.5, indicating a closeness of the corresponding verbs, then a match is considered and a matching score of 0.5 is assigned. Otherwise, the subjectsubject comparison process is applied. c. Subject-Subject Comparison. The system compares hypothesis subject with sentence subject. If a match is found, a score of 0.5 is assigned to the match. d. Object-Verb Comparison. The system compares hypothesis object and verb with retrieved sentence object and verb that are identified through dobj dependency relation. In case of a match, a matching score of 0.5 is assigned. e. WordNet Based Object-Verb Comparison. The system compares hypothesis object with text object. If a match is found then the verb corresponding to the hypothesis object with retrieved sentence object's verb is compared. If the two verbs do not match then the WordNet distance between the two verbs is calculated. If the value of WordNet distance is below 0.5 then a matching score of 0.5 is assigned. f. Cross Subject-Object Comparison. The system compares hypothesis subject and verb with retrieved sentence object and verb or hypothesis object and verb with retrieved sentence subject and verb. In case of a match, a matching sc ore of 0.5 is assigned. g. Number Comparison. The system compares numbers along with units in the hypothesis with similar numbers along with units in the retrieved sentence. Units are first compared and if they match then the corresponding numbers are compared. In case of a match, a matching score of 1 is assigned. h. Noun Comparison. The system compares hypothesis noun words with retrieved sentence noun words that are identified through nn dependency relation. In case of a match, a matching score of 1 is assigned. i. Prepositional Phrase Comparison. The system compares the prepositional dependency relations in the hypothesis with the corresponding relations in the retrieved sentence and then checks for the noun words that are arguments of the relation. In case of a match, a matching score of 1 is assigned. j. Determiner Comparison. The system compares the determiner in the hypothesis and in the retrieved sentence that are identified through det relation. In case of a match, a matching score of 1 is assigned. k. Other relation Comparison. Besides the above relations that are compared, all other remaining relations are compared verbatim in the hypothesis and in the retrieved sentence. In case of a match, a matching score of 1 is assigned.
API for WordNet Searching RiWordnet4 provides Java applications with the ability to retrieve data from the WordNet database.
Each of the matches through the above comparisons is assigned some weight 4 http://www.rednoise.org/rita/wordnet/documentation/index.htm
In this module, we have got the weight from Named Entity Recognition (NER) Module (Section 3.6), Textual Entailment (TE) Module (Section 3.7), Question Type Analysis Module (Section 3.8), Chunk Boundary (Section 3.9) and Syntactic Similarity Module (Section 3.10).
For each question has five hypothesis (H). Hypothesis are ranked by using NER Module, Textual Entailment Module, Chunking Module, Syntactic Similarity Module, Question type analysis Module.
Rank-2 Rank-3 Rank-4
Each sentence in the associated document is assigned an inference score with respect to each generated answer pattern.
Each question has a number of answer options and the task is to identify the best answer to the question given an associated document. Each question in the system is identified as the (question, document) pair represented as {qi, d_id} where i=1…10. There are 10 questions corresponding to each document. The “WH” word in the question is substituted by the given answer option to generate the answer pattern. The set of WH words include WHP ={‘Who’, ‘What’, ‘Where’, ‘Name’, ‘Which’, ‘Whom’, ‘Why’}. Each answer pattern is represented in the system as {d_id, q_idi, a_idj}, where, d_id=document id, q_idi= i th query, where i=1…10, a_idj= j th answer option, where j=1…5. Let us consider an example.
Question: Who is the founder of the SING campaign? Answer Option: Nelson Mandela WHP: who Generated Answer Pattern: Nelson Mandela is the founder of the SING campaign Each answer pattern is stored in the system as the pair (PAT, KL) where, PAT= Probable Answer Text, which is the generated answer pattern and KL= Keyword List, is a list of words after removing the stop words.
For example, the above generated answer pattern is stored as PAT= “Nelson Mandela is the founder of the SING campaign” KL= “Nelson”, “Mandela”, “founder”, “SIGN”, “campaign”.
This module takes query frame as input and returns score as output. The algorithm InferenceScore describes the scoring procedure. keywordmatched = 0 // count no of matched keyword
Algorithm InferenceScore (sentence, PAT, KL) Step 1: [Initialization]
score = 0 Step 2: [Check whether PAT matches in a sentence]
If PAT matches in a sentence then
Score = 1 goto step 5 Step 3: [Check each keyword in KL]
For each keyword in KL
Step 4: [Check whether all the keywords have matched] If (keywordmatched = = total keywords – 1) then
Score =1
End
Now, for each given answer option a score is calculated and the answer option with highest score is taken as correct answer for the given query. The algorithm SelectAnswerOption describes the option selection procedure.
We have prepared some domain knowledgebase for this task. There are four topics in this task: AIDS, Climate Change, Music and society and Alzheimer. So we prepared Named Entity (NE) list, Abbreviation list and MultiWord list (MWE) of these topics except the Climate Change. There are two topics from medical domain: AIDS and Alzheimer. There are lots of medical terms and abbreviation, which can not be identified with a general domain named entity recognizer (NER). So we have prepared these domain based lists manually.
AQi=maximum evaluated score for {S1,S2,…..Sn};
Where AQi is the score of ith option End For
End For END Step 4: [ Applying Validate Factor(Vf)]
For each answer option AQj€ AQ
AQj =InferenceScore(AQj) x Vf Step 5: [Select the answer option]
correct_option= index of maximum AQ={ AQ1, AQ2, AQ3, AQ4, AQ5 } The main measure used in this evaluation campaign is c@1, which is defined in equation 1. (1)
where, nR: the number of correctly answered questions, nU: number of unanswered questions n: the total number of questions Evaluation at question-answering level: - Number of questions ANSWERED: 124 - Number of questions UNANSWERED: 36 - Number of questions ANSWERED with RIGHT candidate answer: 85 - Number of questions ANSWERED with WRONG candidate answer: 39 - Number of questions UNANSWERED with RIGHT candidate answer: 0 - Number of questions UNANSWERED with WRONG candidate answer: 0 - Number of questions UNANSWERED with EMPTY candidate: 36 Accuracy (answered with judgment=correct) calculated over all questions: Overall accuracy = 85/160 = 0.53 Proportion of answers correctly discarded: 0/36 = 0.00
Evaluation at reading-test level: c@1 of all 16 reading tests: Median: 0.66 ; Average: 0.62; Standard Deviation: 0.22
The question answering system has been developed as part of the participation in the QA4MRE track as part of the CLEF 2012 evaluation campaign. The overall system has been evaluated using the evaluation metrics provided as part of the QA4MRE 2012 track. It has been observed from evaluation results that our proposed model works very well on the topics- “Aids”, “Music and Society” and “Alzheimer”. And the system performance decrease to handle “Climate change” documents and questions. As we have only prepared the domain base knwoledgebase only for the medicine domain and a tillte amount of knwoledgebase for Music and Socity. But we could not used any daomain knwoledge for Climate change. So it evaluation result is very poor. Hence it’s proved that domain knwoledgebase has a stong effect on each of our system. But, the overall evaluation results are satisfactory. Future works will be motivated towards improving the performance of the system.
Acknowledgements. We acknowledge the support of the IFCPAR funded Indo-French project “An Advanced Platform for Question Answering Systems” and the DIT, Government of India funded project “Development of Cross Lingual Information Access (CLIA) System Phase II”.