QGSTEC System Description JUQGG: A Rule based approach

1. QGSTEC System DescriptionJUQGG: A Rule basedapproach Santanu Pal, Tapabrata Mondal, Partha Pakray, Dipankar Das and Sivaji Bandyopadhyay Department of Computer Science and Engineering Jadavpur University Kolkata-700032, India

2. Problem Description • Task provides participants with a single sentence and a specific target question type (E.g. WHO? WHY? HOW? WHEN? etc).  • Participants are asked to generate 2 questions from the single sentence of the question type specified. QGSTEC, 2010

3. System Framework 1.Combinator Module 2. Chunker and Clause Detector 3. Question Recognizer (QR) Module 4. Question Generator (QG) Module QGSTEC, 2010

4. Component based Conceptual System Diagram QGSTEC, 2010

5. 1. Combinator Module • Input sentences are passed through the open source tools such as Semantic Role Labeler (SRL) and Stanford Dependency Parser • Keyword lists of causal words, negation words and of ordinal and cardinal numbers. • Identifies the person, temporal and location information from the semantic role labels. • Integrates all the information and forwards it to the Question Recognizer (QR) module. QGSTEC, 2010

6. 2. Chunker and Clause Detector (1) • Chunking done using a Conditional Random Field (CRF) based chunker • Root forms of nouns and verbs are identified using a Morphological Analyzer available with English WordNet • Punctuation marks, Discourse markers identified through mark () type dependency relations • Causal Words (as, because)used for clause detection • Dependency Relations connected directly with each verb chunk are used in clause detection QGSTEC, 2010

7. 2. Chunker and Clause Detector (2) • Each verb chunk and the associated chunks whose head is directly linked with the verb chunk in any dependency relation identifies a clause • Clause detection process helps to generate multiple questions of different types from a single sentence QGSTEC, 2010

8. 3.Question Recognizer (QR) Module (1) • Simple Rules • who type Question - If SRL labels any phrase as first argument (tagged as ARG0) and if any word of that phrase is tagged as NNP and present in nsubj() dependency relation  the phrase is considered as a possible question cue phrase • what type Questions - If a phrase labeled as ARG0 provided that its respective clause contains both ARG0 and ARG1 labeled by SRL  the phrase is considered as a possible question cue phrase QGSTEC, 2010

9. 3.Question Recognizer (QR) Module (2) • when type Question - If SRL labels any phrase as ARGM-TMP or TEMPORAL and the chunk of the phrase contains any word with POS category ‘CD’ and any word of prep_in () or prep_since () dependency relations  the phrase is considered as a possible question cue phrase • where type Questions - If SRL labels any phrase as ARGM-LOC or LOCATION and the chunk of the phrase contains any word of POS category NN/NNP and if any word of that phrase is associated in prep_of () dependency relation the phrase is marked as question cue phrase QGSTEC, 2010

10. 3.Question Recognizer (QR) Module (3) • which type Questions - If any extracted chunk contains multiple noun words that are also present in nn () dependency relation the chunk is considered as question cue phrase. • Dependency relations of prep () type are considered for generating special types of which questions (e.g For which, In which etc.) QGSTEC, 2010

11. 3.Question Recognizer (QR) Module (4) • A slightly different approach considered for generating “how many”, “why” and “yes/no” questions • Keyword lists are used to identify the possible question cue phrases of the three question types • how type Questions - If any phrase contains any word with POS category CD and the phrase is not labeled as ARGM-TMP or TEMPORAL by SRL the phrase is considered as the question cue phrase - A separate knowledgebase is prepared for generating different forms of how questions(e.g. how many, how much, how old, what percentage of etc.) (Analyses made on development set). QGSTEC, 2010

12. 3.Question Recognizer (QR) Module (5) • why type Questions - If any phrase contains key words like as, because that signifies the presence of causal item in a sentence the phrase is considered as the question cue phrase • Yes/No type Questions - If any phrase contains negation word (e.g. not, no) as well as any auxiliary verb  the phrase is considered as the question cue phrase QGSTEC, 2010

13. Example • Text: Nash began work on the designs in 1815, and the Pavilion was completed in 1823. • Q-Type:Who, When • SRL output with [cue phrase]:[ARG0 Nash] [TARGET began] [ARG1 work on the designs] [ARGM-TMP in 1815] and [ARG1 the Pavilion] was [TARGET completed] [ARGM-TMP in 1823] • Generated Questions: 1. Whobegan work on the designs in 1815? 2. WhenNash began work on the designs? 3. When was Pavilion completed in? QGSTEC, 2010

14. 4.Question Generator (QG) Module • Replaces cue phrase in the clause by the specified question word • Reordering module generates possible questions from the respective clause by reordering the chunks in the clause based on rule based grammatical knowledge QGSTEC, 2010

15. Evaluation and Observations • Evaluation on 81 development sentences • Preliminary human evaluation gives satisfactory results • Clause detection using semantic role labels - Filter the adjuncts - Frame questions mostly from argument portions of the input sentences • Handling of Prepositional chunk attachment problem to improve the quality of generated questions QGSTEC, 2010

16. Thank you …. • Questions ? ….. QGSTEC, 2010