1% 0% 55.5% 49.0% DIFFERENT ALGORITHMS OF MACHINE LEARNING WH...

52.1% 52.0% 55.5% 49.0% different algorithms of machine learning which were: Support Vector Machine (SVM), Nearest Neighbors (NN), Nạve Limitations manually created phrases. The automatically created phrases were obtained by extracting nouns, verbs, and propositional The experiments tested only the “what” and “who” questions, phrases, while the manually created phrases were obtained by while other factoid questions such as “when” and “where” hand-correcting these automatic annotations. were not experimented. Table 14: Experimental Results – Stoyanchev et al.  Peng et al. [24], 2005: Contribution MRR Overall first answer Recall Precision of Their research presented an approach to handle the main IR with Lucene on AQUAINT corpus limitations of Ravichandran & Hovy [22]. They explored a Baseline (words disjunction hybrid approach for Chinese definitional question answering from target & question) 31.4% 62.7% 22.3% by combining deep linguistic analysis (e.g. parsing, co-Baseline reference, named-entity) and surface pattern learning in order (+ auto phrases) 33.2% 65.3% 23.6% to capture long-distance dependencies in definitional Words questions. (+ auto NEs & phrases) 31.6% 65.3% 22.0% Experimental environment and results Baseline They produced a list of questions and identified answer (+ manual phrases) 29.1% 60.9% 19.9% snippets from TDT4 data. The overall results showed that combining both pure linguistic analysis and pure pattern-based systems improved the performance of definitional (+ manual NEs & phrases) 29.4% 60.9% 20.2% IR with Yahoo API on WEB corpus questions, which proved that linguistic analysis and pattern learning were complementary to each other, and both were Baseline (words disjunction) 18.3% 57.0% 10.1% helpful for definitional questions. Cascaded (using auto phrases) 22.0% 60.4% 14.0% The pattern matching was based on simple POS tagging which captured only limited syntactic information without providing (using manual phrases) 24.1% 61.4% 15.5% any semantic information. The experimental results showed that the overall accuracy on  Lee et al. (ASQA) [25], 2005: the web was lower than that on the AQUAINT corpus. Their research proposed hybrid architecture for the NTCIR5  Kangavari et al. [21], 2008: CLQA to answer Chinese factoid questions. They presented a knowledge-based approach (InfoMap) and a machine learning The research presented a model for improving QA systems by approach (SVM) for classifying Chinese questions. They query reformulation and answer validation. The model adopted and integrated several approaches from preceding depends on previously asked questions along with the user research, such as question focus [8], coarse-fine taxonomy feedback (voting) to reformulate questions and validate [17], and SVM machine learning [19]. answers through the domain knowledge database. In the CLQA C-C task (Cross-Language Question-Answering The system worked on a closed aerologic domain for from Chinese-to-Chinese) of NTCIR campaign, their system forecasting weather information. Results showed that, from a achieved overall accuracy of 37.5% for correct answers and total of 50 asked questions, the model achieved an