Optimizing question answering systems by Accelerated Particle Swarm Optimization (APSO)

One of the most important research areas in natural language processing is Question Answering Systems (QASs). Existing search engines, with Google at the top, have many remarkable capabilities. However, there is a basic limitation; search engines do not have deduction capability which a QAS is expected to have. In this perspective, a search engine may be viewed as a semi-mechanized QAS. Upgrading a search engine such to a QAS is a task whose complexity is hard to exaggerate. To achieve success, new concepts and ideas are needed to address difficult problems which arise when knowledge has to be dealt with in an environment of imprecision, uncertainty and partial truth.  QASs are search engines that have the ability to provide a brief and accurate answer to each question in natural language for instance, the question that a search engine answers with a set of documents, a QAS answers with a paragraph, sentence or etc. In this paper, a solution is proposed to optimize the performance and speed of web-based QASs for answering English questions.  As evolutionary algorithms are suitable for issues with large search space, in this approach we have used an evolutionary algorithm to optimize QASs. In this regard, we have chosen APSO which is a simplified version of PSO. The proposed method consists of five main stages: question analysis, pre-process, retrieval, extraction and ranking. We have tried to provide a method that would be more accurate in choosing the most probable answer from the documents that have been retrieved by the standard search engine and at the same time, be faster than similar methods. In ranking process, various attributes can be extracted from the text that are used in APSO. For this purpose, in addition to selecting a sentence from the text and examining its attributes, different cut parts of the sentence are selected each time by changing the beginning and end points of the cut part. The attributes which have been used in this study are: 1. Number of unigrams similar to the question words, 2. Number of bigrams similar to the question words, 3. Number of unigrams similar to the question words in the cut part, 4. Number of bigrams similar to the question words in the cut part, 5. Number of synonyms with the question words and 6. Number of synonyms with the question words in the cut part. The fitness function is the weighted sum of these attributes. Top-1 accuracy and MRR are the most valid metrics for measuring the performance of QASs. The proposed method has achieved the accuracy (top-1 accuracy) of 0.527 with respect to the standard dataset and the MRR of it, is 0.711. Both of these results are improved compared to most similar systems. In addition, the time taken to answer the input question in the proposed method, has been significantly reduced compared to similar methods. In general, the accuracy and MRR in this paper have progressed and the system needs less time to find the answer, in comparison with existing QASs.