Haslizatul Mohamed Hanum

A Parallel Latent Semantic Indexing (LSI) Algorithm for Malay Hadith Translated Document Retrieval

Latent Semantic Indexing (LSI) is one of the well-known searching techniques which match queries to documents in information retrieval applications. LSI has been proven to improve the retrieval performance, however, as the size of documents gets larger, current implementations are not fast enough to compute the result on a standard personal computer. In this paper, we proposed a new parallel LSI algorithm on standard personal computers with multi-core processors to improve the performance of retrieving relevant documents. The proposed parallel LSI was designed to automatically run the matrix computation on LSI algorithms as parallel threads using multi-core processors. The Fork-Join technique is applied to execute the parallel programs. We used the Malay Translated Hadith of Shahih Bukhari from Jilid 1 until Jilid 4 as the test collections. The total number of documents used is 2028 of text files. The processing time during the pre-processing phase of the documents for the proposed parallel LSI is measured and compared to the sequential LSI algorithm. Our results show that processing time for pre-processing tasks using our proposed parallel LSI system is faster than sequential system. Thus, our proposed parallel LSI algorithm has improved the searching time as compared to sequential LSI algorithm.

Evaluation of prosody-related features and word frequency for Malay speeches

Changes in energy and pitch under various speaking conditions may corresponds to speaker intention and information content perceived by the speaker. In this paper, two prosody-related features which are energy and pitch of word utterances are experimented and the relationships with word frequency are investigated. Two parliamentary speeches are segmented manually into sentences and then into individual words in this experiment. Next, the maximum, minimum and mean of pitch and energy are extracted from each word to calculate word energy and pitch measure. The frequency of each word is counted to identify the relationship between word frequency with the pitch and energy measure. The study found that there is a significant relationship between word frequency and energy measure of word utterances in Malay speech but no significant relationship between word frequency and pitch measure.

Evaluation of Energy and Duration on Malay Phrase Breaks

This paper presents evaluation of energy and duration features in detection of phrase breaks. The training feature set is developed from evaluation of targeted phrase break in listening test. This cheaper and faster approach is proven useful for under-resource language such as Malay which do not have comprehensive prosodic-annotated corpus. In addition, instead of labeling the phrase break using linguistic and phonetic meaning, the listening test allow labeling of phrase break as perceived by listener. Then, word-related energy and duration features are extracted from the targeted phrase break. Evaluation of the features with RBF, MLP and logistics models reveal best detection accuracy of 84.6% which is comparable to existing context-based algorithm. This simpler approach of using energy patterns and duration features for detection of phrase break, can be used to segment lengthy spontaneous speech into smaller meaningful utterance without analysis of linguistic information. In addition, the results can be use as preliminary information for evaluation of boundary salience at the targeted boundary locations.

Identifying the Qur’anic Segment from Video Recording

This paper describes a system to identify Quran recitation (referred as Qur’anic) segment from speech video recording using the extracted acoustic signal. Identifying the Qur’anic sequence pattern from mixed-combination of speech and Qur’anic signal will contribute to more efficient segmentation of video segments. The random forest classifier algorithm is employed to classify the dynamic pattern of the extracted audio. Two feature sets which are pitch and intensity are extracted from the audio, and constructed into sequence of speech patterns which then classified as Qur’anic or non-Quranic segments. A collection of 40 segmented videos were trained and compared with the segmented videos which have been segmented manually. This project achieves classification accuracy of 57% using pitch and 85% using intensity. While using pitch feature only, 85% of the identified segments match the manually segmented collection while using intensity feature gives 95% match accordingly).

Sentence segmentation and phrase strength estimation in Malay continuous speech

Continuous speech sentences are delivered in several shorter phrasing segments which can be considered as units of information. The paper proposes a technique to improve intonational speech (IP) segmentation into the normal-strong-normal structure. The segmentation process is carried out in two phases. First, each sentence is segmented into arbitrary segments by evaluating the pause duration. Then, phrase strength is estimated using repeated pitch and intensity patterns on each phrasing segments. Phrase strength defines how strong is the pitch or intensity at that particular phrasing segment compared to the adjoining segment. This technique equates the occurrence of local maximum on pitch and intensity contour with occurrences of phrases from Malay continuous speech sentences. The result of this study suggests that the intensity contour on Malay continuous speech vary systematically with the phrase structure. This finding is useful for identifying the phrase segments that a speaker emphasized in content-based classification and retrieval of speech recordings.

Prosodic breaks on Malay speech corpus: Evaluation of pitch, intensity and duration

Prosodic phrasing is useful to segment lengthy spontaneous speech into smaller meaningful utterance without analysis of linguistic information. A simpler approach is presented to identify and classify the boundaries in prosodic phrasing using pitch and intensity patterns and pause duration on Malay speech sentences. We also propose a listening test that allows trained listener to classify the boundaries as minor or major breaks. This cheaper and faster approach is proven useful for under-resource language such as Malay which do not have comprehensive prosodic-annotated corpus. Word-related pitch, intensity and duration features are extracted from the targeted sentence and phrase breaks. A speech corpus is developed from targeted breaks of 100 speech sentences evaluated in the listening test. Instead of labeling the phrase break using linguistic and phonetic meaning, the proposed listening test allows labeling of phrase break as perceived by listener. In addition, the results can be used as preliminary information for evaluation of boundary saliency at the targeted boundary locations.

A Malay Hadith translated document retrieval using parallel Latent Semantic Indexing (LSI)

Latent Semantic Indexing (LSI) is one of the well-known searching techniques where documents are retrieved based on the content similarity or meaning of the documents. LSI is an effective method to improve the retrieval performance, however, as the size of documents gets larger; a better technique is needed to process the documents faster. In this paper, a new parallel LSI algorithm which runs on standard multi-core personal computer (PC) is proposed to improve the performance of retrieving relevant documents. The parallel LSI algorithm uses parallel threads to automatically perform the matrix computations using the Fork-Join approach. 2028 text documents extracted from four volumes of the Malay-translated book of Hadith known as Shahih Bukhari were used as the test collections. We compare the time to process LSI space between both sequential and parallel systems. The percentage of recall, precision and effectiveness for retrieving relevant document are also measured for both systems using the Information Retrieval (IR) metrics which are recall, precision, and effectiveness. The results show that the time taken to create LSI space for parallel system is faster than sequential system. Based on recall, precision and effectiveness measures, our proposed parallel LSI system is comparable to sequential LSI system.

Evaluation of Malay grammar on translation of Al-Quran sentences using Earley algorithm

Grammar is used for describing and analyzing language. As the Arabic verses from the Al-Quran are translated into Malay language, the resulting sentences have different grammar structure than the standard Malay sentences. This paper presents a simple parser to parse translation of the Quranic verses using Malay Context Free Grammar (MCFG). Earley chart parsing algorithm is embedded into the parser to evaluate the sentences and produce parse trees. As a result, additional rules to Malay grammar are proposed to successfully parse the sentences.