One of the key processes to QE is finding relevant words related to a query. Queries in this study consist of one or two sentences in most cases. A query can be considered as a document in IR. A query and retrieved documents by a search engine can be used to collect candidate QE words related to the query. LDA is a popular technique to predict a topic as a concept given a document. LDA topics are generated depending on a collection. In IR on the collection, words selected by LDA would be more appropriate for QE than words chosen from other terminologies that are created based on external sources. Of candidate words, more relevant words for QE can be identified by a classifier. An ANN classifier would be a good choice to identify relevant QE words because ANN classifiers haven shown better than other classifiers in many studies.
Three types of QE models were proposed according to:
Whether the QE model uses only an LDA model (with thresholds for topic/word probability): RQ1 Whether the QE model integrated an ANN classifier with an LDA model: RQ2
Whether the QE model integrated multiple ANN classifiers with multiple LDA models (ensemble QE models): RQ3
To rank topic words for QE, a basic word score (Sw) was calculated using Topic Probability (TP),
Word Probability (WP), and Document Rank (DR):
TP indicates how much a document is related to the topic and WP shows how much a word is related to the topic, therefore, the multiplication value of TP and WP can be used to rank words related to the document in the LDA model. DR means the rank of a document retrieved by a search engine. The first-ranked document would be more relevant to the query than the second-ranked document. LDA Topic words generated by the first-ranked document would be more relevant than topic words generated by the second- ranked document. The power value, 2, was applied to adjust a weight value by document rank (Section 4.1.1).
ANN classifiers were applied to two types of QE models for RQ2 and RQ3: 1) Word Score Weighting (WSW) and Positive Word Selection (PWS). In the WSW model, an ANN classifier was employed to give weight to the original word score (TP * WP / (DR)2). The original word score was
weighted by the probabilities for the three groups (positive/negative/neutral): original word score * (weight for positive/negative/neutral words). Weight values were given to increase the original word score of positive words and to decrease the original word score of negative and neutral words. The power value (pw), 2, which showed better performance than 1, 3, and 4, were applied.
The weight for binary ANN classifier:
- negative words: (1 – the probability to be classified into the negative word group) pw - positive words: (1 + the probability to be classified into the positive word group) pw The weight for 3-class ANN classifier (3 layers and 700 nodes per layer:
- negative words: (1 – the probability to be classified into the negative word group) pw - positive words: (1 + the probability to be classified into the positive word group) pw - neutral words: (1 – the probability to be classified into the negative word group)
In the PWS model, an ANN classifier was used to identify positive words, which were used for QE.
For RQ2, WSW and PWS models, where an LDA model and an ANN classifier were integrated, were applied to QE.
For the ensemble QE models (RQ3), one ANN classifier or multiple ANN classifiers were used to select the top k relevant words for QE, of candidate words recommended by several WSW/PWS models. The best k for QE is different according to QE models.
The overall steps for QE models using LDA models and ANN classifiers were illustrated in Figure 3.
1. Search result generation by the search engine, Terrier.
2. LDA topic word generation by LDA models: topic words were generated with different thresholds for topic probability (TP), word probability (WP), and TP*WP.
The default topic probability (TP) threshold was set up as 0.01. If the topic probability of the retrieved documents is higher than 0.01 or equal to 0.01, the topic was considered as a related topic to the document. Retrieved documents have a rank. The top1 ranked document or the top 2 ranked documents were used to generated LDA topic words. Topic words are scored by (TP*WP/ (document rank)2) and weighted by an ANN classifier. Otherwise, positive (relevant) topic words
for QE are selected by an ANN classifier. The top 7 or top 10 words were used for QE.
Topic words were filtered by specific thresholds for TP (e.g. 0.08 or 0.1), WP (e.g. 0.03), or TP*WP (e.g. 0.08). The threshold values were determined by the result in an LDA model with 1700 topics in terms of infAP, infNDCG, and the ratio of the number of positive words and negative words (Section 4.2). Threshold values generating high infAP and infNDCG scores and high ratio values were preferred. Topic words were sorted by word score. The top 10 words were added to the original query and search results by this new query were evaluated in terms of infAP and infNDCG.
3. QE by a Word Score Weighting (WSW, Figure 4) or Positive Word Selection (PWS, Figure 5) model. A WSW/PWS model consists of an LDA model and an ANN classifier. In the WSW model, an ANN classifier was used to weight topic word scores. If the word is classified into the positive word group, a weight value more than 1 is given, while a weight value smaller than 1 is given for
the negative and neutral words. For more sophisticated weighing, the probabilities of being a positive/negative/neutral word, were used rather than the same values by the classification. The top 7 or top 10 words with highest scores were used for QE. In the PWS model, all or the top 7 positive words selected by the probability of being a positive word were used for QE.
4. In the ensemble QE models, candidate words were recommended by multiple WSW/PWS models. Of the candidate words, the top k words for QE were selected by one classifier and multiple classifiers. The details were explained in the next section.
Figure 3. QE models using LDA models and ANN classifiers
Figure 5. The Positive Word Selection (PWS) model using an LDA models and an ANN classifier