Aspect Extraction Performance With Common Pattern of Dependency Relation in Multi Aspect Sentiment Analysis

Editor's Notes

1. In product review people usually comment on multiple aspect and give different sentiment on various aspects of that product. Issues in product review Specifically, in a review it might consists of four issues: It is a challenge to deal with review sentence that consists of multiple aspects with various polarities expressed to multiple sentiments. Therefore, it is essential to identify and extract each aspect and it specific associated sentiment word correctly. This figure gives an example of review sentence that consist of multiple aspects and multiple sentiments. The first aspect is ‘display’ which is associated with two sentiment words ‘best’ and ‘long time’. Both sentiment words expressed the positive sentiment. Same as for second aspect ‘battery life’ also associated with two sentiment words ‘long’ and ‘convenient’ and expressed the positive sentiment.
2. This issue contributes to the lower precision and recall.
3. Usually the nouns or noun phrases and adjective resulting from POS tag were represented as aspect and sentiment word respectively. In some cases, verbs and adverb also could represent as sentiment word.
4. In Stanford dependencies (SD), it is represented as triplets: name of the relation, governor and dependent. The figure shows that dependency relation nsubj(best-8, display-2), nsubj(long-23, life-20) and compound(life-20, battery-19) will be helpful in identifying aspects. Meanwhile the dependency relation amod(time-16 , long-15 ), conj( long-23, convenient-26) will be helpful in identifying sentiment words.
5. (1) There can be many different types of dependency relations in the sentences, however not all of them are helpful and contribute in identifying aspect and sentiment word. Therefore, in the first step, an experiment on these fifteen TDR (from previous study) has been performed on the SemEval 2014 dataset to select the relevant TDR that could identify the aspect and sentiment word. (2) The first pattern is, governor as adjective and dependent as nouns that represent sentiment word and aspect respectively. Therefore, the POS tag pattern is nsubj (JJ, NN). The second pattern is governor as verb and dependent as nouns that represent sentiment word and aspect respectively. The POS tag pattern is nsubj (VB, NN). TDR ID is given as nsubj1 and nsubj2 to differentiate two POS tag pattern of ‘nsubj’ relation.
6. For example, if dependencies parsing of a review sentence contain ‘nsubj’ relation, and the POS tag of governor and dependent is “JJ” and “NN” respectively, therefore the dependent is an aspect. This example describes the extraction rule of TDR ID nsubj1 as can be seen in third column of Table II. The extraction rule is applied to pre-extract the candidate aspect and sentiment word.
7. The aspect extraction performance between different TDR shows that ‘compound’ relation achieved the best performance for both domain. 41.05% of good precision, 29.70% of highest recall and 34.46% of highest F1-score value for laptop domain. In the case of restaurant domain, ‘compound’ relation achieved good precision of 52.92%, highest recall of 29.02% and highest F1-score value of 37.49%. For precision, ‘nsubj1’ obtained highest precision among the others TDR with 58.12% and 85.99% of precision for laptop and restaurant domain respectively. This shows that the POS tag pattern of ‘nsubj1’ able to extract aspect correctly with the small number of false aspect extracted. The ‘conj’ relation also presents the good performance. As can be seen in Table V, the TDR ID ‘conjA3’ and ‘conjA1’ achieved precision of 46.67% and 46.43% respectively. Meanwhile ‘conjA1’ from Table VI achieved 74.20% of precision. This indicates that POS tag pattern of conjA3 and conjA1 contribute to the high performance of multi aspects extraction.
8. Aspect extraction performance between different TDR shows that ‘compound’ relation achieved the best performance for both domain. The ‘nsubj1’ obtained highest precision among the others TDR with 58.12% and 85.99% of precision for laptop and restaurant domain respectively.