Article Preview
TopIntroduction
With the rapid growth of opinionated reviews available on social media, forums, blogs, and so on, human capabilities have become insufficient to detect the sentiment orientation of such product, service, or other entity. To analyze these data, sentiment analysis systems are required. Sentiment analysis is the computational task for analyzing opinions and feelings expressed in texts. An opinion, according to (Liu, 2010), is made up of the following elements: the opinion object and its aspects, sentiment polarities, the opinion holder who expresses the opinion, and the time of publication of the opinion. Depending on the text granularity, various levels for opinion classification can be considered; the most commonly cited being the document, the sentence, and the aspect levels. The sentiment analysis at the document and sentence levels aims to identify the overall sentiment orientation of each document and sentence, respectively. Indeed, sentiment analysis at these levels does not reflect any sentiment information about the object's aspects (also known as features, entities, or targets). For this reason, the Aspect-Based Sentiment Analysis (ABSA) is introduced. The main focus of ABSA is to identify the targets and their sentiment polarities. As exemplified in the following review “The food is very delicious and the wait staff is friendly”, the terms “food” and “wait staff” are identified as aspects of the restaurant object, while “delicious” and “friendly” are opinion terms that express a positive polarity for both aspects. As a result, this task is divided into two research contexts that have been extensively discussed in the literature. The first one is called Opinion Aspect Extraction (OAE). The Aspect Sentiment Classification (ASC) is the second context. OAE and ASC are two distinct tasks. ASC is a traditional classification task, whereas OAE is an extraction task that is typically studied as a sequence labeling problem. These tasks are addressed with the early works as pipeline methods requiring two different training models. Recent research, on the other hand, has focused on the following challenge: “How to integrate OAE and ASC into a single model for improving ABSA performance?” It is not a trivial task because the goals of the two tasks are so dissimilar. These works investigated the ABSA as a sequence labeling problem (X. Li et al., 2019; Luo et al., 2019; Akhtar et al., 2020; N. Li et al., 2020, Z. Li et al, 2021, Yan et al., 2021) or as a span extraction problem (Y. Zhou et al., 2019; Hu et al., 2019; Lv et al., 2021) in order to construct joint or collapsed models (More details about these methods and models are provided in the related works section). Nonetheless, the sequence labeling methods suffer from two potential drawbacks, which are: (i) the aspect extraction task is essentially built at the token level, which is likely to result in imprecise predictions, particularly with the compound aspects. As regard the aspect “wait staff” of the aforementioned example, if the sentence is treated at the token level, the word “wait” may be considered as a verb rather than as an aspect term. (ii) Furthermore, the token-level-based aspect sentiment classification can lead to wrong predictions within multi-word aspects. If the model fails to identify the correct sentiment polarity from the start of a compound aspect, the error will be disseminated towards the other aspect terms. These issues have been addressed by the span-based approaches. Indeed, such methods typically outperform the sequence tagging methods because the aspect extraction and the sentiment classification tasks are effectively handled at the span level. To avoid the sentiment inconsistency problem, the authors in (Hu et al., 2019; Lv et al., 2021) used the attention mechanism for aggregating the contextual information of each span. The combination of span attention and word attention, on the other hand, may generate noise and reduce prediction accuracy. In this work, the authors propose a new collapsed model named MCLAT-ALSC (Multichannel Convolution & LSTM-based Aspect Tagging- Attention & LSTM-based Sentiment Classification) to avoid the aforementioned problems of both sequence labeling and span methods. To address the first issue of the sequence labeling methods, the n-gram model was used to manage the data at various levels (i.e., token, bigram, n-gram). Indeed, the convolution layer with its kernel size supports these types of models. Following such perspective, a multichannel convolution layer-based model was introduced for the enhancement of the performance of aspect extraction task. Nevertheless, the use of different kernel size values is likely to produce multiple features. Consequently, an optimization method based on aspect position information was developed in order to reduce the number of features and select the most appropriate ones with aspect representations. To sort out the sentiment consistency issue, an attention mechanism based on the Residual Block (RB) architecture and Position Information (PI) was developed to obtain the most significant opinion features for a specific aspect. The proposed model's performance has been demonstrated to outperform ABSA methods found in the literature.