Home

In recent text generation research, controlled text generation is in progress. A sentence generation considering simple positive/negative was proposed based on a pretrained language model [3]. The sentence generation method considered aspects beyond sentence generation through simple sentiment analysis. [4]. However, the segmentation method cannot be easily applied to various review expressions when sentences are segmented based on the aspect. Moreover, by adding target content and target sentiment (e.g., positive or negative) to the input value to control the sentence, the output text was generated by inputting detailed content rather than simple sentence generation. [5]

In addition, a method of generating a rather long sentence by using at least five words as the input was proposed. Keywords are numbered and created sequentially [6]. In the case of sentence generation in this way, it depends on the input keyword without considering the intensity of the sentence. Therefore, in this paper, a detailed segmentation method was derived to generate sentences with intensity.

Aspect-based Sentiment Analysis (ABSA) method [7], which analyzes the sensibility of attributes in detail, can obtain user insights by identifying user sentiment and contributing factors for each significant aspect and aspect in the user's review. In addition, by applying the ABSA to the review of e-commerce, additional aspect-specific vocabulary was constructed [8]. When an aspect (service, location, food, etc.) appears in each sentence and words expressing sentiment appear, they are paired to judge negative and positive. Therefore, sentiment features were extracted for each aspect. However, in recent online reviews, more diverse expressions appear than before, and it is not appropriate to divide them into positive and negative. Therefore, in this paper, the intensity of the sentence is finely adjusted to 1-5 points, and various expressions are created when the sentence is generated for each intensity.

In this study, we crawled Korean review data of Yogiyo [9], which provides Delivery Order Service. The total crawled reviews amount to 823,234, especially for the chicken category. To adjust the intensity based on the existing score, reviews and user scores were also collected.

We refined the collected data to remove existing noises in review texts. First, since there are many sentences with incorrect spacing in the existing review data, spacing was newly performed using the PyKoSpacing library (Python package for automatic Korean word spacing) [10]. Additionally, the Pyhanspell (Hangul Spelling Library) library [11] was used for the spelling check. Second, we've selected reviews of between 30 and 200 characters, including spaces. Reviews of less than 30 characters have many ambiguous words and duplicated words, so we got rid of them from our dataset. In addition, we also deleted reviews of more than 200 characters because most of them were not grammatically correct and had unnatural flows. Third, we removed emoticons and special symbols that are not helpful for the review generation. After those text pre-processing steps, we finally got 233,003 reviews as training raw data.

If we examine the actual reviews in detail, many reviews have negative sentiments, even though the sentences were rated 4 or 5 points, which are close to solid positive. Among the 823,234 reviews collected, 5-point scores occupy 76% of them. Meanwhile, 1-point and 2-point scores are less than 6%. In other words, most online reviews have overwhelmingly positive ratings. Examples such as Table 1 show that some negative expressions are expressed in a written review, even for positive assessments. Therefore, we newly label the scores in two ways to solve the data imbalance of the collected data.

3.4.2 Labeling new scores

In this method, the initial intensity was set to 3points without using the existing score. Based on the review sentiment dictionary, we used a method of adding 1point when positive words appeared and subtracting 1point when negative words appeared according to the number of words by analyzing emotions. The review used in this study was not effective when analyzed using the existing Korean sentiment dictionary because there are many new words and unique words of aspect. So, we constructed the emotional words used in the review separately. The total number of words is 940, with 479 positive words and 461 negative words. Therefore, a result with a score of 5 or more was treated as 5, and a result less than 0 was treated as 1, making a total score between 1 and 5. Finally, when the methods of Section 3.3.2 and 3.4.1 were applied, score distributions were drastically changed from imbalance ones (5 points: 161,884 (76%), 4 points: 27,598 (13%), 3 points: 10,817 (5%), 2 points: 4,394 (2%), and 1 point: 8,620 (4%)) to somewhat balanced ones (5 points: 191,199 (36.45%), 4 points: 172,347 (32.85%), 3 points: 100.925 (19.24%), 2 points: 40,514 (7.72%), and 1 point: 19,611 (3.74%)) were constructed.

Fig. 2. 
Algorithm

To examine the feasibility of our segmentation method and the auto labeling method, we employed two popular text generation methods (i.e., GPT and BART). Moreover, we evaluate Perplexity, BLEU-4 [15], ROUGE score [16], BERT-SCORE [17] together with human judgments to explore the quality of generated review texts.

For the review generation experiments, a total of 524,596 data (after segmentation) were used as training data, and KoGPT-2 ver1, KoGPT-2 ver2, and KoBART models were chosen as automatic generation techniques. The entire collected data went through the text pre-processing to precisely purify the data as in Section 3.2. Since the collected reviews were not official documents, sentences were composed only of emoticons, neologisms, and onomatopoeia. Moreover, words that are difficult to understand and sentences with many overlapping meanings have been removed. To find the most appropriate combination of segmentation and scoring schemes that are introduced in Section 3.2 and 3.3, we've tested the following five methods.

• No Segmentation & Existing Scores (NE): It was trained in KoGPT-2 ver1, ver2, and KoBART models without data refinement process.
• Simple-Segmentation & labeling based on Existing Scores (SSLE): Data using simple-segmentation and labeling techniques using existing scores
• Simple-Segmentation & Labeling New Score (SSLN): Combination using simple segmentation and labeling new score.
• Segmentation Considering Sentence Intensity & Labeling based on Existing scores (SCILE): Data using Segmentation Techniques Considering Intensity and Labeling Techniques Using Existing Scores
• Segmentation Considering Sentence Intensity & Labeling New score (SCILN): Segmentation considering sentence Intensity and Data with labeling New scores without using existing scores

When trained with the raw data without the pre-processing, both GPT-2 and BART-based review generation models produce sentences that do not conform to intensity and sentences that do not conform to grammars. Meanwhile, when we applied the text pre-processing, our proposed segmentation and labeling techniques, GPT-2 based models succeeded in generating sentences appropriate to the intensity with varying lengths.

The sentence is not natural when entering Negative Keyword because the number of negative data is small before pre-processing. SSLE and SSNL produce relatively short sentences in length and have a natural flow of sentences, but some created sentences are not related to the input keyword. It also creates sentences that are not appropriate for intensity. SCILE adds longer sentences and more words than previous methods. In particular, SSILN produces a natural sentence compared to other methods, and the expression of the sentence is abundant

BART requires a data set with an encoder-decoder structure, so the data is manually constructed, and the number of data is 7823. KoBART applied the simple-segmentation method and auto-labeling method. As mentioned earlier, BART is an input-output data set, so if segmentation is applied to one output (review), sentences that do not match the input (keyword) are derived. Due to these problems, it was challenging to construct the data set and apply the proposed five methods. Therefore, the datasets are relatively small, and only the Simple-Segmentation method and Auto-Labeling are applied. Unlike GPT, BART produced sentences similar to GPT or relatively short and produced slightly poor results even when the segmentation method considering the strength of Section 3.3.2 was applied.

In Table 2, SCILN and SCILE methods showed excellent BLEU-4, Rouge Score, and BERTSCORE [22], and SCILE had the lowest perplexity. However, the SSLN method also yielded similar figures. We conducted human judgments to evaluate the grammar and natural context flow of the sentences generated by each method. A total of 203 college students participated, and they evaluated the automatically generated review sentences by filling out an internet questionnaire. For a single query, nine sentences generated by GPT-2 and 1 sentence generated by BART are randomly listed and evaluated for naturalness (comprehension of the entire content), grammar, strength, and goodness of fit. Among the ten sentences, the sentence that best fits the item was selected.

Table 3 is the Human Evaluation result table for the results generated by GPT-2 and BART. The Segmentation Considering Sentence Intensity & Labeling New score (SCILN) method received the highest score in all aspects of the overall generation result. In human judgments, SCILN showed higher values than other methods.

In fact, the SCILN-based generation-based results in Table 4 show more complete results than those of other methods.