Application of fine-grained sentiment analysis in food-to-meal scenarios

The classic aspect-based Sentiment Analysis (ABSA) mainly includes three sub-tasks, namely, three cascaded tasks of attribute extraction, opinion extraction and judgment of Sentiment tendency of attribution-opinion pair. Meituan to shop to eat application are introduced in this algorithm the team by combining field the most advanced entity extraction of reading comprehension, attention mechanism and so on, analysis of emotional experience, solve the meal (dishes, attributes, point of view, emotional) quad extraction problem, and fall to the ground in multiple business scenario application, hoping to help students engaged in related work or inspiration.

The background,

As an online life service e-commerce platform, Meituan is committed to connecting consumers and merchants through technology and striving to provide consumers with quality life. As one of the core businesses of Meituan, In-store catering (hereinafter referred to as in-store catering) is an important platform to meet users’ in-store consumption needs and enable catering merchants to operate online. In the process of serving millions of catering merchants and c-end users, meituan has accumulated massive User Generated Content (UGC). It contains the true feelings of users after their consumption experience in the store. If the key emotional polarity and opinion expression can be effectively extracted, it can not only assist more users to make consumption decisions, but also help merchants to collect user feedback information about their business conditions.

In recent years, large-scale pre-training model (BERT), Prompt learning and other NLP technologies have developed rapidly. The application effect of various natural language processing tasks such as text classification, sequence tagging and text generation has been significantly improved, among which sentiment analysis is one of the most common application forms. Its task objective is to analyze, process, induce and reason the input text by MEANS of NLP technology, and give the result of judging the polarity of text emotion.

The granularity can be subdivided into chapter/whole sentence granularity Sentiment Analysis and fine-grained Sentiment Analysis (ABSA, aspect-based Sentiment Analysis) according to emotional polarity [1]. In general, the task goals of fine-grained Sentiment analysis revolve around Aspect Term, Opinion Term, and Sentiment Sentiment. It can be divided into three cascade tasks: attribute extraction, viewpoint extraction and attribute-viewpoint pair emotion orientation judgment [2-5]. For example, for a given user comment “this store has a good environment but bad service”, the expected output is (environment, good, positive), (service, bad, negative).

In combination with the business scenarios of the supply side, platform side and demand side of the catering business, the catering algorithm team provides efficient and high-quality algorithm solutions for the intellectualization of core business links, and assists the business to reduce costs and improve efficiency through the algorithm ability. This paper explores the application practice of fine-grained sentiment analysis technology in user evaluation mining by combining with B/C business scenarios.

Second, objective review

2.1 Service Problems

Adhering to the “help you eat better, better life” mission, to eat for consumers, including package, vouchers, such as pay, booking, rich products and services, and through the black pearl restaurant guides, public review will eat list list, such as, query and search, evaluation, etc., to help consumers better consumption decisions. At the same time, it provides one-stop marketing services for merchants, helping catering merchants to accumulate reputation, acquire users, increase re-purchase, and so on, so as to easily manage restaurants.

With the acceleration of restaurant chain and the fierce competition in the industry, the breadth and difficulty of business management are gradually increasing, and the business requirements are more detailed, and the awareness of data management is more urgent. Historical user comments contain a large number of users’ feedback after consumption, which is an important part of sentiment analysis. It can not only describe consumption feelings, but also reflect the quality of the dining environment. Therefore, good emotional analysis is conducive to helping restaurants improve service quality and promote consumption experience.

UGC evaluation analysis mainly mines relevant information about dishes, services and food safety (food safety for short) from comment texts, obtains users’ fine-grained emotions in various dimensions, and describes the service status of merchants in detail, as shown in Figure 2. For catering merchants, the evaluation and analysis of dishes, services and food safety can be disassembled as follows:

1. Dish evaluation mainly includes dish recognition, evaluation attribute extraction, dish viewpoint extraction and viewpoint emotion classification in user comments.
2. Service evaluation mainly includes the extraction of evaluation attributes from user comments, the extraction of views on service and the classification of sentiment of views.
3. Food safety evaluation mainly includes the extraction of evaluation attributes from user reviews, the extraction of food safety views, and the classification of opinion emotions.

Questions 2 and 3 are typical triplet extraction tasks, that is, identifying aspects of service or food (attribute, viewpoint, emotion). For question 1, based on the evaluation of service and food safety, dish evaluation needs to identify the dishes mentioned in the comments. Compared with the quad (attribute, viewpoint, attribute category and emotion) in the industry [6], the extraction task is mainly the recognition of quad (dish, attribute, viewpoint and emotion) in the meal scene.

2.2 Technical Research

In Meituan, we investigated the relevant work results for UGC evaluation and Analysis, mainly developed a multi-task model based on MT-Bert pre-training model, and attempted to solve ACSA (aspect-category Setiment Analysis) and (attribute, viewpoint, In addition, a sentence granularity emotion classification tool was developed. Meanwhile, ASAP, a Chinese attribute-level emotion analysis data set based on real scenes, was opened source [7-9]. However, for Meituan catering business, we need to put forward targeted solutions based on specific scenarios, such as quad extraction task, and cannot directly reuse related technologies and tools of other teams. Therefore, it is necessary to build fine-grained sentiment analysis technology for catering business scenarios.

In the industry, we also investigated other teams in the industry, such as Tencent and Alibaba, in the field of fine-grained sentiment analysis. In 2019, Tencent AI Lab and Ali Dharma Institute collaborated [3] and proposed a model based on two stacked LSTM and three components (boundary guidance, emotional consistency and opinion enhancement). The “BIOES” labeling system is combined with Positive, Neutral and Negative emotions to form a unified label, which can identify attributes and emotions at the same time. In the same year, Ali Dharma Institute proposed the BERT+E2E-ABSA model structure to further solve the joint extraction problem of attribute and emotion [10], and proposed the triplet extraction task (attribute, viewpoint, emotion) [2], and gave a two-stage solution framework. Firstly, attribute (emotion fusion as unified label) and viewpoint were identified respectively. Then determine if the attribute-viewpoint is paired.

Since then, subsequent studies in the industry began to focus on joint triplet extraction [11-14]. In February 2021, Huawei Cloud [6] proposed a quad extraction multi-task model (attribute, opinion, attribute category, emotion), in which one task identifies attribute and opinion, and another task identifies attribute category and emotion. In April 2021, Tencent [15] introduced aspect-sentiment-opinion Triplet Extraction (ASOTE) task and proposed a BERT three-stage model of position perception, which solved the Triplet Extraction problem of (attribute, viewpoint and emotion).

Research institutions	industry	Pretraining model	Fine – grained emotion analysis problems	Reading comprehension problems	Triplet problem	The quad problem	Joint extraction problem
Ali Dharma Courtyard(2, 10]	The electronic commerce	✓	✓	✗	✓	✗	✓
Huawei cloud[6]	The cloud service	✓	✓	✗	✗	✓	✓
tencent[15]	social	✓	✓	✗	✓	✗	✗
Meituan to meals	The local life	✓	✓	✓	✓	✓	✓

From academia, more focused on how to better for entity extraction, the emotional classification and joint extraction, multitasking may ignore the computing efficiency of the industry is more focused on landing (such as multi-dimensional annotation and emotional dimension integration, increase the computing and storage resources consumption, time delay) under limited resources, effects the accuracy (such as end-to-end development task module, As a result, relevant technical methods cannot be directly applied to business scenarios, and further development and improvement are required to achieve the implementation of business.

As shown in the table above, in view of the above research, we draw lessons from the Meituan search and the ministry of NLP in triples fine-grained emotional experience analysis, dismantling to eat quad extraction problem, and combining the educational world’s most advanced entity extraction of reading comprehension, attention mechanism and so on, the classification of the emotional experience, designed and developed is applied to the meal business fine-grained sentiment analysis solution.

2.3 Technical Objectives

As mentioned above, dish evaluation mainly focuses on dishes, evaluation attributes, dish views and sentiment, while service and food safety evaluation mainly focuses on evaluation attributes, views and emotions of service or food safety. As far as the fine-grained sentiment analysis task is concerned, it can be seen that the former question involves quad information, while the latter two questions only involve triad information.

2.4 Main Challenges

Since the triplet problem can be viewed as a subproblem of the quad problem without loss of generality, the technical challenges related to quads will be highlighted below.

Question 3: How to reduce the impact of pipeline method error accumulation by quad extraction and identification simultaneously?

A typical solution to reduce the cumulative impact of pipeline methods is to deal with information extraction and classification tasks simultaneously, that is, multi-task learning. The traditional method is to directly try the idea of multi-task learning, but the dependency relationship between entities and even the remote association relationship are ignored in the process [2]. At present, we are also trying to directly transform the quintuple into a multi-task learning process. In the future, we hope to establish a pair or triplet relationship between entities for joint extraction and recognition.

To sum up, for problem 1 and problem 2, we will optimize the extraction results according to the results identified by pipeline using strategies. For problem 3, integrating entities, relationships, and classification tasks for joint learning will help reduce the cumulative impact of errors in the pipeline approach.

3. Practice of fine-grained emotion analysis

3.1 method of Pipeline

As mentioned in Question 2 of 2.3 above, we adopted the method of Pipeline to disassemble the quad extraction problem into three tasks, including entity recognition, viewpoint extraction, viewpoint category and emotion classification, as shown in Figure 4 below:

3.1.1 Entity recognition

Since BERT[16] appeared in 2018, NER model has been replaced by BERT+CRF (or BERT+LSTM+CRF) from traditional LSTM+CRF. It was once the SOTA model of NER tasks in the industry. In recent years, NER tasks have been improved mainly from the following two aspects:

1. Additional features are added [17-19], such as character feature, word feature, part of speech feature, syntactic feature and knowledge map representation;
2. Conversion of task form [20-21] : NER task was transformed into QA (Question Answering) task or machine translation task.

Considering that the introduction of additional features requires the construction of an artificial dictionary, and the high cost of transforming the form of question answering task depends on the manual template, BERT+CRF model is adopted.

Learning rate adjustment, model strategy tuning. During the experiment, we found that the effect of BERT+CRF was slightly improved compared with that of simple BERT+Softmax. The reason was that the pre-training model could learn features with obvious differentiation after fine-tuning, so the addition of CRF layer had almost no effect on the entity recognition result. However, a good CRF transition matrix is obviously helpful for prediction and can add constraints to the labels of final prediction to ensure the rationality of prediction results. Further experiments found that by adjusting the learning rate of BERT and CRF layer, for example, BERT uses a smaller learning rate while CRF layer uses a learning rate 100 times that of BERT (e2/ E1 > 100E2 / E1 > 100E2 / E1 >100, as shown in FIG. 5), Finally, the effect of BERT+CRF was significantly improved compared to BERT+Softmax. In addition, based on the traditional NER model LSTM+CRF, we also experimented with BERT+LSTM+CRF, but the effect actually decreased slightly and the prediction time also increased, so LSTM layer was not introduced in the end.

3.1.2 Opinion extraction

Opinion Extraction task is also known as target-oriented Opinion Words Extraction (TOWE), which aims to extract Opinion Words corresponding to a given objective from comment sentences. Opinion extraction can also be viewed as a NER task, but it is a technical challenge to accurately extract all entity-opinion relationships when reviews involve multiple entities and viewpoints. Based on the idea of Machine Reading Comprehension (MRC) task, a priori knowledge is introduced by constructing a reasonable Query to assist opinion extraction.

QA task form, viewpoint extraction modeling. As shown in Figure 6, the model as a whole is composed of two parts: pre-training layer and output layer. For the output layer, we used the regular QA task outputs, including Start Label and End Label, but manually designed Quey. Referring to the experience of the paper [20], and taking Figure 3 as an example, the experiment found that Query design designed to “find out the taste, taste, weight, ingredients, appearance, price, hygiene and overall evaluation of fresh shrimp dumpling” had the best effect, which may incorporate viewpoint description information and be more conducive to viewpoint extraction. Considering that QA task naturally has the problem of category imbalance, Focal Loss for category imbalance is introduced into Loss function to improve the effect of viewpoint extraction model. Since point of view extraction can also be regarded as NER task, we tried to design the output layer as CRF layer, but the experimental effect was not ideal, possibly because the length of point of view statements was different and personalized, which affected model recognition. On the other hand, considering that Google Chinese pre-training model BERT is word granularity segmentation, without considering Chinese word segmentation in traditional NLP, we replace BERT model with Hit open source Chinese pre-training model, such as Bert-WWM-ext and Roberta-WWM, etc., in the pre-training layer. The final effect of the model was further improved.

3.1.3 Classification of viewpoints and emotions

View category and emotion classification can be classified as two tasks, including food evaluation quad task view category contain taste, flavor, components, ingredients, presentation, price, health, food as a whole, such as eight tags, and emotion contains positive, neutral, negative, not mentioned, such as four labels, are predefined. Considering that user comments on a certain dish may involve multiple dimensions, if each dimension is modeled separately, multiple models need to be built, which is complicated and difficult to maintain. Combined with the experience of ATAE-LSTM[22] and NLP Center [7-9] in sentiment analysis and the characteristics of catering business, the overall structure of the model is designed as a multi-task and multi-classification learning framework.

Multi-task and multi-classification models, jointly modeling idea categories and emotions. As shown in Figure 7, the model as a whole is divided into two parts: BERT shared layer and Attention exclusive layer. Learning view Embedding in BERT shared layer represents the emotional tendency of learning view Embedding in Attention exclusive layer in each view category. Considering that each section of the review focuses on different perspectives, the Attention structure is introduced to make the model pay more Attention to the text information related to specific perspectives, thus improving the overall effect.

3.2 Joint learning

The advantage of pipeline method is that the target problem is divided into several sub-module problems, and the sub-module is optimized respectively. Through post-processing, the many-to-many relationship between entities can be solved to a certain extent. However, the pipeline approach can have some fatal flaws, including:

1. Error propagation, the error of entity recognition module will affect the performance of opinion extraction model;
2. Ignore the correlation between tasks, such as entities and ideas often appear together, if the idea can be known, then it can also determine the described entity, and the pipeline method obviously can not use this information;
3. Information redundancy, because it is necessary to extract views of the identified entities and classify the extracted views, some invalid matching pairs are generated and the error rate is improved.

Referring to the current situation of joint learning of sentiment analysis in the industry, the main method is joint extraction of triples (attribute, viewpoint and emotion). Combined with the characteristics of food business scenarios (as described in Question 2 of Challenge 2.3), the overall design is a two-stage model. The first stage is the joint training of food entities, viewpoints and emotions, and the second stage is the classification of viewpoints, so as to obtain the results of quad recognition.

3.2.1 Triplet joint extraction

At present, the methods of triplet (attribute, viewpoint and emotion) joint extraction mainly include sequence annotation method [11], QA method [5,12] and generative method [13,14]. Combining the experience of dish analysis scene and viewpoint extraction module in pipeline method, we adopted qA-style joint extraction method, mainly referring to the model Dual-MRC[5].

Improvement of Dual-MRC model, triplet joint extraction modeling. In the process of model design, the dual-MRC model classiifies the emotional orientation as an overall evaluation of a certain attribute, that is, one attribute corresponds to only one emotion. However, in the scene of catering business, the recognition of dish entities is added, and UGC comments contain different views and emotional tendencies for the same dish entity. As shown in Figure 3, “very good taste” expressed positive feelings towards “fresh shrimp dumpling”, while “a little expensive” obviously expressed negative feelings. Therefore, we modified the Dual-MRC model to integrate opinion and emotion labels into a unified label. As shown in Figure 8, the overall structure of dual-MRC is based on the two-tower BERT model. By introducing two queries, the left side is responsible for extracting dish entities, and the right side is responsible for extracting views and emotions, so as to realize the joint extraction of triples.

Description of model structure:

1. The whole is composed of two parts. On the left, BERT extracts food entities, and on the right, BERT extracts viewpoints and emotions to form a unified label B-{POS, NEU, NEG}, I-{POS, NEU, NEG} and O. Emotions are not mentioned in the label O.
2. Referring to the experience of pipeline method, two Quey were constructed. Quey1 on the left was constructed as “find the dishes in the comments”, and Quey2 on the right was constructed as “find the taste, taste, weight, ingredients, appearance, price, hygiene and overall evaluation of fresh shrimp dumpling”.
3. In the training stage, for each dish entity marked on the left, the process on the right needs to be repeated, and the models on both sides share parameters for training; In the prediction stage, because entities are unknown, pipeline method is adopted to extract all dishes entities from the left part first, and then input each entity to the right part to extract views and feelings.

On this basis, we also explore the possibility of joint extraction of quads. The specific operation is to transform Query2 on the right, such as “find out the taste evaluation of fresh shrimp dumpling”. For each opinion category, we need to construct Query for prediction, so as to achieve joint extraction of quads. However, considering the large calculation magnitude and long time consuming, the opinion category is finally made another prediction.

3.2.2 Classification of viewpoints

Embedding in [CLS] position is based on BERT classification, and then a full connection layer and Softmax layer are added. In the catering business scenario, the problem of small samples is mainly faced. The solution of small samples of NLP in the industry is referred to, represented by the r-DROP [23] method based on comparative learning and the fourth paradigm based on Prompt[24]. Based on the BERT model structure, Prompt template method (as shown in Figure 9) and R-DROP data enhancement (as shown in Figure 10) were respectively tested. Among them, Prompt template mainly refers to the idea of P-Tuning [25], adopts the method of automatic template construction, and solves problems based on MLM tasks.

In Figure 9, [U1]~[U6] represents [UNused1]~[unused6] in the BERT thesaurus, that is, the template is constructed using the unknown tokens, and the number of tokens is a hyperparameter. The experimental results show that the bert-based pre-training model, combined with P-TUNING or R-DROP structure, can improve the classification effect to some extent, and the p-TUNING effect is slightly better than R-DROP. The follow-up will continue to explore solutions for small samples.

Four, in the application of catering business

4.1 Comparison of model effects

The UGC labeling data of the food were used to evaluate the overall effect of quad recognition. Finally, the accuracy rate and recall rate of the whole quad were used to calculate the F1 value as the performance evaluation index. As shown in Figure 11, the classical BERT+CRF model is used for entity extraction. After the Baseline Tuning optimization, the F1 value increases by 2.61% when the annotation data of meal reviews only reaches F1 of 0.61. As mentioned above, in the opinion extraction module, after the sequence labeling question was transformed into QA question, the BERT+MRC model was adopted, and F1 was significantly improved to 0.64, an increase of 5.9%, indicating that the question transformation had gained great benefits. In addition, the BERT pre-training of Hit Chinese still improved to a certain extent, and F1 was improved to 0.65. Note that the model iteration in Figure 11 represents the model of the key optimization module in the quad problem, and the overall effect of the quad is finally evaluated for comparative analysis.

4.2 Service Application Scenarios

Brand dashboard

As an important part of the flagship store’s capabilities, brand dashboards provide brand-level data services to drive business growth. The product is positioned as the data center of the head catering brand, with basic data disclosure ability, and guide business decisions by quantifying business effects. Due to the abundant online information (a large amount of transaction/traffic/comment data) deposited by big customers on the platform, there is a large space for mining and analysis. Fine-grained emotion analysis technology is applied to mine information related to dishes dimension, service dimension and food safety dimension from review data, quantify business performance of merchants, and guide business actions. As for the user feedback monitoring of dishes, brand merchants pay more attention to the user feedback of dishes, taste, taste and other dimensions. After iteration of the model mentioned above, the recognition accuracy of dish emotion, taste emotion and taste emotion has been improved to some extent.

Food information optimization for restaurant merchants

Along with The food to strengthen the construction of dish information, mainly at the production level, integrating the data of dishes from various sources of merchants, building the dish center of merchants, and optimizing the UGC uploading function of C-end dishes to effectively supplement UGC dish production; On the consumption level, the general dishes of merchants and recommended dishes of netizens are integrated, and the content aggregation and display consumption of C-end dishes information are optimized based on the perfection of dishes information. At the same time, the restaurant business will continue to be empowered by the production construction of evaluation information, and more users will be guided to describe and introduce the dishes of merchants from the evaluation production level. Mainly for the evaluation information related to dishes of restaurant merchants, the optimization of information linkage and display level was carried out. Compared with before iteration, the coverage of evaluation dishes was greatly improved.

Store treasure evaluation management

Businesses gain users by providing catering services. After consumption, users give feedback to businesses through evaluation, prompting businesses to constantly optimize and provide better services, so as to obtain more users and achieve a positive cycle. The significance of evaluation analysis lies in establishing the channel between evaluation and catering service, and realizing the positive promotion cycle of evaluation to service. Through the analysis and evaluation of the content, to help businesses find restaurants in dishes, services, environment and other aspects, do good and bad, and then targeted improvement. Compared with before iteration, the number of related comments of dishes, service and environment has been greatly improved.

5. Future prospects

After nearly a year of construction, the ability of sentiment analysis has not only been successfully applied to restaurant business operation, supply chain and other businesses, but also optimized multi-source dish information, assisted brand merchants in user feedback monitoring, and improved their service capabilities. In terms of joint learning exploration, the quad problem is mainly transformed into a two-stage model at present, as shown in FIG. 11, where F1 value drops to 0.63. The reason may be that in the triplet joint extraction model, the relationship between entities is ignored, especially the long range relationship (as mentioned in question 3 of 2.4 above), resulting in the performance is not expected. Next, the extraction ability of sentiment analysis quad will be further improved, and the core needs and important feedbacks of UGC users will be mined. In terms of technology, the model will continue to evolve iteratively, mainly involving:

1. Continue to optimize existing models to ensure quality and improve efficiency

   There is still a lot of room for improvement in the experimental results, and it is necessary to further explore model optimization methods, such as optimizing the pre-training model, using MT-BERT, and further introducing inter-entity relations in joint extraction to improve the performance of quintuple extraction.

2. Deeply explore the field of emotion analysis and construct a four-tuple joint extraction model

   The quatertuple extraction is mainly realized through the transformation of Query, but the calculation magnitude is large, so it is necessary to explore the optimization of model structure, reduce the amount of redundant calculation, and make it meet the joint extraction of quatertuples.

3. Construct a general framework for fine – grained sentiment analysis

   The meal scenario involves multiple emotion analysis scenarios, and a flexible and convenient general framework needs to be built to help quickly support businesses and reduce resource consumption.

In the future, the team will continue to optimize the application technology to address the needs of emotion analysis in catering business scenarios. Fine-grained sentiment analysis is a challenging and promising task, and the in-store catering algorithm team will continue to explore and research together with readers.

Vi. References

• [1] Liu, B. 2012. Sentiment analysis and opinion Mining. Synthesis lectures on Human Language Technologies 5(1):1 — 167.
• [2] Peng, H. Xu, L. Bing, L. Huang, F. Lu, W. and Si, L.2020. Knowing What, How and Why: A Near Complete Solution for aspect-based Sentiment Analysis. In AAAI, 8600 — 8607.
• [3] Li, X. Bing, L. Li, P. and Lam, W. 2019a. A unified model for opinion target Extraction and target sentiment prediction. In AAAI, 6714 — 6721.
• [4] Zhao, H. Huang, L. Zhang, R. Lu, Q. and Xue, H. 2020. SpanMlt: A span-based multi-task Learning Framework for pair-wise Aspect and Opinion Terms Extraction. In ACL, 3239 — 3248.
• [5] Y. Mao, Y. Shen, C. Yu, and L. Cai. 2021. A joint training dual-mrc framework for aspect based sentiment analysis. arXiv preprint ArXiv: 2101.00816.
• [6] Huawei Cloud fine-grained text sentiment analysis and application
• [7] Yang Yang, Jia Hao et al. Meituan BERT’s exploration and practice.
• [8] Ren Lei, BU Jiahao et al. Exploration and application of emotion analysis technology in Meituan.
• [9] Bu J, Ren L, Zheng S, et al. ASAP: A Chinese Review Dataset Towards Aspect Category Sentiment Analysis and Rating Prediction. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.
• [10] Xin Li, Lidong Bing, Wenxuan Zhang, and Wai Lam. Exploiting BERT for end-to-end aspect-based sentiment analysis. In W-NUT@EMNLP, 2019.
• [11] Xu, L. Li, H. Lu, W. and Bing, L. 2020. Position-Aware Tagging for Aspect Sentiment Triplet Extraction. In EMNLP, 2339–2349.
• [12] Chen, S. Wang, Y. Liu, J. and Wang, Y. 2021a. Bidirectional Machine Reading Comprehension for Aspect Sentiment Triplet Extraction. In AAAI.
• [13] Yan, H. Dai, J. Qiu, X. Zhang, Z. et al.2021. A Unified Generative Framework for aspect-based Sentiment Analysis. ArXiv Preprint arXiv:2106.04300.
• [14] Wenxuan Zhang, Xin Li, Yang Deng, Lidong Bing, And Wai Lam. 2021. Research on Generative Aspect — Based Sentiment Analysis. ACL/IJCNLP 2021, 504 — 510.
• [15] Li Yuncong, Fang Wang, Zhang Wenjun, Sheng-hua Zhong, Cunxiang Yin, & Yancheng He. 2021. A More Fine-Grained Aspect-Sentiment-Opinion Triplet Extraction Task. arXiv: Computation and Language.
• [16] Devlin, J. Chang, M.-W. Lee, K. and Toutanova, K. 2019. BERT: Pre-training for Deep Bidirectional Transformers for Language Understanding. In NAACL-HLT, 4171 — 4186.
• [17] Yue Zhang and Jie Yang. 2018. Chinese Lattice Using LSTM. ArXiv Preprint arXiv:1805.02023.
• [18] Li, X. Yan, H. Qiu, X. and Huang, X. 2020. FLAT: Chinese NER Using flat-lattice Transformer. ArXiv preprint arXiv:2004.11795.
• [19] Tareq Al-Moslmi, Marc Gallofre Ocana, Andreas L. Opdahl, and Csaba Veres. 2020. Named entity extraction for knowledge graphs: IEEE Access 8 (2020), 32862 — 32881.
• [20] X. Li, J. Feng, Y. Meng, Q. Han, F. Wu, And J. Li.2020. A Unified MRC framework for Named Entity Recognition. In ACL, 5849 — 5859.
• [21] Jana Strakova, Milan Straka, and Jan Hajic. 2019. Neural architectures for nested ner through linearization. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 5326 — 5331.
• [22] Yequan Wang, Minlie Huang, Li Zhao, and Xiaoyan Zhu. 2016. Attention-based lstm for aspect-level sentiment classification. In Proceedings of the conference on empirical methods in natural language processing, 606–615.
• [23] Liang X, Wu L, Li J, et al. R-Drop: A Regularized Dropout for Neural Networks[J]. ArXiv PrePrint arXiv:2106.14448, 2021.
• [24] P. Liu, W. Yuan, J. Fu, Z. Jiang, H. Hayashi, and G. Neubig. 2021. Pre-train, prompt, and predict: Only A systematic survey of heavily-guarded methods in natural language processing. ArXiv Preprint arXiv:2107.13586.
• [25] X. Liu, Y. Zheng, Z. Du, M. Ding, Y. Qian, Z. Yang, and J. Tang. 2021. Gpt understands, Too. The arXiv preprint arXiv: 2103.10385.

7. Explanation of terminology

The term	explain
ABSA	Fine-grained Sentiment Analysis, aspect-based Sentiment Analysis
NER	Named Entity Recognition
TOWE	Target-oriented Opinion Words Extraction
MRC	Machine Reading Comprehension
MLM	Masked Language Model
BERT	Bidirectional Encoder Representations from Transformers
CRF	Conditional Random Fields; Conditional Random Fields
LSTM	Long -Term Memory
R-drop	Regularization strategy based on Dropout, Regularization strategy upon dropout

8. Introduction to the author

Chu Zhe, Wang Lu, Run Yu, Ma Ning, Jian Lin, Zhang Kun, Liu Qiang, all from Meituan to Store business group/platform technology Department.

Recruitment information

Meituan in-service platform Technology Department’s dish knowledge map direction of in-service business data strategy group is mainly responsible for the application of dish knowledge to in-service business, the mission is to provide efficient, high-quality and intelligent application algorithm solutions for in-service business. Based on massive incoming service data, cutting-edge entity extraction, relationship mining, entity representation learning, fine-grained emotion analysis, small sample learning, semi-supervised learning and other algorithm technologies are applied to provide algorithm capability support for incoming service.

We are looking for a natural language processing (NLP)/machine learning (ML) expert. Please send your resume to [email protected].

Read more technical articles from meituan’s technical team

Front end | | algorithm back-end | | | data security operations | iOS | Android | test

| in the public bar menu dialog reply goodies for [2020], [2019] special purchases, goodies for [2018], [2017] special purchases such as keywords, to view Meituan technology team calendar year essay collection.

| this paper Meituan produced by the technical team, the copyright ownership Meituan. You are welcome to reprint or use the content of this article for non-commercial purposes such as sharing and communication. Please mark “Content reprinted from Meituan Technical team”. This article shall not be reproduced or used commercially without permission. For any commercial activity, please send an email to [email protected] for authorization.