Application of fine-grained sentiment analysis in meal arrival scenarios

The classic fine-grained sentiment analysis (ABSA, Aspect-based Sentiment Analysis) mainly includes three sub-tasks, which are attribute extraction, opinion extraction, and attribute-view pair sentiment tendency judgment three cascaded tasks. This article introduces the Meituan-to-store-to-dining application algorithm team through combining the most advanced reading comprehension, attention mechanism and other aspects of entity extraction and sentiment analysis experience to solve the problem of four-tuple extraction of meals (dishes, attributes, opinions, emotions) Problems, and applied in multiple business scenarios, I hope it can be helpful or inspiring to students who are engaged in related work.

1. Background

As an online e-commerce platform for life services, Meituan is committed to connecting consumers and merchants through technology and striving to provide consumers with a quality life. As one of Meituan’s core businesses, in-store catering (to-dining for short) is an important platform that satisfies users’ dine-in consumption needs and empowers catering merchants’ online operations. In the process, a large amount of User Generated Content (UGC) has been accumulated, which contains the true feelings of users after the consumer experience in the store. If the key emotional polarity and opinion expression can be extracted effectively, it will not only assist more Users make consumption decisions and can also help merchants collect user feedback on business conditions.

In recent years, NLP technologies such as large-scale pre-training models (BERT) and prompt learning (Prompt) have developed rapidly. The application effects of various natural language processing tasks such as text classification, sequence labeling, and text generation have been significantly improved, and sentiment analysis is one of the most common application forms. Its task goal is to analyze, process, summarize, and reason the input text through NLP technology, and give the result of text sentiment polarity judgment.

According to the granularity of sentiment polarity, it can be subdivided into chapter/sentence granular sentiment analysis and fine-grained sentiment analysis (ABSA, Aspect-based Sentiment Analysis) ^[1] . Generally speaking, the task goals of fine-grained sentiment analysis are mainly developed around the three elements of attribute (Aspect Term), opinion (Opinion Term), and sentiment (Sentiment Polarity), which can be divided into attribute extraction, opinion extraction, and attribute-view pair emotion Tendency to determine three cascading tasks ^[2-5] . For example, for a given user comment "This store has a good environment, but the service is terrible", the expected output is (environment, good, positive), (service, bad, negative).

The restaurant algorithm team combines the business scenarios of the restaurant business supply side, platform side, and demand side to provide efficient and high-quality algorithm solutions for the intelligentization of core business links, and helps reduce business costs and improve efficiency through algorithm capabilities. This article combines the dining B/C end business scenarios to explore the application and practice of fine-grained sentiment analysis technology in the direction of user evaluation mining.

2. Goal review

2.1 Business issues

Adhering to the mission of "Helping everyone eat better and live better", Daedian provides consumers with a wealth of products and services including set meals, vouchers, billing, reservations, etc., and has passed the Black Pearl restaurant guide and public comment. Lists such as eating lists, as well as search, query, evaluation, etc., help consumers make better consumption decisions. At the same time, it provides one-stop marketing services for merchants to help catering merchants accumulate reputation, acquire users, increase repurchase, etc., and then easily manage restaurants.

With the acceleration of restaurant chains and fierce competition in the industry, the breadth and difficulty of merchant management have gradually increased, the business requirements of merchants have become more refined, and the awareness of data management has become more urgent. User historical comments contain a large number of user 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, a good sentiment analysis is helpful to help catering stores improve service quality, and it can also better promote the consumer experience.

UGC evaluation analysis is mainly to dig out relevant information about dishes, services, food safety (abbreviated as food safety) from the review text, obtain the fine-grained emotions of users in various dimensions, and meticulously portray the current status of the merchant's service, as shown in Figure 2 above. . For catering merchants, the evaluation and analysis of dishes, services, and food safety can be disassembled as follows:

1. Dishes evaluation mainly includes the recognition of dishes in user reviews, the extraction of evaluation attributes, the extraction of opinions on dishes, and the classification of opinions and sentiments;
2. Service evaluation mainly includes the extraction of evaluation attributes from user reviews, the extraction of opinions on services, and the classification of opinions and sentiments;
3. Food safety evaluation mainly includes the extraction of evaluation attributes from user reviews, the extraction of food safety opinions, and the classification of opinions and emotions.

Among them, questions 2 and 3 are typical triad extraction tasks, that is, to identify service or food safety aspects (attributes, opinions, emotions). For question 1, on the basis of service and food safety evaluation issues, dish evaluation needs to identify the dishes mentioned in the reviews, compared to the industry quadruple (attributes, opinions, attribute categories, emotions) ^[6] extraction task, In the meal scene, it is mainly the identification of (dishes, attributes, opinions, emotions) quadruples.

2.2 Technical research

In Meituan, we investigated the related work results for UGC evaluation and analysis problems, mainly developed a multi-task model based on the MT-BERT pre-training model, and tried to solve the ACSA (Aspect-Category Setiment Analysis) problem in sentiment analysis and ( Attribute, opinion, sentiment) triples extraction problem, and realized the development of sentence-granular sentiment classification tools, and open source the Chinese attribute-level sentiment analysis data set ASAP ^[7-9] based on real scenes. But for Meituan’s to-dining business, we need to propose targeted solutions based on specific scenarios, such as the quadruple extraction task. It is not possible to directly reuse related technologies and tools of other teams. Therefore, it is necessary to build services for the to-dining business. Fine-grained sentiment analysis technology for scenes.

In the industry, we have also investigated relevant research on fine-grained sentiment analysis by other teams in the industry, such as Tencent and Ali. In 2019, Tencent AI Lab and Ali Dharma Academy collaborated ^[3] , and proposed a model based on two stacked LSTMs and three components (boundary guidance, emotional consistency, and opinion enhancement), which combined the "BIOES" labeling system with Positive, neutral, and negative emotions combine 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 attributes and emotions ^[10] , and also proposed (attributes, opinions, emotions) ^[2] triplet extraction task , And gave a two-stage solution framework, firstly identify attributes (emotion fusion into a unified label) and viewpoints, and then judge whether the attribute-viewpoints are paired.

Since then, follow-up research in the industry has begun to expand ^[11-14] joint extraction of triples. In February 2021, Huawei Cloud ^[6] proposed a (attribute, opinion, attribute category, emotion) four-tuple extraction multi-task model, in which one task recognizes attributes and opinions, and the other task recognizes attribute categories and emotions. In April 2021, Tencent ^[15] introduced the Aspect-Sentiment-Opinion Triplet Extraction (ASOTE) task, and proposed a location-aware BERT three-stage model to solve the (attribute, opinion, emotion) triplet extraction problem.

Research institutions	Industry	pre-trained model	Fine-grained sentiment analysis problem	reading comprehension problem	triplet problem	Quadruple problem	Joint extraction problem
Ali Dharma Institute [2,10]	E-commerce	✓	✓	✗	✓	✗	✓
HUAWEI CLOUD [6]	cloud service	✓	✓	✗	✗	✓	✓
Tencent [15]	social contact	✓	✓	✗	✓	✗	✗
Meituan to Meal	local life	✓	✓	✓	✓	✓	✓

From the academic perspective, more attention is paid to how to better perform entity extraction, sentiment classification, and multi-task joint extraction, and may ignore the calculation timeliness that the industrial world pays more attention to (such as multi-dimensional labeling and emotional dimension integration, increasing calculations) , Storage resource consumption, long delays under limited resources), effect accuracy (such as end-to-end development of task modules, neglect of business personalization, direct reuse leads to reduced accuracy) and other requirements, resulting in related technical methods that cannot be directly Applied to business scenarios, further development and improvement are needed to achieve business landing.

As shown in the above table, in response to the above research, we draw on the experience of Meituan Search and NLP in the fine-grained sentiment analysis of triples, disassembled to the problem of quadruple extraction, and combined with the most advanced reading comprehension and attention in the academic world. Based on the experience of entity extraction and sentiment classification in terms of force mechanism, etc., we have designed and developed a fine-grained sentiment analysis solution applied to the meal business.

2.3 Technical goals

As mentioned above, dish evaluation mainly focuses on dishes, evaluation attributes, dish opinions and opinions and emotions, while service and food safety evaluation issues mainly focus on evaluation attributes, opinions and emotions in terms of service or food safety. As far as fine-grained sentiment analysis tasks are concerned, it can be seen that the former question involves quadruple information, while the latter two questions only involve triple information.

2.4 Main challenges

Since the triad problem can be regarded as a sub-problem of the quadruple problem, without loss of generality, the following will focus on the technical challenges related to the quadruple.

Question 3 : How to extract and identify four-tuples at the same time to reduce the cumulative impact of errors in the pipeline method?

To reduce the cumulative impact of errors in the pipeline method, the typical solution is to propose simultaneous processing of information extraction and classification tasks, that is, multi-task learning. The traditional method is to directly try the idea of multi-task learning, but the process of ignoring the dependency relationship between entities, and even the remote association relationship ^[2] . At present, it is also trying to directly transform the quaternion into a multi-task learning process. In the future, it is expected to conduct joint extraction and recognition by establishing a pair or triplet relationship between entities.

In summary, for question 1 and question 2 , we will use the results of pipeline recognition and then use strategies to optimize the extraction results; for problem 3 161b6f68e91d6f problem 3, 161b6f68e91d6f, integrated learning, and classification of entities, joint relationships and classification It helps to reduce the cumulative impact of errors in the pipeline method.

3. Fine-grained sentiment analysis practice

3.1 Pipeline method

As mentioned in question 2 of 2.3 above, we use the pipeline method to disassemble the quaternion extraction problem into three tasks, which are divided into entity recognition, opinion extraction, opinion category and sentiment classification, as shown in Figure 4 below:

3.1.1 Entity recognition

Since the ^[16] in 2018, the NER model has been replaced from the traditional LSTM+CRF to BERT+CRF (or BERT+LSTM+CRF). It was once the SOTA model of the industry’s NER task. In the past two years, NER tasks have been mainly from Make improvements in the following two areas:

1. Add additional features ^[17-19] : such as character features, word features, part-of-speech features, syntactic features, and knowledge graph representations;
2. Conversion task form ^[20-21] : Convert NER tasks into QA (Question Answering) tasks or machine translation tasks.

Taking into account that the introduction of additional features requires the construction of an artificial dictionary, and the conversion of the question and answer task form depends on the manual template, the cost is high, so the BERT+CRF model is adopted.

learning rate adjustment, model strategy adjustment . In the course of the experiment, we found that the effect of BERT+CRF is very small compared with the simple BERT+Softmax. The reason is that the pre-training model can learn features with obvious discrimination after fine-tuning, which leads to the increase of the CRF layer for entity recognition The result has almost no effect. However, a good CRF transition matrix is obviously helpful for prediction, and constraints can be added to the final predicted label to ensure the reasonableness of the prediction result. After further experiments, it is found that by adjusting the learning rate of the BERT and CRF layers, for example, BERT uses a smaller learning rate and the CRF layer uses 100 times the BERT learning rate (ie $e2/e1>100$, as shown in Figure 5) In the end, the effect of BERT+CRF has been significantly improved compared to BERT+Softmax. In addition, on the basis of 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 the LSTM layer was not finally introduced.

3.1.2 Opinion extraction

Opinion extraction task is also called Target-oriented Opinion Words Extraction (TOWE) in the industry, which aims to extract opinion words corresponding to a given target from comment sentences. Opinion extraction can also be regarded as a kind of NER task, but if a review involves multiple entities and opinions, how to accurately extract all "entity-viewpoint" relationships is a technical challenge. Learn from the idea of MRC (Machine Reading Comprehension) task, introduce prior knowledge by constructing reasonable Query, and assist in viewpoint extraction.

QA task format, opinion extraction modeling . As shown in Figure 6, the model as a whole consists of two parts: the pre-training layer and the output layer. In the output layer, we use conventional QA task output, including Start Label and End Label, but we need to manually design Quey. Refer to the paper ^[20] experience. Taking Figure 3 as an example, the experiment found that Query is designed to "find out the taste, texture, weight, ingredients, appearance, price, hygiene, and overall evaluation of shrimp dumplings." The effect is the best, maybe Incorporating viewpoint description information, it is more helpful for viewpoint extraction. Considering that the QA task naturally has the problem of category imbalance, the loss function introduces Focal Loss for category imbalance to improve the effect of the opinion extraction model. Since opinion extraction can also be regarded as a NER task, we tried to design the output layer as a CRF layer, but the experimental results were not ideal. It may be that the length of the opinion sentence is different and more personalized, which affects the model recognition. On the other hand, considering that the Google Chinese pre-training model BERT is segmented by word granularity, and the Chinese word segmentation in traditional NLP is not considered, in the pre-training layer we replace the BERT model with the open-source Chinese pre-training model of Harbin Institute of Technology, such as BERT -wwm-ext, RoBERTa-wwm, etc., the final model effect has been further improved.

3.1.3 Viewpoint categories and sentiment classification

Opinion category and sentiment classification can be regarded as two classification tasks. Among them, the opinion category of the four-tuple task of dish evaluation includes 8 labels such as taste, taste, weight, ingredients, appearance, price, hygiene, and overall dishes, and emotion contains positive The four labels of orientation, neutral, negative, and unmentioned are all predefined by the business. Considering that the opinions of a certain dish mentioned in user comments may involve multiple dimensions, if each dimension is modeled separately, multiple models need to be constructed, which is more complicated and difficult to maintain. Combining the ATAE-LSTM ^[22] and NLP Center ^[7-9] sentiment analysis experience and the characteristics of the meal business, the overall structure of the model is designed as a multi-task and multi-class learning framework.

Multi-task and multi-classification model, joint modeling of opinion categories and emotions . As shown in Figure 7, the model as a whole is divided into two parts, namely the BERT shared layer and the Attention exclusive layer. The BERT shared layer learning viewpoint Embedding indicates that the Attention exclusive layer learns the emotional tendency of viewpoints in each viewpoint category. Considering that each part of the comment will focus on different perspective dimensions, the introduction of the Attention structure makes the model pay more attention to text information related to a specific dimension, thereby improving the overall effect.

3.2 Joint learning

The advantage of the pipeline method is to split the target problem into multiple sub-module problems, optimize the sub-modules separately, and solve the problem of many-to-many relationships between entities to a certain extent through post-processing. However, the pipeline method also has some fatal flaws, including:

1. Error propagation, errors in the entity recognition module will affect the performance of the opinion extraction model;
2. Ignore the relevance between tasks, such as entities and opinions often appear together, if the opinions can be known, then the described entities can also be judged, and the pipeline method obviously cannot use this information;
3. Information redundancy, due to the need to extract opinions for the identified entities, and to classify the extracted opinions, resulting in some invalid matching pairs and increasing the error rate.

Refer to the current situation of joint learning in sentiment analysis in the industry, mainly for (attributes, opinions, sentiments) triple joint extraction. Combining the characteristics of the dining business scene (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 dish entities, opinions and emotions, and the second stage is the classification of opinions, and then four The result of tuple recognition.

3.2.1 Joint extraction of triples

At present, in academia, the methods for joint extraction of triples (attributes, opinions, emotions) mainly include sequence labeling method ^[11] , QA method ^[5,12] , generative method ^[13,14] Wait. Combining the dish analysis scenario and the experience of the opinion extraction module in the pipeline method, we adopted a QA-style joint extraction method, the main reference model Dual-MRC ^[5] .

Dual-MRC model, triple joint extraction modeling . In the process of model design, since the Dual-MRC model classifies sentiment tendency as an overall evaluation of a certain attribute, that is, one attribute corresponds to only one emotion. However, in the dining business scenario, the recognition of the dish entity is newly added, and at the same time, there are different opinions and emotional tendencies on the same dish entity in the UGC comments. As shown in Figure 3, "taste is very good" expresses positive feelings for "fresh shrimp dumplings", while "a bit expensive" obviously expresses negative feelings. Therefore, we modified the Dual-MRC model to integrate opinions and emotion tags into a unified tag. As shown in Figure 8, the overall structure of the Dual-MRC is based on the Twin Tower BERT model. By introducing two Query, the left side is responsible for extracting dish entities, and the right side is responsible for extracting opinions and opinion sentiments, so as to achieve joint extraction of triples.

Model structure description:

1. The whole is composed of two parts. The BERT on the left extracts dish entities, and the BERT on the right extracts opinions and sentiments. The opinions and sentiments form unified labels B-{POS, NEU, NEG}, I-{POS, NEU, NEG} and O , Which does not mention that emotion is integrated into the O tag;
2. Refer to the experience of the pipeline method to construct two Queys. Quey1 on the left is constructed to "find out the dishes in the review", and Quey2 on the right is constructed to "find out the taste, texture, weight, ingredients, presentation, price, hygiene and overall of the shrimp dumplings." Evaluation";
3. In the training stage, for each dish entity marked on the left, it is necessary to repeat the process on the right, and the two models share parameters for training; in the prediction stage, because the entity is unknown, the pipeline method is used. Each entity is input to the right part to extract opinions and opinion emotions.

On this basis, we also explored the possibility of joint extraction of four-tuples. The specific operation is to modify the Query2 on the right, such as "find out the taste evaluation of shrimp stuffed dumplings". For each opinion category, a Query needs to be constructed to predict. So as to realize the joint extraction of four-tuples. However, considering that the calculation is relatively large and time-consuming, the opinion category is finally forecasted separately.

3.2.2 View category classification

View category classification is obviously a text classification problem. The usual method is based on BERT classification, taking Embedding at the position of [CLS], and then a fully connected layer and a Softmax layer. In the meal business scenario, the main problem is the small sample problem. Refer to the industry's NLP small sample solution, 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, we experimented with the Prompt template method (shown in Figure 9) and R-drop data enhancement (shown in Figure 10). Among them, the Prompt template mainly draws on the idea of P-tuning ^[25] , adopts the method of automatically constructing the template, and solves the problem based on the MLM task.

In Figure 9 [u1]~[u6] represent [unused1]~[unused6] in the BERT vocabulary, that is, use unused tokens to construct a template, and the number of tokens is a hyperparameter. The experimental results found that, based on the pre-training model of BERT, combined with the P-tuning or R-drop structure, the classification effect can be improved to a certain extent, and the effect of P-tuning is slightly better than R-drop, and the follow-up will continue to explore fewer samples Solution.

Fourth, the application in the meal business

4.1 Comparison of model effects

Using the UGC annotation data of the meal, the overall effect of the quaternion recognition was evaluated, and finally the F1 value was calculated based on the accuracy and recall rate of the overall quaternion as the performance evaluation index. As shown in Figure 11, the classic BERT+CRF model is used for entity extraction. After the meal comment annotation data only reaches 0.61 F1, after the optimization of the learning rate and other tuning parameters (Baseline Tuning), the F1 value is increased by 2.61%. As mentioned above, in the opinion extraction module, after converting sequence labeling questions into question-and-answer (QA) questions, using the BERT+MRC model, F1 is significantly increased to 0.64, which is an increase of 5.9%, indicating that the problem conversion has achieved greater benefits. In addition, the use of Harbin Institute of Technology Chinese pre-training BERT still achieved a certain degree of improvement, and F1 increased to 0.65. Note that the model iteration in Figure 11 represents the model of the key optimization module in the quaternion problem, and the overall effect of the quaternion is finally evaluated for comparative analysis.

4.2 Business application scenarios

brand dashboard

As an important part of flagship store capabilities, brand dashboards provide brand-level data services to help business growth. The product is positioned as the data center of the leading catering brand, with basic data disclosure capabilities, and guides merchants' business decisions by quantifying business effects. Because Dake has accumulated a wealth of online information (a large amount of transaction/traffic/comment data) on the platform, there is a lot of room for mining and analysis. Fine-grained sentiment analysis technology is used to dig out information related to dish dimension, service dimension, and food safety dimension from review data, quantify business performance of merchants, and guide business actions. Regarding user feedback monitoring of dishes, brand merchants pay more attention to user feedback in the dimensions of dishes, taste, and taste. After the model iteration described above, the accuracy of recognition of dish emotion, taste emotion, and taste emotion has been improved to a certain extent.

Optimized the dish information of the restaurant merchants

With the arrival of the meal, the establishment of dish information has been strengthened, mainly including at the production level, the integration of the various source dish data of the merchants, the construction of the merchant dish center, and the optimization of the C-side dish UGC upload function, effectively supplementing the UGC dish production; at the consumption level In the above, it integrates the dishes from merchants and the dishes recommended by netizens, and optimizes the content aggregation and display consumption of the C-end dish information based on the improvement of dish information. At the same time, it cooperates with the catering business and continues to empower the production and construction of evaluation information, and more guide users to describe and introduce merchant dishes from the evaluation production level. Mainly aiming at the evaluation information related to the dishes of the restaurant merchants, the information linkage and the optimization of the display level are carried out. Compared with the previous iteration, the coverage of the evaluated dishes has been greatly improved.

shop treasure evaluation management

Merchants obtain users by providing catering services. After users consume, they provide feedback to the merchants through evaluations, urging them to continuously optimize and provide better services, so as to acquire more users and achieve a positive cycle. The significance of evaluation analysis is to establish a channel between evaluation and catering services to realize the positive promotion cycle of evaluation and service. By analyzing and evaluating the content, it helps businesses find the restaurant's dishes, services, environment, etc., where they are doing well and not doing well, and then make targeted improvements. Compared with the previous iteration, the number of comments related to dishes, services, and the environment has been greatly improved.

5. Future prospects

After nearly a year of construction, sentiment analysis related capabilities have not only been successfully applied to restaurant business operations, supply chain and other businesses, but also optimized multi-source dish information, assisted brand merchants in user feedback monitoring, and improved merchant service capabilities. In the joint learning exploration, currently the four-tuple problem is mainly transformed into a two-stage model. As shown in Figure 11, the F1 value has dropped to only 0.63. The reason may be that in the triple joint extraction model, the relationship between entities is ignored, especially the long-range relationship (as described in question 3 of 2.4 above), resulting in insufficient performance. Next, it will further improve the ability of sentiment analysis quadruple extraction, and explore the core needs and important feedback of users in UGC. In terms of technology, the model will continue to evolve iteratively, mainly involving:

1. continues to optimize existing models to ensure quality while improving efficiency

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

2. Explore the field of sentiment analysis in depth, and build a four-tuple joint extraction model

   The quadruple extraction is mainly achieved through the transformation of Query, but the magnitude of the calculation is relatively large. It is necessary to explore the optimization of the model structure and reduce the redundant calculation to meet the requirements of the combined extraction of the quadruple.

3. Building a general framework for fine-grained sentiment analysis

   The meal-to-dining scene involves multiple sentiment analysis scenarios, and a flexible and convenient general framework needs to be built to help quickly support the business and reduce resource consumption.

In the future, the team will continue to optimize the application technology to solve the emotional analysis needs in the dining business scenario. Fine-grained sentiment analysis is a challenging and promising task. The in-store dining algorithm team will continue to explore and research together with readers.

6. 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, etc. Exploration and practice of BERT in Meituan.
• [8] Ren Lei, Bu Jiahao, etc. The exploration and application of sentiment 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. Towards Generative Aspect-Based Sentiment Analysis. In 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 of Deep Bidirectional Transformers for Language Understanding. In NAACL-HLT, 4171–4186.
• [17] Yue Zhang and Jie Yang. 2018. Chinese ner using lattice 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 Gallofré Ocaña, Andreas L. Opdahl, and Csaba Veres. 2020. Named entity extraction for knowledge graphs: A literature overview. 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: 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: A systematic survey of prompting 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. arXiv preprint arXiv:2103.10385.

Seven, term explanation

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

8. Author introduction

Chu Zhe, Wang Lu, Runyu, Ma Ning, Jianlin, Zhang Kun, and Liu Qiang are all from the Meituan Dadian Business Group/Platform Technology Department.

Nine, recruitment information

Meituan’s to-store platform technology department’s meal business data strategy group dish knowledge map direction is mainly responsible for applying dish knowledge to meal-related businesses, and its mission is to provide efficient, high-quality, and intelligent application algorithm solutions for the meal business. Based on the massive amount of catering business data, the cutting-edge entity extraction, relationship mining, entity representation learning, fine-grained sentiment analysis, small sample learning, semi-supervised learning and other algorithm technologies are applied to provide algorithmic support for the catering business.

The business data strategy group has long-term recruitment of natural language processing algorithm experts/machine learning algorithm experts in the direction of dish knowledge map. Interested students can send their resumes to [e-mail].

Read more technical articles from the

the front | algorithm | backend | data | security | operation and maintenance | iOS | Android | test

161b6f68e92c4d | . You can view the collection of technical articles from the Meituan technical team over the years.

| This article is produced by the Meituan technical team, and the copyright belongs to Meituan. Welcome to reprint or use the content of this article for non-commercial purposes such as sharing and communication, please indicate "the content is reproduced from the Meituan technical team". This article may not be reproduced or used commercially without permission. For any commercial activity, please send an email to [e-mail] to apply for authorization.