• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    TSCL-SQL:Two-Stage Curriculum Learning Framework for Text-to-SQL

    2023-09-22 14:30:46YINFengCHENGLuyi程路易WANGQiuyue王秋月WANGZhijun王志軍DUMingXUBo
    關鍵詞:王志軍路易

    YIN Feng(尹 楓), CHENG Luyi (程路易), WANG Qiuyue(王秋月), WANG Zhijun(王志軍), DU Ming(杜 明), XU Bo(徐 波)

    School of Computer Science and Technology, Donghua University, Shanghai 201620, China

    Abstract:Text-to-SQL is the task of translating a natural language query into a structured query language. Existing text-to-SQL approaches focus on improving the model’s architecture while ignoring the relationship between queries and table schemas and the differences in difficulty between examples in the dataset. To tackle these challenges, a two-stage curriculum learning framework for text-to-SQL(TSCL-SQL) is proposed in this paper. To exploit the relationship between the queries and the table schemas, a schema identification pre-training task is proposed to make the model choose the correct table schema from a set of candidates for a specific query. To leverage the differences in difficulty between examples, curriculum learning is applied to the text-to-SQL task, accompanied by an automatic curriculum learning solution, including a difficulty scorer and a training scheduler. Experiments show that the framework proposed in this paper is effective.

    Key words:text-to-SQL; curriculum learning; semantic parsing

    Introduction

    Text-to-SQL is the task of mapping a natural language query to a structured query language, which enables general users to query relational databases with natural languages. Limited by the scale of the dataset, early work can only complete the task on a single database with a few tables[1]. Recently, the release of the WikiSQL[2]dataset, which consists of more than 20 000 tables and about 80 000 natural language queries, presents a new challenge. The model is required to be generalized to unseen table schemas[2]and different kinds of queries.

    To tackle this challenge, existing text-to-SQL approaches cast the problem as a slot-filling[3]task. Xuetal.[4]utilized a multi-task model to fill the predicted values into a pre-defined grammar template. Heetal.[5]and Lyuetal.[6]further improved the model architecture and achieved better performance. However, the current text-to-SQL models still suffer from two challenges.

    The first challenge is that current approaches do not leverage the differences in difficulty between examples in the dataset. As shown in Fig.1(a), a simple query is related to fewer columns in the table, and the names of all related columns are mentioned in the query. In the simple example, the Winner and Runner-up columns are directly mentioned in the query. A complex query is shown in Fig.1(b). It is related to more columns, and some of the columns’ names are not mentioned in the query. In this complex example, the query is related to the Goals, Matches, Average and Team columns. However, the Team column is not mentioned in the query. The model must infer the column name from potential cell values. It makes sense that we can use the differences in difficulty to guide the training process.

    The second challenge is that current approaches do not utilize the relationship between queries and table schemas. As shown in Fig.1, a column name might be mentioned directly or indirectly in the query. The model is required to ground these potential mentions to the table schema. However, existing methods only consider the query’s corresponding table schema, which makes it difficult for the model to learn query-schema alignment.

    To address these shortcomings, a two-stage curriculum learning framework for text-to-SQL is proposed. Specifically, to leverage the differences in difficulty between examples, curriculum learning[7]is applied to the text-to-SQL task and an automatic curriculum learning solution is designed, including a difficulty scorer and a training scheduler. To exploit the relationship between queries and table schemas, a schema identification pre-training task is proposed to make the model choose the correct table schema from a set of candidates for a specific query. Experiments, including comprehensive ablation studies conducted on the WikiSQL dataset would demonstrate the effectiveness of the proposed method.

    Fig.1 Examples of text-to-SQL task:(a) simple example; (b) complex example

    1 Framework Overview

    In this section, the text-to-SQL problem is formulated and the two-stage curriculum learning framework for the text-to-SQL problem is introduced.

    1.1 Problem formulation

    Given a natural language queryQand a table schemaS=, the text-to-SQL task aims to output the corresponding SQL query. The table schema consists of the names of the columnsC={c1,c2, ,cn} and their corresponding typesT={t1,t2, ,tn}.

    1.2 Overall architecture

    As shown in Fig.2, the TSCL-SQL framework split the training process of the text-to-SQL task into two stages. Firstly, the query-schema alignment model was built at the pre-training stage. Specifically, a schema identification task was designed to retrieve the table schema for a specific natural language query. Based on the cross-encoder[8]with an in-batch negative[9]sampling strategy, the model chose the most similar table schema from the candidates for a specific query. Secondly, the curriculum learning was adopted, and the training process of the text-to-SQL task was re-designed with a difficulty scorer and a training scheduler at the curriculum learning stage. The difficulty scorer scored the difficulty of each training sample. The training scheduler organized training samples according to the score, from simple to complex, and split them into buckets to guide the optimization process.

    Fig.2 TSCL-SQL framework:(a) pre-training stage; (b) curriculum learning stage

    2 Pre-training Stage

    The objective of the pre-training stage is to enhance the encoder for the text-to-SQL task by establishing a strong alignment between the natural language query and the table schema. In order to build the alignment, a novel schema identification task is proposed to retrieve the relevant table schema for a given query. To facilitate this task, a dataset specifically designed for schema identification is constructed based on the WikiSQL dataset. The schema identification task is completed using a cross-encoder approach with an in-batch negative sampling strategy, effectively leveraging the power of the model to accurately identify and match query-table schema pairs.

    2.1 Dataset construction

    As shown in Table 1, the dataset mainly consists of the query and the table schema’s meta information. Since Wikipedia is the data source of the WikiSQL dataset, the corresponding table ID, article title, and section title from Wikipedia are concatenated as descriptions for each table schema. Figure 3 shows an example of the data source.

    Table 1 Information of schema identification dataset

    Fig.3 Data source of descriptions for a table schema

    2.2 Query-schema alignment model

    The query-schema alignment model aims to build a better encoder representation for the text-to-SQL task. A retrieval-based schema identification task of selecting the most similar table schema from a set of candidates for the given query is proposed. Figure 4 shows the architecture of the query-schema alignment model. It took the query and the description of the table schema as input and output a score representing the semantic consistency between the query and the table schema. The one with the highest score was chosen as the query’s corresponding table schema.

    sim(Qi,Di)=Linear(red(Encoder

    ([CLS]Qi[SEP]Di[SEP]))),

    (1)

    whereEncoder(·) represents the encoder of a pre-trained language model based on transformers[10];red(·) is a function that takes the representation of the first token from the last layer of the encoder;Linear(·) is a fully connected layer; [CLS] and [SEP] are special tokens.

    Fig.4 Architecture of query-schema alignment model

    (2)

    3 Curriculum Learning Stage

    The curriculum learning stage aims to use a curriculum learning framework to train a text-to-SQL model. A curriculum learning framework for the text-to-SQL task is introduced. Then, the implementation of two core components of the framework is described in detail.

    3.1 Curriculum learning framework

    The curriculum learning framework consists of a difficulty scorer and a training scheduler. The difficulty of each training sample is measured by an automatic difficulty scorer to avoid the uncertainty of hand-crafted rules and consider more about the feedback from the model. The overall process is as follows.

    Firstly, the difficulty scorer scores the samples and sorts them from easy to complex. Secondly, the training scheduler initializes an empty subset of the training set as a training stage starts. Sorted samples are split into buckets. For each training stage, a new bucket is added to the subset according to the difficulty. If the training on the subset is converged, the scheduler moves to the next stage until all the buckets are trained. Finally, the full training set is used for training for another few epochs.

    3.2 Difficulty scorer

    The difficulty scorer aims to score every training sample based on its difficulty. Due to the lack of existing information about which training samples are more difficult, instead of hand-craft rules, the model’s training loss is used as a measurement of the difficulties. A higher loss indicates that the sample is more complex and difficult.

    (3)

    Fig.5 Template used for slot-filling task

    3.3 Training scheduler

    The training scheduler aims to arrange the scored training samples for curriculum learning. As shown in Fig.6, the scheduler first sorts the training samples from easy to difficult and splits them intoNbuckets. Then it starts training with the easiest one. If the training process reaches convergence or a pre-defined number of epochs, a more difficult bucket will be merged. The scheduler will shuffle the data in the bucket and start training. After all the bucket is merged, it will train for several extra epochs on the complete training set.

    Fig.6 Training scheduler

    4 Experiments

    4.1 Dataset and metrics

    The proposed framework is evaluated on the WikiSQL dataset. It consists of tables from Wikipedia, natural language queries and their corresponding SQL statements. The basic characteristic of the WikiSQL dataset is shown in Table 2.

    Table 2 Basic characteristics of WikiSQL dataset

    Specifically, the natural language queries and their corresponding SQL statements are stored with JavaScript object notation. The tables are managed with SQLite database. Figure 7 shows an example of the training set.

    Fig.7 Example of training sample

    In Fig.7, table_id represents the corresponding table of a query; question is the natural language query, sql is the annotated SQL statement; agg and sel represent the column name and the aggregate function of the SELECT statement, respectively; conds are triplets (column-name, operator, value) of the WHERE statement.

    Logic form accuracyAccland execution accuracyAcceare used to evaluate the performance. Logic form accuracy considers whether the predicted SQL statement matches the ground truth. Execution accuracy considers if the execution result of the predicted SQL statement is the same as the execution result of the ground truth one. The formulas are as follows.

    (4)

    (5)

    (6)

    (7)

    whereNdenotes the size of a specific split of the dataset;SQL′ andSQLdenote the predicted SQL statement and the ground truth one, respectively;Y′ andYrepresent the execution result of the predicted SQL statement and the ground truth one, respectively.

    4.2 Execution-guided decoding

    Execution-guided decoding[11]uses beam search to expand the search space of candidate SQL statements and uses the execution result to filter the candidates. The main idea of execution-guided decoding is as follows. If the execution result of the predicted SQL statement is empty or the SQL parser cannot parse the statement, it is believed that the predicted SQL statement is wrong and another SQL statement will be selected from the candidates. In the following experiments, execution-guided decoding is used to further improve the performance.

    4.3 Parameter settings

    All the experiments were conducted on an entry-level server. Hardware and software configurations are shown in Table 3.

    Table 3 Hardware and software configurations

    Due to the limitation of the hardware, the implementation is based on RoBERTabase[13]. At the pre-training stage, the encoder was trained for three epochs. The initial learning rate was 3×10-6. At the curriculum learning stage, the model was first trained on the full training set for two epochs to get the difficulty scorer. Then the scored training samples were split into four buckets. After that the model was trained for three epochs for each training stage until all buckets were trained. Finally, the model was further trained on the full training set until converge.

    4.4 Baselines

    In order to establish a comprehensive performance comparison, multiple baseline methods for the text-to-SQL task are implemented and evaluated.

    1) HydraNet:HydraNet[6]proposes a hybrid ranking network, which casts the text-to-SQL problem as a column-wise ranking and decoding problem. It uses RoBERTa[12]as the encoder.

    2) HydraNet+Pt:the query-schema alignment pre-training method is implemented on that of the original HydraNet.

    3) HydraNet+CL:curriculum learning is applied to the original HydraNet.

    4) TSCL-SQL:the proposed method utilizes both query-schema alignment pre-training and curriculum learning.

    4.5 Results and analyses

    The results are shown in Tables 4 and 5, which demonstrate the framework’s performance under two scenarios, namely with execution-guided decoding (EG) and without EG.

    1) When EG is not applied, the logic form accuracy and the execution accuracy of the re-produced HydraNet model on the test set are 80.8% and 86.4%, respectively. The proposed model, TSCL-SQL, improves performance by 1.5% and 1.4%, respectively.

    2) When EG is applied, although the execution accuracy on the test set is already 91.0%, TSCL-SQL still improves the logic form accuracy and the execution accuracy by 0.9% and 0.5%, respectively.

    3) Ablation studies are conducted to investigate the effectiveness of the pre-training stage and the curriculum learning stage. If the pre-training stage is removed, the logic form accuracy and the execution accuracy will drop 0.5% and 0.6%, respectively, on the test set when EG is not applied. When EG is applied, there is still a slight decrease on the logic form and the execution accuracy if the pre-training stage is removed. It demonstrates that the pre-training stage would help the model initialize a better representation.

    Table 4 Performance of TSCL-SQL framework without EG

    Table 5 Performance of TSCL-SQL framework with EG

    Tables 6 and 7 show the performance comparison on all sub-tasks. TSCL-SQL achieves a performance improvement of 0.4% on the S-AGG sub-task compared to the baseline on the test set. On the other sub-tasks, the performance is still comparable. Therefore, TSCL-SQL is effective.

    Table 6 Development accuracy and test accuracy of various sub-tasks on Wiki SQL dataset without EG

    Table 7 Development accuracy and test accuracy of various sub-tasks on Wiki SQL dataset with EG

    Through analysis, it is found that both the pre-training stage and the curriculum learning stage are important. The pre-training stage provides a better representation for downstream tasks. The curriculum learning stage lets the model learn from easy tasks to complex tasks. It is beneficial for the model to approach the global minimum gradually and smoothly.

    5 Conclusions

    In this paper, a two-stage curriculum learning framework for text-to-SQL (TSCL-SQL) is proposed. At the pre-training stage, a schema identification pre-training task is proposed to build an alignment between queries and schemas. At the curriculum learning stage, an automatic curriculum learning solution is proposed for the text-to-SQL task. Experimental results demonstrate the effectiveness of the framework proposed in this paper.

    猜你喜歡
    王志軍路易
    國畫:慕思春雨
    當代作家(2024年3期)2024-06-29 22:18:20
    Effect of interface anisotropy on tilted growth of eutectics:A phase field study
    臭臭的路易
    王志軍 油畫作品
    真情本無語
    故事會(2018年3期)2018-02-07 15:28:22
    每個小孩都能說出天使一樣的句子
    最佳維權(quán)
    最佳維權(quán)
    愛你(2016年19期)2016-04-12 07:25:00
    對腎病患者的臨終關懷(短篇小說)
    南方文學(2015年3期)2015-07-15 08:03:15
    Numerical Simulation on New Perforator
    亚洲精品成人久久久久久| 插逼视频在线观看| 久久精品国产a三级三级三级| 国产亚洲精品久久久com| 男男h啪啪无遮挡| 久久影院123| 综合色av麻豆| 五月玫瑰六月丁香| 免费av毛片视频| 99热国产这里只有精品6| 美女主播在线视频| 在线观看一区二区三区| 菩萨蛮人人尽说江南好唐韦庄| 亚洲真实伦在线观看| 国产亚洲5aaaaa淫片| 欧美日韩精品成人综合77777| 黄片wwwwww| 精品熟女少妇av免费看| 日本爱情动作片www.在线观看| 在线观看国产h片| 各种免费的搞黄视频| 成人漫画全彩无遮挡| 婷婷色综合www| 精品人妻熟女av久视频| 啦啦啦中文免费视频观看日本| 亚洲国产最新在线播放| 成人二区视频| 男女无遮挡免费网站观看| 国产午夜精品一二区理论片| 久久鲁丝午夜福利片| 国产av码专区亚洲av| eeuss影院久久| av一本久久久久| 久久综合国产亚洲精品| 97热精品久久久久久| 久久久色成人| 少妇人妻精品综合一区二区| 少妇 在线观看| 狠狠精品人妻久久久久久综合| tube8黄色片| 老女人水多毛片| 国产精品一区二区三区四区免费观看| 欧美日韩一区二区视频在线观看视频在线 | 性色av一级| 黄色视频在线播放观看不卡| 99久久精品国产国产毛片| 亚洲成人一二三区av| 女的被弄到高潮叫床怎么办| 男女国产视频网站| 男人爽女人下面视频在线观看| 亚洲,一卡二卡三卡| 69av精品久久久久久| 免费观看的影片在线观看| 少妇 在线观看| 亚洲精品中文字幕在线视频 | av在线亚洲专区| 久久鲁丝午夜福利片| 欧美最新免费一区二区三区| av在线蜜桃| 身体一侧抽搐| 神马国产精品三级电影在线观看| 深夜a级毛片| 久久99热这里只有精品18| 黑人高潮一二区| 有码 亚洲区| 色视频在线一区二区三区| 午夜福利视频精品| 97热精品久久久久久| 国产黄色视频一区二区在线观看| 久热久热在线精品观看| 欧美日韩亚洲高清精品| 亚洲久久久久久中文字幕| 亚洲av成人精品一区久久| 免费人成在线观看视频色| 久久亚洲国产成人精品v| 亚洲欧美日韩东京热| 日本免费在线观看一区| 在线a可以看的网站| 成人毛片a级毛片在线播放| 男人舔奶头视频| 2022亚洲国产成人精品| 欧美日韩视频精品一区| eeuss影院久久| 99热网站在线观看| 久久久久久伊人网av| 2018国产大陆天天弄谢| 欧美另类一区| 天天躁日日操中文字幕| 久久97久久精品| 狂野欧美白嫩少妇大欣赏| 香蕉精品网在线| 国产乱人偷精品视频| 国产午夜精品一二区理论片| 久久久久国产网址| 精品国产露脸久久av麻豆| 国产乱人视频| 国产极品天堂在线| 男女下面进入的视频免费午夜| 中国三级夫妇交换| 舔av片在线| 91在线精品国自产拍蜜月| 一区二区三区免费毛片| 国产亚洲av嫩草精品影院| 亚洲欧美一区二区三区国产| 欧美 日韩 精品 国产| 别揉我奶头 嗯啊视频| 成人一区二区视频在线观看| 可以在线观看毛片的网站| 免费看不卡的av| 白带黄色成豆腐渣| 亚洲精品自拍成人| 精品人妻视频免费看| 少妇人妻久久综合中文| 亚洲最大成人av| 亚洲精品久久午夜乱码| 亚洲va在线va天堂va国产| 国产亚洲一区二区精品| 亚洲精品乱码久久久v下载方式| 免费黄频网站在线观看国产| 在线播放无遮挡| 国产av码专区亚洲av| 亚洲精品,欧美精品| 少妇人妻久久综合中文| 国产精品爽爽va在线观看网站| 亚洲国产精品成人综合色| 亚洲国产精品成人久久小说| 亚洲国产成人一精品久久久| av在线app专区| 亚洲人成网站在线观看播放| 亚洲av二区三区四区| 国产精品一区二区性色av| 亚洲精品亚洲一区二区| 九九久久精品国产亚洲av麻豆| 久久久久久久午夜电影| 能在线免费看毛片的网站| 国产中年淑女户外野战色| 亚洲国产成人一精品久久久| 亚洲在线观看片| 国产又色又爽无遮挡免| 网址你懂的国产日韩在线| 天堂网av新在线| 亚洲成人一二三区av| 高清毛片免费看| 久久亚洲国产成人精品v| 国产免费一级a男人的天堂| 亚洲成人精品中文字幕电影| 2018国产大陆天天弄谢| 日本一二三区视频观看| 久久精品夜色国产| 亚洲av成人精品一区久久| 亚洲色图av天堂| 男人舔奶头视频| 国产精品秋霞免费鲁丝片| 日韩强制内射视频| 国精品久久久久久国模美| 亚洲精品中文字幕在线视频 | 国产精品偷伦视频观看了| 一级片'在线观看视频| 啦啦啦中文免费视频观看日本| 大香蕉97超碰在线| 久热久热在线精品观看| 欧美潮喷喷水| 久久女婷五月综合色啪小说 | 成人漫画全彩无遮挡| 在现免费观看毛片| 婷婷色av中文字幕| 天堂网av新在线| 欧美激情国产日韩精品一区| 边亲边吃奶的免费视频| 又大又黄又爽视频免费| 偷拍熟女少妇极品色| 简卡轻食公司| 在线观看人妻少妇| 欧美最新免费一区二区三区| 亚洲自拍偷在线| 一级毛片久久久久久久久女| 在线 av 中文字幕| 国产在线一区二区三区精| 麻豆成人av视频| 午夜福利高清视频| 精品一区二区免费观看| 一级毛片电影观看| 大香蕉97超碰在线| 亚洲久久久久久中文字幕| 麻豆成人av视频| 欧美精品人与动牲交sv欧美| 菩萨蛮人人尽说江南好唐韦庄| 亚洲av一区综合| 久久久色成人| tube8黄色片| 亚洲成色77777| 精品国产乱码久久久久久小说| 日韩强制内射视频| 国产精品久久久久久精品电影小说 | 听说在线观看完整版免费高清| 久久国内精品自在自线图片| 欧美成人精品欧美一级黄| 午夜视频国产福利| 永久网站在线| av免费在线看不卡| 欧美日韩视频高清一区二区三区二| 色婷婷久久久亚洲欧美| 丝袜喷水一区| 亚洲色图av天堂| 欧美日韩精品成人综合77777| 青春草亚洲视频在线观看| 丝袜脚勾引网站| 日韩电影二区| 日本-黄色视频高清免费观看| 国产极品天堂在线| 国国产精品蜜臀av免费| 18禁裸乳无遮挡免费网站照片| 欧美成人午夜免费资源| 色5月婷婷丁香| 中文字幕av成人在线电影| 国产亚洲91精品色在线| 日韩免费高清中文字幕av| 精品亚洲乱码少妇综合久久| 亚洲国产日韩一区二区| 中文欧美无线码| 免费大片18禁| 青春草国产在线视频| 亚洲人成网站高清观看| 色网站视频免费| 黄片wwwwww| 欧美激情在线99| 嘟嘟电影网在线观看| 波多野结衣巨乳人妻| 日韩人妻高清精品专区| 国产一区亚洲一区在线观看| 久久午夜福利片| 18禁动态无遮挡网站| 国产精品99久久久久久久久| av网站免费在线观看视频| 国产成人福利小说| 91久久精品国产一区二区成人| 建设人人有责人人尽责人人享有的 | 啦啦啦在线观看免费高清www| 久久午夜福利片| 乱码一卡2卡4卡精品| 夫妻性生交免费视频一级片| 精华霜和精华液先用哪个| 日韩成人av中文字幕在线观看| 亚洲欧美中文字幕日韩二区| 男女那种视频在线观看| 九九在线视频观看精品| 国内揄拍国产精品人妻在线| 视频中文字幕在线观看| 国产精品一二三区在线看| 日日摸夜夜添夜夜添av毛片| 麻豆久久精品国产亚洲av| 亚洲一级一片aⅴ在线观看| 最后的刺客免费高清国语| 亚洲精品一区蜜桃| 人人妻人人澡人人爽人人夜夜| 久久综合国产亚洲精品| 日韩制服骚丝袜av| 日韩av免费高清视频| 亚洲最大成人手机在线| 国产精品麻豆人妻色哟哟久久| 免费观看av网站的网址| 国产精品久久久久久精品电影| 超碰av人人做人人爽久久| 美女高潮的动态| 伊人久久精品亚洲午夜| 国产真实伦视频高清在线观看| 欧美日韩视频高清一区二区三区二| 美女高潮的动态| 一本—道久久a久久精品蜜桃钙片 精品乱码久久久久久99久播 | 91在线精品国自产拍蜜月| 国产v大片淫在线免费观看| 岛国毛片在线播放| 免费观看a级毛片全部| 午夜亚洲福利在线播放| 国产v大片淫在线免费观看| 女人久久www免费人成看片| 国产一区二区在线观看日韩| 久久韩国三级中文字幕| 女的被弄到高潮叫床怎么办| 成人亚洲精品一区在线观看 | 久久久久久久久久久免费av| 日韩制服骚丝袜av| 在线观看免费高清a一片| 亚洲不卡免费看| 欧美区成人在线视频| 色吧在线观看| 51国产日韩欧美| 中文资源天堂在线| 国产亚洲5aaaaa淫片| 亚洲精品久久午夜乱码| 亚洲美女搞黄在线观看| 国产探花极品一区二区| 免费播放大片免费观看视频在线观看| 免费观看性生交大片5| 中文字幕制服av| 免费av不卡在线播放| 黑人高潮一二区| 97精品久久久久久久久久精品| 天天一区二区日本电影三级| 亚洲怡红院男人天堂| 欧美日韩国产mv在线观看视频 | 欧美一级a爱片免费观看看| 亚洲精品亚洲一区二区| 99九九线精品视频在线观看视频| 国产精品蜜桃在线观看| 99视频精品全部免费 在线| 亚洲综合色惰| 亚洲精品国产av成人精品| 国产在线男女| 久久99蜜桃精品久久| 日韩成人av中文字幕在线观看| 成年av动漫网址| 九九在线视频观看精品| 久久久国产一区二区| 香蕉精品网在线| 欧美日韩综合久久久久久| 久久久国产一区二区| 麻豆精品久久久久久蜜桃| 国产伦在线观看视频一区| 日韩伦理黄色片| 久久久久九九精品影院| 亚洲丝袜综合中文字幕| 国产真实伦视频高清在线观看| 欧美日韩国产mv在线观看视频 | 欧美变态另类bdsm刘玥| 黄片无遮挡物在线观看| 性插视频无遮挡在线免费观看| 欧美高清性xxxxhd video| 久久女婷五月综合色啪小说 | 丰满少妇做爰视频| 亚洲av成人精品一区久久| 亚洲精品日韩在线中文字幕| 久久人人爽av亚洲精品天堂 | 中文在线观看免费www的网站| 亚洲av日韩在线播放| 日韩强制内射视频| 国产欧美日韩一区二区三区在线 | 日本与韩国留学比较| 日本av手机在线免费观看| 日本三级黄在线观看| 大话2 男鬼变身卡| 欧美一区二区亚洲| 欧美高清性xxxxhd video| 亚洲av日韩在线播放| 国产精品爽爽va在线观看网站| 嘟嘟电影网在线观看| 熟女电影av网| 激情五月婷婷亚洲| 亚洲av不卡在线观看| 国产毛片在线视频| 少妇人妻久久综合中文| 乱系列少妇在线播放| 不卡视频在线观看欧美| 亚洲伊人久久精品综合| 国产亚洲精品久久久com| 人人妻人人爽人人添夜夜欢视频 | 国产淫片久久久久久久久| 特级一级黄色大片| 亚洲美女视频黄频| 日本一二三区视频观看| 亚洲欧美成人精品一区二区| 少妇丰满av| 毛片女人毛片| 尾随美女入室| 高清日韩中文字幕在线| 少妇 在线观看| 久久鲁丝午夜福利片| 国产亚洲91精品色在线| 亚洲成人一二三区av| 熟女电影av网| 久久ye,这里只有精品| 国产乱人偷精品视频| av网站免费在线观看视频| av在线播放精品| 国产 一区 欧美 日韩| 天天躁夜夜躁狠狠久久av| 69av精品久久久久久| 国产午夜精品一二区理论片| 精品国产三级普通话版| 乱系列少妇在线播放| 伊人久久精品亚洲午夜| 99久久中文字幕三级久久日本| 国产av国产精品国产| 午夜爱爱视频在线播放| 亚洲色图av天堂| a级毛片免费高清观看在线播放| 亚洲人成网站在线播| 天天躁夜夜躁狠狠久久av| 性插视频无遮挡在线免费观看| 亚洲av中文字字幕乱码综合| 又大又黄又爽视频免费| 成年人午夜在线观看视频| 人人妻人人爽人人添夜夜欢视频 | 亚洲精品日本国产第一区| 国产毛片在线视频| 国产精品一及| 99热这里只有是精品在线观看| 成年女人看的毛片在线观看| 青青草视频在线视频观看| 尤物成人国产欧美一区二区三区| 男的添女的下面高潮视频| 啦啦啦中文免费视频观看日本| 久久这里有精品视频免费| 婷婷色综合大香蕉| 下体分泌物呈黄色| 国产欧美亚洲国产| 欧美国产精品一级二级三级 | 超碰av人人做人人爽久久| 观看免费一级毛片| 亚洲天堂av无毛| 成年人午夜在线观看视频| 国产精品.久久久| a级毛片免费高清观看在线播放| 亚洲av在线观看美女高潮| 免费看a级黄色片| 亚洲最大成人手机在线| 人妻制服诱惑在线中文字幕| 男女边吃奶边做爰视频| 国产在线一区二区三区精| 久久99蜜桃精品久久| 热99国产精品久久久久久7| 免费av不卡在线播放| 久久久久久久精品精品| 大又大粗又爽又黄少妇毛片口| 久久精品国产自在天天线| 午夜激情福利司机影院| 日本免费在线观看一区| 日本av手机在线免费观看| 日韩视频在线欧美| 99久久精品热视频| 国产v大片淫在线免费观看| 在线观看av片永久免费下载| 能在线免费看毛片的网站| 亚洲国产日韩一区二区| 亚洲av男天堂| 99久久精品国产国产毛片| 久久精品国产亚洲网站| 99热这里只有精品一区| 国产乱人偷精品视频| 极品少妇高潮喷水抽搐| 直男gayav资源| 国产乱来视频区| 麻豆乱淫一区二区| 91精品一卡2卡3卡4卡| 国产男人的电影天堂91| 成人亚洲精品av一区二区| 在线天堂最新版资源| 国产高清国产精品国产三级 | 精品亚洲乱码少妇综合久久| 午夜福利在线在线| 日本欧美国产在线视频| 高清视频免费观看一区二区| 自拍欧美九色日韩亚洲蝌蚪91 | 亚洲综合色惰| 人妻系列 视频| 欧美日韩综合久久久久久| 午夜亚洲福利在线播放| 99久久精品一区二区三区| 九九久久精品国产亚洲av麻豆| 成人二区视频| 精品人妻偷拍中文字幕| 亚洲国产精品成人综合色| 人妻一区二区av| 亚洲国产欧美人成| 国产日韩欧美在线精品| 精品国产乱码久久久久久小说| 秋霞在线观看毛片| 三级国产精品欧美在线观看| 国产成人免费无遮挡视频| 国产毛片在线视频| 男女边吃奶边做爰视频| 真实男女啪啪啪动态图| 99热国产这里只有精品6| 亚洲欧美日韩另类电影网站 | 五月天丁香电影| 色网站视频免费| 色哟哟·www| 建设人人有责人人尽责人人享有的 | 性色avwww在线观看| 亚洲最大成人中文| 欧美性猛交╳xxx乱大交人| 成人综合一区亚洲| 亚洲av成人精品一二三区| 日本黄色片子视频| 日韩欧美 国产精品| 国语对白做爰xxxⅹ性视频网站| 国产精品一区二区三区四区免费观看| 99九九线精品视频在线观看视频| 午夜视频国产福利| 亚州av有码| 欧美成人午夜免费资源| 婷婷色麻豆天堂久久| 国产老妇伦熟女老妇高清| 成年版毛片免费区| 亚洲欧美日韩卡通动漫| 免费av毛片视频| 18禁裸乳无遮挡动漫免费视频 | 国国产精品蜜臀av免费| 国产大屁股一区二区在线视频| 99热国产这里只有精品6| 欧美丝袜亚洲另类| 日韩av不卡免费在线播放| 五月天丁香电影| 最近中文字幕2019免费版| 大片免费播放器 马上看| 网址你懂的国产日韩在线| 欧美激情在线99| 国产一区二区三区综合在线观看 | 国产av码专区亚洲av| 我要看日韩黄色一级片| 成人一区二区视频在线观看| 日韩三级伦理在线观看| 又爽又黄a免费视频| 少妇人妻久久综合中文| 熟妇人妻不卡中文字幕| 精品国产乱码久久久久久小说| 国产精品伦人一区二区| 2021天堂中文幕一二区在线观| 在线免费观看不下载黄p国产| 久久热精品热| 精品人妻熟女av久视频| 成人毛片60女人毛片免费| 一区二区三区乱码不卡18| 黄色日韩在线| 午夜福利视频精品| 成人亚洲精品一区在线观看 | 搡女人真爽免费视频火全软件| 国产精品久久久久久久久免| 亚洲天堂av无毛| 亚洲精品中文字幕在线视频 | 日韩免费高清中文字幕av| 色网站视频免费| 97热精品久久久久久| 欧美区成人在线视频| 一级毛片我不卡| 美女脱内裤让男人舔精品视频| 成人无遮挡网站| 欧美激情久久久久久爽电影| 国产高清有码在线观看视频| 亚洲无线观看免费| 深爱激情五月婷婷| 97超碰精品成人国产| 成人漫画全彩无遮挡| 亚洲性久久影院| 亚洲精品国产成人久久av| 成年女人在线观看亚洲视频 | 联通29元200g的流量卡| 久久人人爽人人爽人人片va| 黄色欧美视频在线观看| 日本猛色少妇xxxxx猛交久久| 可以在线观看毛片的网站| 欧美少妇被猛烈插入视频| 亚洲精华国产精华液的使用体验| 久久久国产一区二区| 久热这里只有精品99| 97在线视频观看| 另类亚洲欧美激情| 国产精品精品国产色婷婷| 久久久久久久久久成人| 在线观看国产h片| 国产精品不卡视频一区二区| av在线天堂中文字幕| 一个人看视频在线观看www免费| 一区二区三区四区激情视频| 色综合色国产| 最后的刺客免费高清国语| 成年女人看的毛片在线观看| 麻豆成人午夜福利视频| 久久久亚洲精品成人影院| 男人爽女人下面视频在线观看| 男女边摸边吃奶| 一级黄片播放器| 深夜a级毛片| 九九爱精品视频在线观看| av免费在线看不卡| 成年女人在线观看亚洲视频 | 欧美高清成人免费视频www| 观看美女的网站| 日本熟妇午夜| 一区二区三区免费毛片| 国产精品久久久久久久电影| 国产亚洲精品久久久com| 1000部很黄的大片| 97精品久久久久久久久久精品| 日本与韩国留学比较| 欧美成人午夜免费资源| 亚洲精品乱久久久久久| 午夜视频国产福利| 日韩成人av中文字幕在线观看| 日本熟妇午夜| 丰满少妇做爰视频| 欧美日韩一区二区视频在线观看视频在线 | 欧美日韩国产mv在线观看视频 | 久久99精品国语久久久| 亚洲欧美日韩另类电影网站 | 国内精品美女久久久久久| 免费黄色在线免费观看| 久久99精品国语久久久| 成人亚洲精品av一区二区| 久久久久久久久大av| 亚洲av中文字字幕乱码综合| 黄色怎么调成土黄色| 久久女婷五月综合色啪小说 | 青春草视频在线免费观看| 精品久久国产蜜桃| 精品少妇黑人巨大在线播放| 天天躁日日操中文字幕| 国产成人精品久久久久久| a级毛色黄片| 搡女人真爽免费视频火全软件| 免费黄频网站在线观看国产| 久久久久久久午夜电影| 在线精品无人区一区二区三 | 性色avwww在线观看| 中文天堂在线官网|