fsauor2018

Code for Fine-grained Sentiment Analysis of User Reviews of AI Challenger 2018.

Single model can achieve 0.71 marco-f1 score.

Testa rank: 27

Testb rank: 16

The final result is achieved by ensemble 10 models by simple voting.

Issues and starts are welcomed!

Train from scratch

For those who don't want to preprocess data, refer to scripts.

Data

For those who want to get the raw dataset, please refer to this link data.

Requirements

tensorflow == 1.4.1

Model Architecture

The model architecture is simple. Basiclly, you can think of it as a seq2seq model.

Some details of the model:

• Embedding layer + 3 * Bi-LSTM layers as encoder
• Residual connection is added on the second and third Bi-LSTM layers
• The final encoder outputs are weighted sum of outputs of each layer. Scalars and weight are learned variables. This idea is copied from ELMO.
• Use a simple LSTM cell + Attention as decoder, decode 20 steps to get 20 outputs for each label
  • Inputs to decoder are learnable embeddings
• Outputs of decoder are fed to two FC layers to get the final sentiment logits

Data preprocess

The data preprocess code is not provided here, I may release it later.

To use this project, you need fowllowing files:

• train.json / validataion.json / testa.json
• vocab.txt
• embedding.txt
• label.txt

Training files

You need to preprocess the orginal data to json files, each line of the json line should be like fowllowing:

{"id": "0", "content": "吼吼吼 ， 萌 死 人 的 棒棒糖 ， 中 了 大众 点评 的 霸王餐 ， 太 可爱 了 。 一直 就 好奇 这个 棒棒 糖 是 怎么 个 东西 ， 大众 点评 给 了 我 这个 土老 冒 一个 见识 的 机会 。 看 介绍 棒棒 糖 是 用 <place> 糖 做 的 ， 不 会 很 甜 ， 中间 的 照片 是 糯米 的 ， 能 食用 ， 真是 太 高端 大气 上档次 了 ， 还 可以 买 蝴蝶 结扎口 ， 送 人 可以 买 礼盒 。 我 是 先 打 的 卖家 电话 ， 加 了 微信 ， 给 卖家传 的 照片 。 等 了 几 天 ， 卖家 就 告诉 我 可以 取 货 了 ， 去 <place> 那 取 的 。 虽然 连 卖家 的 面 都 没 见到 ， 但是 还是 谢谢 卖家 送 我 这么 可爱 的 东西 ， 太 喜欢 了 ， 这 哪 舍得 吃 啊 。", "location_traffic_convenience": "-2", "location_distance_from_business_district": "-2", "location_easy_to_find": "-2", "service_wait_time": "-2", "service_waiters_attitude": "1", "service_parking_convenience": "-2", "service_serving_speed": "-2", "price_level": "-2", "price_cost_effective": "-2", "price_discount": "1", "environment_decoration": "-2", "environment_noise": "-2", "environment_space": "-2", "environment_cleaness": "-2", "dish_portion": "-2", "dish_taste": "-2", "dish_look": "1", "dish_recommendation": "-2", "others_overall_experience": "1", "others_willing_to_consume_again": "-2"}

To be specific:

• content should be tokeninzed words
  • You can use jieba/ltp to do the segmentation
  • Use NER toolkits to replace place and orginaztion to special tokens '<place>','<org>'
• other fields are same as the original data files
• for test files, which labels are unknow, you can leave them to be empty string("")

Vocab file

I choose the top 50k most common words in training file.

The top 3 words are special tokens, which are:

• <unk>: unknow token
• <sos>: start of content
• <eos>: end of content, also used as padding token

Embedding file

This is a glove-format embedding file, I use Chinese-Word-Vectors as pretrained embedding file(which is Sogou News word2vec word embedding).

Label file

All the label names.

Train

Refer to bash/elmo_train.sh

Inference

Refer to bash/elmo_inference.sh