--- license: mit --- # UCTopic This repository contains the code of model UCTopic and an easy-to-use tool UCTopicTool used for Topic Mining, Unsupervised Aspect Extractioin or Phrase Retrieval. Our ACL 2022 paper [UCTopic: Unsupervised Contrastive Learning for Phrase Representations and Topic Mining](https://arxiv.org/abs/2202.13469). # Quick Links - [Overview](#overview) - [Pretrained Model](#pretrained-model) - [Getting Started](#getting-started) - [UCTopic Model](#uctopic-model) - [UCTopicTool](#uctopictool) - [Experiments in Paper](#experiments) - [Requirements](#requirements) - [Datasets](#datasets) - [Entity Clustering](#entity-clustering) - [Topic Mining](#topic-mining) - [Pretraining](#pretraining) - [Contact](#contact) - [Citation](#citation) # Overview We propose UCTopic, a novel unsupervised contrastive learning framework for context-aware phrase representations and topic mining. UCTopic is pretrained in a large scale to distinguish if the contexts of two phrase mentions have the same semantics. The key to pretraining is positive pair construction from our phrase-oriented assumptions. However, we find traditional in-batch negatives cause performance decay when finetuning on a dataset with small topic numbers. Hence, we propose cluster-assisted contrastive learning(CCL) which largely reduces noisy negatives by selecting negatives from clusters and further improves phrase representations for topics accordingly. # Pretrained Model Our released model: | Model | Note| |:-------------------------------|------| |[uctopic-base](https://drive.google.com/file/d/1XQzi4E9ctdI373CK5O-pXQyBvOONssp1/view?usp=sharing)| Pretrained UCTopic model based on [LUKE-BASE](https://arxiv.org/abs/2010.01057) Unzip to get `uctopic-base` folder. # Getting Started We provide an easy-to-use phrase representation tool based on our UCTopic model. To use the tool, first install the uctopic package from PyPI ```bash pip install uctopic ``` Or directly install it from our code ```bash python setup.py install ``` ## UCTopic Model After installing the package, you can load our model by just two lines of code ```python from uctopic import UCTopic model = UCTopic.from_pretrained('JiachengLi/uctopic-base') ``` The model will automatically download pre-trained parameters from [HuggingFace's models](https://huggingface.co/models). If you encounter any problem when directly loading the models by HuggingFace's API, you can also download the models manually from the above table and use `model = UCTopic.from_pretrained({PATH TO THE DOWNLOAD MODEL})`. To get pre-trained phrase representations, our model inputs are same as [LUKE](https://huggingface.co/docs/transformers/model_doc/luke). Note: please input only ONE span each time, otherwise, will have performance decay according to our empirical results. ```python from uctopic import UCTopicTokenizer, UCTopic tokenizer = UCTopicTokenizer.from_pretrained('JiachengLi/uctopic-base') model = UCTopic.from_pretrained('JiachengLi/uctopic-base') text = "Beyoncé lives in Los Angeles." entity_spans = [(17, 28)] # character-based entity span corresponding to "Los Angeles" inputs = tokenizer(text, entity_spans=entity_spans, add_prefix_space=True, return_tensors="pt") outputs, phrase_repr = model(**inputs) ``` `phrase_repr` is the phrase embedding (size `[768]`) of the phrase `Los Angeles`. `outputs` has the same format as the outputs from `LUKE`. ## UCTopicTool We provide a tool `UCTopicTool` built on `UCTopic` for efficient phrase encoding, topic mining (or unsupervised aspect extraction) or phrase retrieval. ### Initialization `UCTopicTool` is initialized by giving the `model_name_or_path` and `device`. ```python from uctopic import UCTopicTool topic_tool = UCTopicTool('JiachengLi/uctopic-base', device='cuda:0') ``` ### Phrase Encoding Phrases are encoded by our method `UCTopicTool.encode` in batches, which is more efficient than `UCTopic`. ```python phrases = [["This place is so much bigger than others!", (0, 10)], ["It was totally packed and loud.", (15, 21)], ["Service was on the slower side.", (0, 7)], ["I ordered 2 mojitos: 1 lime and 1 mango.", (12, 19)], ["The ingredient weren't really fresh.", (4, 14)]] embeddings = topic_tool.encode(phrases) # len(embeddings) is equal to len(phrases) ``` **Note**: Each instance in `phrases` contains only one sentence and one span (character-level position) in format `[sentence, span]`. Arguments for `UCTopicTool.encode` are as follows, * **phrase** (List) - A list of `[sentence, span]` to be encoded. * **return_numpy** (bool, *optional*, defaults to `False`) - Return `numpy.array` or `torch.Tensor`. * **normalize_to_unit** (bool, *optional*, defaults to `True`) - Normalize all embeddings to unit vectors. * **keepdim** (bool, *optional*, defaults to `True`) - Keep dimension size `[instance_number, hidden_size]`. * **batch_size** (int, *optional*, defaults to `64`) - The size of mini-batch in the model. ### Topic Mining and Unsupervised Aspect Extraction The method `UCTopicTool.topic_mining` can mine topical phrases or conduct aspect extraction from sentences with or without spans. ```python sentences = ["This place is so much bigger than others!", "It was totally packed and loud.", "Service was on the slower side.", "I ordered 2 mojitos: 1 lime and 1 mango.", "The ingredient weren't really fresh."] spans = [[(0, 10)], # This place [(15, 21), (26, 30)], # packed; loud [(0, 7)], # Service [(12, 19), (21, 27), (32, 39)], # mojitos; 1 lime; 1 mango [(4, 14)]] # ingredient # len(sentences) is equal to len(spans) output_data, topic_phrase_dict = tool.topic_mining(sentences, spans, \ n_clusters=[15, 25]) # predict topic for new phrases phrases = [["The food here is amazing!", (4, 8)], ["Lovely ambiance with live music!", (21, 31)]] topics = tool.predict_topic(phrases) ``` **Note**: If `spans` is not given, `UCTopicTool` will extract noun phrases by [spaCy](https://spacy.io/). Arguments for `UCTopicTool.topic_mining` are as follows, Data arguments: * **sentences** (List) - A List of sentences for topic mining. * **spans** (List, *optional*, defaults to `None`) - A list of span list corresponding sentences, e.g., `[[(0, 9), (5, 7)], [(1, 2)]]` and `len(sentences)==len(spans)`. If None, automatically mine phrases from noun chunks. Clustering arguments: * **n_clusters** (int or List, *optional*, defaults to `2`) - The number of topics. When `n_clusters` is a list, `n_clusters[0]` and `n_clusters[1]` will be the minimum and maximum numbers to search, `n_clusters[2]` is the search step length (if not provided, default to 1). * **meric** (str, *optional*, defaults to `"cosine"`) - The metric to measure the distance between vectors. `"cosine"` or `"euclidean"`. * **batch_size** (int, *optional*, defaults to `64`) - The size of mini-batch for phrase encoding. * **max_iter** (int, *optional*, defaults to `300`) - The maximum iteration number of kmeans. CCL-finetune arguments: * **ccl_finetune** (bool, *optional*, defaults to `True`) - Whether to conduct CCL-finetuning in the paper. * **batch_size_finetune** (int, *optional*, defaults to `8`) - The size of mini-batch for finetuning. * **max_finetune_num** (int, *optional*, defaults to `100000`) - The maximum number of training instances for finetuning. * **finetune_step** (int, *optional*, defaults to `2000`) - The number of training steps for finetuning. * **contrastive_num** (int, *optional*, defaults to `5`) - The number of negatives in contrastive learning. * **positive_ratio** (float, *optional*, defaults to `0.1`) - The ratio of the most confident instances for finetuning. * **n_sampling** (int, *optional*, defaults to `10000`) - The number of sampled examples for cluster number confirmation and finetuning. Set to `-1` to use the whole dataset. * **n_workers** (int, *optional*, defaults to `8`) - The number of workers for preprocessing data. Returns for `UCTopicTool.topic_mining` are as follows, * **output_data** (List) - A list of sentences and corresponding phrases and topic numbers. Each element is `[sentence, [[start1, end1, topic1], [start2, end2, topic2]]]`. * **topic_phrase_dict** (Dict) - A dictionary of topics and the list of phrases under a topic. The phrases are sorted by their confidence scores. E.g., `{topic: [[phrase1, score1], [phrase2, score2]]}`. The method `UCTopicTool.predict_topic` predicts the topic ids for new phrases based on your training results from `UCTopicTool.topic_mining`. The inputs of `UCTopicTool.predict_topic` are same as `UCTopicTool.encode` and returns a list of topic ids (int). ### Phrase Similarities and Retrieval The method `UCTopicTool.similarity` compute the cosine similarities between two groups of phrases: ```python phrases_a = [["This place is so much bigger than others!", (0, 10)], ["It was totally packed and loud.", (15, 21)]] phrases_b = [["Service was on the slower side.", (0, 7)], ["I ordered 2 mojitos: 1 lime and 1 mango.", (12, 19)], ["The ingredient weren't really fresh.", (4, 14)]] similarities = tool.similarity(phrases_a, phrases_b) ``` Arguments for `UCTopicTool.similarity` are as follows, * **queries** (List) - A list of `[sentence, span]` as queries. * **keys** (List or `numpy.array`) - A list of `[sentence, span]` as keys or phrase representations (`numpy.array`) from `UCTopicTool.encode`. * **batch_size** (int, *optional*, defaults to `64`) - The size of mini-batch in the model. `UCTopicTool.similarity` returns a `numpy.array` contains the similarities between phrase pairs in two groups. The methods `UCTopicTool.build_index` and `UCTopicTool.search` are used for phrase retrieval: ```python phrases = [["This place is so much bigger than others!", (0, 10)], ["It was totally packed and loud.", (15, 21)], ["Service was on the slower side.", (0, 7)], ["I ordered 2 mojitos: 1 lime and 1 mango.", (12, 19)], ["The ingredient weren't really fresh.", (4, 14)]] # query multiple phrases query1 = [["The food here is amazing!", (4, 8)], ["Lovely ambiance with live music!", (21, 31)]] # query single phrases query2 = ["The food here is amazing!", (4, 8)] tool.build_index(phrases) results = tool.search(query1, top_k=3) # or results = tool.search(query2, top_k=3) ``` We also support [faiss](https://github.com/facebookresearch/faiss), an efficient similarity search library. Just install the package following [instructions](https://github.com/facebookresearch/faiss/blob/main/INSTALL.md) here and `UCTopicTool` will automatically use `faiss` for efficient search. `UCTopicTool.search` returns the ranked top k phrases for each query. ### Save and Load finetuned UCTopicTool The methods `UCTopicTool.save` and `UCTopicTool.load` are used for save and load all paramters of `UCTopicTool`. Save: ```python tool = UCTopicTool('JiachengLi/uctopic-base', 'cuda:0') # finetune UCTopic with CCL output_data, topic_phrase_dict = tool.topic_mining(sentences, spans, \ n_clusters=[15, 25]) tool.save(**your directory**) ``` Load: ```python tool = UCTopicTool('JiachengLi/uctopic-base', 'cuda:0') tool.load(**your directory**) ``` The loaded parameters will be used for all methods (for encoding, topic mining, phrase similarities and retrieval) introduced above. # Experiments In this section, we re-implement experiments in our paper. ## Requirements First, install PyTorch by following the instructions from [the official website](https://pytorch.org). To faithfully reproduce our results, please use the correct `1.9.0` version corresponding to your platforms/CUDA versions. Then run the following script to install the remaining dependencies, ```bash pip install -r requirements.txt ``` Download `en_core_web_sm` model from spacy, ```bash python -m spacy download en_core_web_sm ``` ## Datasets The downstream datasets used in our experiments can be downloaded from [here](https://drive.google.com/file/d/1dVIp9li1Wdh0JgU8slsWm0ObcitbQtSL/view?usp=sharing). ## Entity Clustering The config file of entity clustering is `clustering/consts.py` and most arguments are self-explained. Please setup `--gpu` and `--data_path` before running. The clustering scores will be printed. Clustering with our pre-trained phrase embeddings. ```bash python clustering.py --gpu 0 ``` Clustering with our pre-trained phrase embeddings and Cluster-Assisted Constrastive Learning (CCL) proposed in our paper. ```bash python clustering_ccl_finetune.py --gpu 0 ``` ## Topic Mining The config file of entity clustering is `topic_modeling/consts.py`. **Key Argument Table** | Arguments | Description | |:-----------------|:-----------:| | --num_classes |**Min** and **Max** number of classes, e.g., `[5, 15]`. Our model will find the class number by [silhouette_score](https://scikit-learn.org/stable/modules/generated/sklearn.metrics.silhouette_score.html).| | --sample_num_cluster |Number of sampled phrases to confirm class number.| | --sample_num_finetune|Number of sampled phrases for CCL finetuning.| | --contrastive_num|Number of negative classes for CCL finetuning.| | --finetune_step | CCL finetuning steps (maximum global steps for finetuning).| **Tips**: Please tune `--batch_size` or `--contrastive_num` for suitable GPU memory usage. Topic mining with our pre-trained phrase embeddings and Cluster-Assisted Constrastive Learning (CCL) proposed in our paper. ```bash python find_topic.py --gpu 0 ``` **Outputs** We output three files under `topic_results`: | File Name | Description | |:-----------------|:-----------:| | `merged_phraes_pred_prob.pickle` |A dictionary of phrases and their topic number and prediction probability. A topic of a phrase is merged from all phrase mentioins. `{phrase: [topic_id, probability]}`, e.g., {'fair prices': [0, 0.34889686]}| | `phrase_instances_pred.json`| A list of all mined phrase mentions. Each element is `[[doc_id, start, end, phrase_mention], topic_id]`.| | `topics_phrases.json`|A dictionary of topics and corresponding phrases sorted by probability. `{'topic_id': [[phrase1, prob1], [phrase2, prob2]]}`| ### Pretraining **Data** For unsupervised pretraining of UCTopic, we use article and span with links from English Wikipedia and Wikidata. Our processed dataset can be downloaded from [here](https://drive.google.com/file/d/1wflsmhPI9J0ZA6aVRl2mQjHIE6JIvzAv/view?usp=sharing). **Training scripts** We provide example training scripts and our default training parameters for unsupervised training of UCTopic in `run_example.sh`. ```bash bash run_example.sh ``` Arguments description can be found in `pretrain.py`. All the other arguments are standard Huggingface's `transformers` training arguments. **Convert models** Our pretrained checkpoints are slightly different from the checkpoint `uctopic-base`. Please refer `convert_uctopic_parameters.py` to convert it. # Contact If you have any questions related to the code or the paper, feel free to email Jiacheng (`j9li@eng.ucsd.edu`). If you encounter any problems when using the code, or want to report a bug, you can open an issue. Please try to specify the problem with details so we can help you better and quicker! # Citation Please cite our paper if you use UCTopic in your work: ```bibtex @inproceedings{Li2022UCTopicUC, title = "{UCT}opic: Unsupervised Contrastive Learning for Phrase Representations and Topic Mining", author = "Li, Jiacheng and Shang, Jingbo and McAuley, Julian", booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)", month = may, year = "2022", address = "Dublin, Ireland", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2022.acl-long.426", doi = "10.18653/v1/2022.acl-long.426", pages = "6159--6169" } ```