AutoPhrase: Automated Phrase Mining from Massive Text Corpora

Publications

Please cite the following two papers if you are using our tools. Thanks!

• Jingbo Shang, Jialu Liu, Meng Jiang, Xiang Ren, Clare R Voss, Jiawei Han, "Automated Phrase Mining from Massive Text Corpora", accepted by IEEE Transactions on Knowledge and Data Engineering, Feb. 2018.

• Jialu Liu*, Jingbo Shang*, Chi Wang, Xiang Ren and Jiawei Han, "Mining Quality Phrases from Massive Text Corpora”, Proc. of 2015 ACM SIGMOD Int. Conf. on Management of Data (SIGMOD'15), Melbourne, Australia, May 2015. (* equally contributed, slides)

Recent Changes

2020.06.14

• Updates docker image with the git master.

2018.03.04

• Fix a few bugs during the pre-processing and post-processing, i.e., Tokeninzer.java. Previously, when the corpus contains characters like /, the results could be wrong or errors may occur.
• When the phrasal segmentation is serving new text, for the phrases (every token is seen in the traning corpus) provided in the knowledge base (wiki_quality.txt), the score is set as 1.0. Previously, it was kind of infinite.

2017.10.23

• Support extremely large corpus (e.g., 4GB or more). Please comment out the // define LARGE in the beginning of src/utils/parameters.h before you run AutoPhrase on such a large corpus.
• Quality phrases (every token is seen in the raw corpus) provided in the knowledge base will be incorporated during the phrasal segmentation, even their frequencies are smaller than MIN_SUP.
• Stopwords will be treated as low quality single-word phrases.
• Model files are saved separately. Please check the variable MODEL in both auto_phrase.sh and phrasal_segmentation.sh.
• The end of line is also a separator for sentence splitting.

New Features

(compared to SegPhrase)

• Minimized Human Effort. We develop a robust positive-only distant training method to estimate the phrase quality by leveraging exsiting general knowledge bases.
• Support Multiple Languages: English, Spanish, and Chinese. The language in the input will be automatically detected.
• High Accuracy. We propose a POS-guided phrasal segmentation model incorporating POS tags when POS tagger is available. Meanwhile, the new framework is able to extract single-word quality phrases.
• High Efficiency. A better indexing and an almost lock-free parallelization are implemented, which lead to both running time speedup and memory saving.

Related GitHub Repositories

• SegPhrase
• SegPhrase-MultiLingual

Requirements

Linux or MacOS with g++ and Java installed.

Ubuntu:

• g++ 4.8 $ sudo apt-get install g++-4.8
• Java 8 $ sudo apt-get install openjdk-8-jdk
• curl $ sudo apt-get install curl

MacOS:

• g++ 6 $ brew install gcc6
• Java 8 $ brew update; brew tap caskroom/cask; brew install Caskroom/cask/java

Default Run

Phrase Mining Step

$ ./auto_phrase.sh

The default run will download an English corpus from the server of our data mining group and run AutoPhrase to get 3 ranked lists of phrases as well as 2 segmentation model files under the MODEL (i.e., models/DBLP) directory.

• AutoPhrase.txt: the unified ranked list for both single-word phrases and multi-word phrases.
• AutoPhrase_multi-words.txt: the sub-ranked list for multi-word phrases only.
• AutoPhrase_single-word.txt: the sub-ranked list for single-word phrases only.
• segmentation.model: AutoPhrase's segmentation model (saved for later use).
• token_mapping.txt: the token mapping file for the tokenizer (saved for later use).

You can change RAW_TRAIN to your own corpus and you may also want change MODEL to a different name.

Phrasal Segmentation

We also provide an auxiliary function to highlight the phrases in context based on our phrasal segmentation model. There are two thresholds you can tune in the top of the script. The model can also handle unknown tokens (i.e., tokens which are not occurred in the phrase mining step's corpus).

In the beginning, you need to specify AutoPhrase's segmentation model, i.e., MODEL. The default value is set to be consistent with auto_phrase.sh.

$ ./phrasal_segmentation.sh

The segmentation results will be put under the MODEL directory as well (i.e., model/DBLP/segmentation.txt). The highlighted phrases will be enclosed by the phrase tags (e.g., <phrase>data mining</phrase>).

Incorporate Domain-Specific Knowledge Bases

If domain-specific knowledge bases are available, such as MeSH terms, there are two ways to incorporate them.

• (recommended) Append your known quality phrases to the file data/EN/wiki_quality.txt.
• Replace the file data/EN/wiki_quality.txt by your known quality phrases.

Handle Other Languages

Tokenizer and POS tagger

In fact, our tokenizer supports many different languages, including Arabics (AR), German (DE), English (EN), Spanish (ES), French (FR), Italian (IT), Japanese (JA), Portuguese (PT), Russian (RU), and Chinese (CN). If the language detection is wrong, you can also manually specify the language by modify the TOKENIZER command in the bash script auto_phrase.sh using the two-letter code for that language. For example, the following one forces the language to be English.

TOKENIZER="-cp .:tools/tokenizer/lib/*:tools/tokenizer/resources/:tools/tokenizer/build/ Tokenizer -l EN"

We also provide a default tokenizer together with a dummy POS tagger in the tools/tokenizer. It uses the StandardTokenizer in Lucene, and always assign a tag UNKNOWN to each token. To enable this feature, please add the -l OTHER" to the TOKENIZER command in the bash script auto_phrase.sh.

TOKENIZER="-cp .:tools/tokenizer/lib/*:tools/tokenizer/resources/:tools/tokenizer/build/ Tokenizer -l OTHER"

If you want to incorporate your own tokenizer and/or POS tagger, please create a new class extending SpecialTagger in the tools/tokenizer. You may refer to StandardTagger as an example.

stopwords.txt

You may try to search online or create your own list.

wiki_all.txt and wiki_quality.txt

Meanwhile, you have to add two lists of quality phrases in the data/OTHER/wiki_quality.txt and data/OTHER/wiki_all.txt. The quality of phrases in wiki_quality should be very confident, while wiki_all, as its superset, could be a little noisy. For more details, please refer to the tools/wiki_enities.

Use an already tokenized/preprocessed and POS tagged corpus

You can also use AutoPhrase with an already tokenized and tagged corpus. For this, you need to:

• Set POS_TAGGING_MODE=${POS_TAGGING_MODE:- 2} in both auto_phrase.sh and phrasal_segmentation.sh scripts
• Place a pos_tags.txt file inside your data directory (eg. data/EN/pos_tags.txt)
• Assure that the count of tags in pos_tags.txt is equal to the count of tokens in dataset.txt.
• Separate your dataset.txt (input file) tokens using one-char delimiters. Set the delimiters in both auto_phrase.sh and phrasal_segmentation.sh scripts (search for -delimiters).
  • Eg.: If \n, \t and WhiteSpace are used as delimiters, set:

auto_phrase.sh:
time java $TOKENIZER -m train -i $RAW_TRAIN -o $TOKENIZED_TRAIN -t $TOKEN_MAPPING -c N -thread $THREAD -delimiters "\n\t "

phrasal_segmentation.sh
time java $TOKENIZER -m direct_test -i $TEXT_TO_SEG -o $TOKENIZED_TEXT_TO_SEG -t $TOKEN_MAPPING -c N -thread $THREAD -delimiters "\n\t "

Note also that, by using such custom input, you can lemmatize or stemm your tokens beforehand and keep the already computed POS tags unchanged.

Docker

Default Run

sudo docker run -v $PWD/models:/autophrase/models -it \
    -e ENABLE_POS_TAGGING=1 \
    -e MIN_SUP=30 -e THREAD=10 \
    remenberl/autophrase

./auto_phrase.sh

The results will be available in the models folder. Note that all of the environment variables above have their default values--leaving the assigments out here would produce exactly the same results. (However, in this case, using default values, the results of phrasal_segmentation.txt would be saved to the internal default_models directory--this is unavoidable, since the phrasal segmentation app reads from and writes to the same model directory.)

User Specified Input

Assuming the path to input file is ./data/input.txt.

sudo docker run -v $PWD/data:/autophrase/data -v $PWD/models:/autophrase/models -it \
    -e RAW_TRAIN=data/input.txt \
    -e ENABLE_POS_TAGGING=1 \
    -e MIN_SUP=30 -e THREAD=10 \
    -e MODEL=models/MyModel \
    -e TEXT_TO_SEG=data/input.txt \
    remenberl/autophrase

./auto_phrase.sh

"RAW_TRAIN" is the training corpus, and "TEXT_TO_SEG" is a corpus whose phrases are to be highlighted--typically, this is the same corpus, but training and phrasal segmentation use two different scripts. When the user wants to segment a new corpus with an existing model, only the latter script need be used (and setting "RAW_TRAIN" isn't necessary).

Note that, in a Docker deployment, the (default) data and models directories are renamed to default_data and default_models, respectively, to avoid conflicts with mounted external directories with the same names. It should be noted as well that there's litle point in saving a model to the default models directory, since all new files are erased when the container is exited (and if an external directory is mounted as "models", and no value is specified for "MODEL", the results will be saved in the "models/DBP" subdirectory). The same wrinkle also means that there's little point to running a container with the "FIRST_RUN" variable set to 0.

Because the original data directory will have been been renamed, it's perfectly fine for the user to mount an external directory called "data" and read the corpus from there--and in most cases, there's no need for a user to change the supplied files stored in the default data directory. If such a change is necessary, though, the environment variable that specifies the directory in question is "DATA_DIR".

In Windows

The sudo command won't work in a Windows bash shell, and in any case isn't needed in an elevated window--replace it with winpty.

In addition, the PWD variable works a little oddly in MinGW (the Git bash shell), appending ";C" to the end of the path. To prevent this, replace $PWD/models:/autophrase/models with "/${PWD}/models":/autophrase/models, and $PWD/data/autophrase/data with "/${PWD}/data:/autophrase/data.