adaptation_specialization_material/code/etm.py

from AbstractSemantics.terms import TermExtractor
from AbstractSemantics.embeddings import GensimWord2Vec
from gensim.models import KeyedVectors

import nltk

import pandas as pd
import numpy as np

from scipy.sparse import csr_matrix

from os.path import join as opj
from os.path import exists

import itertools
from functools import partial
from collections import defaultdict

import re

import multiprocessing as mp

# from matplotlib import pyplot as plt

import argparse
import yaml
import sys
import pickle

from gensim.models.callbacks import CallbackAny2Vec


class MonitorCallback(CallbackAny2Vec):
    def __init__(self, test_words):
        self._test_words = test_words
        self.epoch = 0

    def on_epoch_end(self, model):
        loss = model.get_latest_training_loss()
        if self.epoch == 0:
            print("Loss after epoch {}: {}".format(self.epoch, loss))
        else:
            print(
                "Loss after epoch {}: {}".format(
                    self.epoch, loss - self.loss_previous_step
                )
            )

        self.epoch += 1
        self.loss_previous_step = loss

        for word in self._test_words:  # show wv logic changes
            print(f"{word}: {model.wv.most_similar(word)}")


def filter_ngrams(l, wl):
    return [ngram for ngram in l if ngram in wl]


def construct_bow(l, n):
    items = list(set(l))
    return np.array(items), np.array([l.count(i) for i in items])


def ngram_inclusion(i, js):
    return [
        j.find(i) >= 0  # matching
        and bool(re.search(f"(^|\_){re.escape(i)}($|\_)", j))
        and (j.count("_") == i.count("_") + 1)
        and (
            (i.count("_") >= 1)
            or bool(re.search(f"(^|\_){re.escape(i)}$", j))
            or bool(re.search(f"^{re.escape(i)}($|\_)", j))
        )
        for j in js
    ]


if __name__ == "__main__":

    parser = argparse.ArgumentParser("CT Model")
    parser.add_argument("location", help="model directory")
    parser.add_argument(
        "filter", choices=["categories", "keywords", "no-filter"], help="filter type"
    )
    parser.add_argument("--values", nargs="+", default=[], help="filter allowed values")

    parser.add_argument("--dataset", default="inspire-harvest/database")


    # sample size
    parser.add_argument("--samples", type=int, default=50000)
    parser.add_argument("--constant-sampling", type=int, default=0)

    # text pre-processing
    parser.add_argument(
        "--add-title", default=False, action="store_true", help="include title"
    )
    parser.add_argument(
        "--remove-latex", default=False, action="store_true", help="remove latex"
    )
    parser.add_argument(
        "--lemmatize", default=False, action="store_true", help="lemmatize"
    )
    parser.add_argument(
        "--limit-redundancy",
        default=False,
        action="store_true",
        help="limit redundancy",
    )
    parser.add_argument("--blacklist", default=None, help="blacklist")

    # embeddings
    parser.add_argument("--dimensions", type=int, default=50)
    parser.add_argument("--pre-trained-embeddings", default=False, action="store_true")
    parser.add_argument("--use-saved-embeddings", default=False, action="store_true")

    # topic model parameters
    parser.add_argument("--topics", type=int, default=25)
    parser.add_argument("--min-df", type=float, default=0.001)
    parser.add_argument("--max-df", type=float, default=0.15)

    parser.add_argument("--threads", type=int, default=4)

    args = parser.parse_args(
        [
            "output/acl_2002_2022",
            "no-filter",
            "--dataset",
            "../acl",
            "--constant-sampling",
            "12000",
            "--samples",
            "300000",
            "--threads",
            "30",
            "--add-title",
            "--remove-latex",
            "--dimensions",
            "50",
            "--topics",
            "20",
            "--min-df",
            "0.00075",
            "--lemmatize",
            "--pre-trained-embeddings",
            # "--limit-redundancy"
            "--use-saved-embeddings"
            # "--blacklist",
            # "output/medialab/blacklist",
        ]
    )

    with open(opj(args.location, "params.yml"), "w+") as fp:
        yaml.dump(args, fp)

    articles = pd.read_parquet(
        opj(args.dataset, "articles.parquet")
    )[["title", "abstract", "article_id", "date_created", "categories"]]

    if args.add_title:
        articles["abstract"] = articles["abstract"].str.cat(articles["title"], sep=". ")

    articles.drop(columns=["title"], inplace=True)

    if args.remove_latex:
        articles["abstract"] = articles["abstract"].apply(
            lambda s: re.sub("$[^>]+$", "", s)
        )

    articles["abstract"] = articles["abstract"].apply(
        lambda s: re.sub(r"\b\\\w+", "", s)
    )
    articles["abstract"] = articles["abstract"].apply(
        lambda s: re.sub("[^0-9a-zA-Z--- -\.]+", "", s)
    )
    # articles["abstract"] = articles["abstract"].str.replace("-", " ") # NEW

    articles = articles[articles["abstract"].map(len) >= 100]
    articles["abstract"] = articles["abstract"].str.lower()

    articles = articles[articles["date_created"].str.len() >= 4]

    if "year" not in articles.columns:
        articles["year"] = articles["date_created"].str[:4].astype(int) - 2000
        articles = articles[(articles["year"] >= 0) & (articles["year"] <= 40)]
    else:
        articles["year"] = articles["year"].astype(int)
        articles = articles[articles["year"]>=2002]
    
    articles["year_group"] = articles["year"] // 5
    keep = pd.Series([False] * len(articles), index=articles.index)

    print("Applying filter...")
    if args.filter == "keywords":
        for value in args.values:
            keep |= articles["abstract"].str.contains(value)
    elif args.filter == "categories":
        for value in args.values:
            keep |= articles["categories"].apply(lambda l: value in l)
    elif args.filter == "no-filter":
        keep |= True

    articles = articles[keep == True].sample(frac=1)

    if args.constant_sampling > 0:
        articles = articles.groupby("year").head(args.constant_sampling)

    articles = articles.sample(frac=1).head(args.samples)

    articles.reset_index(inplace=True)
    articles[["article_id"]].to_csv(opj(args.location, "articles.csv"))
    print(articles)

    print("Extracting n-grams...")
    extractor = TermExtractor(
        articles["abstract"].tolist(),
        limit_redundancy=args.limit_redundancy,
        patterns=[
            ["JJ.*"],
            ["NN.*"],
            ["JJ.*", "NN.*"],
            ["JJ.*", "NN.*", "NN.*"],
            # ["JJ.*", "NN", "CC", "NN.*"],
            # ["JJ.*", "NN.*", "JJ.*", "NN.*"],
            # ["RB.*", "JJ.*", "NN.*", "NN.*"],
        ],
    )

    ngrams = extractor.ngrams(
        threads=args.threads,
        lemmatize=args.lemmatize,
        lemmatize_ngrams=args.lemmatize,
        split_sentences=args.pre_trained_embeddings and not args.use_saved_embeddings,
    )

    del extractor
    del articles["abstract"]

    if args.pre_trained_embeddings and not args.use_saved_embeddings:
        ngrams = map(
            lambda l: [
                [
                    ("_".join(n))
                    .strip()
                    .replace("-", "_")
                    .replace(".._", "")
                    .replace("_..", "")
                    for n in sent
                ]
                for sent in l
            ],
            ngrams,
        )
        ngrams = list(ngrams)

        print("Pre-training embeddings...")
        emb = GensimWord2Vec(
            [sentence for sentences in ngrams for sentence in sentences]
        )
        model = emb.create_model(
            vector_size=args.dimensions,
            window=5,
            workers=args.threads,
            compute_loss=True,
            # epochs=90,
            # min_count=30,
            epochs=80,
            min_count=15,
            sg=1,
            callbacks=[
                MonitorCallback(
                    [
                        "transformer",
                        "embedding",
                        "syntax",
                        "grammar"
                    ]
                )
            ],
        )
        model.wv.save_word2vec_format(opj(args.location, "embeddings.bin"), binary=True)
        del model

        ngrams = [
            list(itertools.chain.from_iterable(article_sentences))
            for article_sentences in ngrams
        ]
    else:
        ngrams = map(
            lambda l: [
                "_".join(n)
                .strip()
                .replace("-", "_")
                .replace(".._", "")
                .replace("_..", "")
                for n in l
            ],
            ngrams,
        )

    ngrams = list(ngrams)

    print("Deriving vocabulary...")
    voc = defaultdict(int)

    for article_ngrams in ngrams:
        _ngrams = set(article_ngrams)
        for ngram in _ngrams:
            voc[ngram] += 1

    voc = pd.DataFrame({"ngram": voc.keys(), "count": voc.values()})

    voc["df"] = voc["count"] / len(articles)
    voc.set_index("ngram", inplace=True)

    if args.min_df < 1:
        voc = voc[voc["df"] >= args.min_df]
    else:
        voc = voc[voc["count"] >= args.min_df]

    if args.max_df < 1:
        voc = voc[voc["df"] <= args.max_df]
    else:
        voc = voc[voc["count"] <= args.max_df]

    voc["len"] = voc.index.map(len)
    voc = voc[voc["len"] >= 2]

    stop_words = nltk.corpus.stopwords.words("english")
    voc = voc[~voc.index.isin(stop_words)]

    if args.blacklist is not None:
        print("Filtering black-listed keywords...")
        blacklist = pd.read_csv(args.blacklist)["ngram"].tolist()
        voc = voc[
            voc.index.map(lambda s: not any([ngram in s for ngram in blacklist]))
            == True
        ]
        print("Filtering completed.")

    voc = voc.sort_values("df", ascending=False)
    voc.to_csv(opj(args.location, "ngrams.csv"))

    voc = pd.read_csv(opj(args.location, "ngrams.csv"), keep_default_na=False)[
        "ngram"
    ].tolist()

    vocabulary = {n: i for i, n in enumerate(voc)}

    print("Filtering n-grams...")
    with mp.Pool(processes=args.threads) as pool:
        ngrams = pool.map(partial(filter_ngrams, wl=voc), ngrams)

    print("Constructing bag-of-words...")
    bow = [[vocabulary[ngram] for ngram in _ngrams] for _ngrams in ngrams]

    # if args.limit_redundancy:
    #     print("Building 'within' matrix...")
    #     with mp.Pool(processes=args.threads) as pool:
    #         within = pool.map(partial(ngram_inclusion, js=voc), voc)

    #     within = np.array(within).astype(int)

    #     print("Removing double-counting...")
    #     bow = csr_matrix(bow)
    #     double_counting = bow.dot(csr_matrix(within.T))
    #     bow = bow - double_counting
    #     print(double_counting.sum(), "redundant keywords removed")
    #     del double_counting
    #     bow = bow.todense()
    #     print((bow <= -1).sum(), "keywords had negative counts after removal")

    del ngrams

    with mp.Pool(processes=args.threads) as pool:
        bow = pool.map(partial(construct_bow, n=len(voc)), bow)

    keep = [i for i in range(len(bow)) if len(bow[i][0]) > 0]

    articles = articles.iloc[keep]
    articles[["article_id"]].to_csv(opj(args.location, "articles.csv"))

    bow = [bow[i] for i in keep]

    dataset = {
        "tokens": [bow[i][0] for i in range(len(bow))],
        "counts": [bow[i][1] for i in range(len(bow))],
        "article_id": articles["article_id"],
    }

    del bow

    with open(opj(args.location, "dataset.pickle"), "wb") as handle:
        pickle.dump(dataset, handle, protocol=pickle.HIGHEST_PROTOCOL)

    print("Training...")
    from embedded_topic_model.models.etm import ETM

    etm_instance = ETM(
        voc,
        num_topics=args.topics,
        rho_size=args.dimensions,
        emb_size=args.dimensions,
        epochs=25,
        debug_mode=True,
        train_embeddings=not args.pre_trained_embeddings,
        model_path=opj(args.location, "model"),
        embeddings=opj(args.location, "embeddings.bin")
        if args.pre_trained_embeddings
        else None,
        use_c_format_w2vec=True,
    )

    etm_instance.fit(dataset)

    with open(opj(args.location, "etm_instance.pickle"), "wb") as handle:
        pickle.dump(etm_instance, handle, protocol=pickle.HIGHEST_PROTOCOL)

    topics = etm_instance.get_topics(20)
    print(topics)
    topic_coherence = etm_instance.get_topic_coherence()
    print(topic_coherence)
    topic_diversity = etm_instance.get_topic_diversity()
    print(topic_diversity)