Generating synonyms or similar words using BERT word embeddings

user_2599383 0
Viewed 2317

I want to generate synonyms or similar words using BERT words embeddings. I started to do this using BERT. For later software integration, it has to be done in JAVA, so I went for easy-bert (https://github.com/robrua/easy-bert).

It appears I can get word embeddings this way:

try(Bert bert = Bert.load(new File("com/robrua/nlp/easy-bert/bert-uncased-L-12-H-768-A-12"))) {
    float[][] embedding = bert.embedTokens("A sequence");
    float[][][] embeddings = bert.embedTokens("Multiple", "Sequences");

}

Do you know how I could get similars words from these word embeddings ?

Thanks for your help !

word-embedding nlp
2 Answers

Similar task for this subject (lexical substitution) would belong to LS07 and LS14. One researcher achieved the SOTA in those benchmarks using the BERT. You'd be interested in reading this paper. https://www.aclweb.org/anthology/P19-1328.pdf

The author says as below.

applies dropout to the target word’s embedding for partially masking the word, allowing BERT to take balanced consideration of the target word’s semantics and contexts for proposing substitute candidates, and then validates the candidates based on their substitution’s influence on the global contextualized representation of the sentence."

I don't know how to reproduce the same result because the implementation is not open to public. But here's the hint - the embedding dropout could be applied to generate substitute candidates.

user_148341740

I developed a way to do this using Luminoso. I work for them so this is a bit of an ad, but it does exactly what you want it to do.

https://www.luminoso.com/search

Luminoso is really good at understanding conversational text like product reviews, product descriptions, survey results and trouble tickets. It doesn't require ANY kind of training or ontology building and will build a language model around your language. You feed the text for your pages into Luminoso and it will generate a set synonyms for the concepts used in your text.

As an example project I built a search using Amazon.com beauty products. I'll just copy a couple of the automatically generated synonyms around three concepts. There were 17851 synonyms generated from this dataset.

scent, rose-like, not sickeningly, not nauseating, not overwhelming, herb-y, no sweetness, cucumber-y, not too citrus-y, no gardenia, not lemony, pachouli, vanilla-like, fragarance, not spicy, flowerly, musk, perfume-like, floraly, not cloyingly => scent
recommend, recommende, advice, suggestion, highly recommend, suggest, recommeded, recommendation, recommend this product, reccommended, advise, suggest, indicated, suggestion, advice, agree, recommend, say, considering, mentioned => recommend
bottle, no sprayer, 8-oz, beaker, decanter, push-down, dispenser, pipet, pint, not the bottle, no dropper, keg, gallon, jug, pump-top, liter, half-full, decant, tumbler, vial => bottle
eczema, non-steroidal, ulcerative, dematitis, ecsema, Elidel, dermititis, inflammation, pityriasis, hydrocortizone, dyshidrotic, chickenpox, Stelatopia, perioral, rosacea, dry skin, nummular, ecxema, mild-moderate, ezcema => eczema

There were 800k products in this search index so the results were large as well, but this will work on small datasets as well.

Besides the synonym format there you can also place this directly into elasticsearch and associated the synonyms for a specific page with that page.

This is a sample of an Elasticsearch index enhanced with the same technology. It's dialed up super high so there are too many concepts added, but just to show you how well it finds relationships between concepts.

{"index": {"_index": "amzbeauty", "_type": "_doc", "_id": "130414089X"}}
{"title": "New Benefit Waterproof Automatic Eyeliner Pen - Black - BAD Gal Liner", "text": "Length : 13.5 cm\nColor: Black\n100% Brand new and unused.\nSmudge free.\nFine-tip. Easy to blend and smooth to apply\nCan make fine and bold eyeline with new texture and furnishing.\nProvide rich and consistant colour\nLongwearing and waterproof\nFregrance Free", "primary_concepts": ["not overpoweringly", "concoction", "equipped", "fine-tip", "water-resistant", "luxuriant", "make", "fixture", "☆", "not lengthen", "washable", "not too heady", "blendable", "doesn't collect", "shade", "niche", "supple", "smudge-proof", "sumptuous", "movable", "black", "over-apply", "quick", "silky", "colored", "sweatproof", "opacity", "accomodate", "fuchsia", "furnishes", "meld", "sturdily", "smear", "inch", "mid-back", "chin-length", "smudge", "alredy", "not cheaply", "long-wearing", "eyeline", "texture", "steady", "no-name", "audacious", "easy", "edgy", "is:A", "marketers", "greys", "decadent", "applicable", "Crease-free", "magenta", "free", "itIn", "stay-true", "racy", "application", "glides", "smooth", "sleek", "taupe", "grainy", "dark", "wealthy", "JP7506CF", "gray", "grayish", "width", "newness", "purfumes", "Lancme", "blackish", "easily", "doesn't smudge", "maroon", "blend", "convenient", "smoother", "Moschino", "long-wear", "mauve", "medium-length", "no raccoon", "revamp", "demure", "richly", "white", "brand", "offers", "lenght", "soft", "doesn't smear", "provide", "provides", "unusable", "eye-liner", "unopened", "straightforward", "silky-smooth", "uniting", "compactness", "bold", "fearless", "mix", "indulgent", "brash", "serviceable", "unmarked", "not musky", "constructed", "racoon", "smoothly", "sealant", "merged", "boldness", "reuse", "unused", "long", "Kors", "effortless", "luscious", "stain", "rich", "discard", "richness", "opulent", "short", "consistency", "fine", "sents", "newfound", "fade-resistant", "mixture", "hue", "sassy", "apply", "fragnance", "heathy", "adventurous", "not enthusiastic", "longwearing", "fregrance", "non-waterproof", "empty", "lashline", "simple", "newly", "you'r", "combined", "no musk", "mingle", "waterproof", "painless", "pinkish", "thickness", "clump-free", "gos", "consistant", "color", "smoothness", "name-brand", "new", "smudgeproof", "yaaay", "water-proof", "eyemakeup", "not instant", "spidery", "furnish", "tint", "product", "reapply", "not black", "no globs", "imitators", "blot", "cinch", "uncomplicated", "untouched", "length"], "related_concepts": ["eyeliner", "no goofs", "doesn't smear", "pen", "hundreds"]}
{"index": {"_index": "amzbeauty", "_type": "_doc", "_id": "130414643X"}}
{"title": "Goodskin Labs Eyliplex-2 Eye Life and Circle Reducer - 10ml", "text": "Eyliplex-2 is a dual solution that focuses on the problematic eye area. This breakthrough, 24-hour system from the scientists at good skin pharmacy visibly tightens eye areas while reducing dark circles. 0.34 oz. each. 64% of subjects reported younger looking eyes immediately and a 20% reduction in the appearance of dark circles in clinical studies.", "primary_concepts": ["coloration", "Laboratories", "oncology", "cornea", "undereye", "eye", "immediately", "☆", "teen", "dry-skin", "good", "eyelids", "puffiness", "behold", "research", "temperamental", "dermatological", "breakthrough", "study", "store", "nice", "lasik", "instantaneously", "teenaged", "multi", "rheostat", "dermatology", "chemist", "invisibly", "PhD", "pharmacy", "alredy", "not cheaply", "optional", "pharmacist", "Obagi-C", "topic", "supermarket", "reversible", "studies", "Younger", "medically", "report", "thermo", "tightness", "dual", "eliminate", "researcher", "Minimization", "cutaneous", "hydration", "O2", "taupe", "increase", "moisturization", "dark", "preliminary", "excellent", "Quad", "well", "appearance", "dusky", "quickly", "instantly", "CVS", "Dermal", "great", "revolutionary", "biologist", "epidermis", "blackish", "disclosed", "problem", "youngsters", "murky", "scientific", "teenager", "oz", "dark circles", "clinically", "emphasis", "absorption", "skin", "loosen", "intractable", "technological", "reduction", "clinician", "nutritional", "forthwith", "grocer", "scientifically", "swiftly", "examination", "state-of-the-art", "not acne prone", "zone", "decrease", "younger-looking", "excellently", "troublesome", "system", "radius", "tighten", "FDA", "decent", "noticeably", "WD-40", "clearer", "scientist", "saggy", "significantly", "improvement", "Teamine", "interchangeable", "visible", "visable", "no fine line", "shortly", "minimize", "survey", "problematic", "young", "glance", "racoon", "vicinity", "youthful", "exacerbated", "focal", "region", "groundbreaking", "reddish", "focus", "reduce", "increments", "nad", "fasten", "area", "soon", "complexion", "squinting", "look", "grocery", "eyliplex-2", "Eyliplex-2", "subsequently", "even-toned", "bothersome", "eyes", "mitigate", "markedly", "philosophy:you", "difficult", "darkish", "bluish", "satisfactory", "darken", "epidermal", "lessen", "appearence", "ocular", "ergonomically", "diminished", "progression", "purplish", "sun-damaged", "Cellex-C", "visibly", "diagnosis", "drugstore", "under-eye", "apothecary", ":-D", "terrific", "clinical", "oz.", "Endocrinology", "time-released", "Nouriva", "tight", "adolescent", "subject", "eyeballs", "sking", "Pro-Retinol", "aggravate", "younger", "shortcomings", "solution", "assess", "promptly", "teenage", "Kinetin", "24-hour", "Mart", "youth", "visibility", "scientists", "taut", "better", "eyesight", "no dark circles", "not reduce", "photoaging", "Pending"], "related_concepts": ["A22", "A82", "Amazon", "daytime", "HK", "nighttime", "smell", "dark circles", "purchased"]}
{"index": {"_index": "amzbeauty", "_type": "_doc", "_id": "1304146537"}}

Luminoso uses word embeddings from ConceptNet which it also develops and the technology is above and beyond what ConceptNet gives you. I'm biased, but every time I've run data through it I'm amazed. Not free, but it really works with absolutely zero pre-training of the data and nothing is actually free.

user_148664750
Related