Why Deep Learning Is a Perfect Match for Natural Language Processing
Natural language processing is set to change the way artificial intelligence understands human desires and behaviors, and deep learning is a game-changer for the field. How will business and marketing benefit?
Sometimes it takes an attempt to build artificial intelligence (AI) to truly appreciate how complex the human mind is. That has certainly been the case for natural language processing (NLP), a branch of technology devoted to the nuances of how an AI understands human language.
For AI, the seemingly simple task of learning the difference between ‘plaster’ (a substance used for wall covering) and ‘plasters’ (Band-Aid) proves quite laborious. It means feeding the AI many pre-identified contexts for plasters so the AI would have sufficient experience to then identify whether a buyer is looking for home furnishing materials or first aid products in a search request.
For the classical method to work effectively, explains Appier’s Chief AI Scientist Min Sun, “it usually relies on humans to explicitly clarify meaning of each word and the relations between words”. It required a level of human intervention to define exceptions, labor that undercuts the gains in efficiency AI was supposed to deliver.
The classical method does result in some basic language processing, but it is like having the AI learn language and meaning by rote. Many pre-identified contexts need to be curated and a significant amount of human intervention is required. This prevents the wide adaptation of language processing in business domain.
The way humans communicate is much more complex than that. Language is highly specific; it has a context that exists outside of words – for example the relationship between a writer and an intended reader – that influences vocabulary and word choice, syntax, spelling and punctuation.
Furthermore, “language evolves through the years and the way we use language in formal documents and online will be quite different,” says Sun. “If you want the classical method to be effective, you would need to have dedicated knowledge bases for formal English, American English, Australian English… and you would also have a dedicated knowledge base for, say, Reddit types of English. That’s not scalable.”
The Relationship Between NLP and Deep Learning
Deep learning changes that. In this field, deep learning enables AI to learn the meaning of words or phrases through directly observing how they are used in a paragraph. As a result, instead of requiring pre-identified contexts or humans to clarify meanings and define relations, the meaning and relations of words or phrases are learned simply from raw paragraphs. This means that when a person searches for ‘clear plaster’, the AI understands they are more likely to mean the first aid item than the furnishing material, which brings the NLP level closer to a human learner’s level.
Deep learning does need input of a large amount of text (corpus), usually obtained from sources such as news sites, Wikipedia and Reddit comments, before the computer “learns the vector representation for each discrete token,” Sun says. A vector is where similar words are mapped in proximity to one another to indicate how those of the same context are likely to share semantic meaning.
Deep learning and vector-mapping make AI’s language processing more accurate without much human intervention, as well as opens up new opportunities for keyword-based marketing and sharper sentiment analysis.
The vectors become significant when leveraged for keyword-based marketing. A human marketer would need to ideate a list of keywords to target, but with deep learning you would just need to give the AI a ‘seed’, “and then in the vector space it could find similar keywords,” says Sun.
For example, imagine you are a travel company creating a list of keywords to target potential customers. You might add words like ‘vacation’, ‘holiday’, ‘flight’, ‘cruise’, ‘getaway’ on your list. An AI well versed in NLP would also understand that a user searching for ‘Cebu’ and ‘Philippines’ in the right context would also be interested in travel.
Secondly, deep learning-based NLP allows for better sentiment analysis, which means it is more accurate at telling whether users feel positively or negatively about their keywords. If a user searched for ‘Cebu’ and ‘earthquake’ rather than ‘Cebu’ and ‘diving’, then the AI would recognize that this user may not be a target candidate for a travel brand. “You will get a 10 to 20 percent accuracy boost right away,” Sun asserts.
Powerfully, deep learning-based NLP allows marketers to gain a detailed understanding of users, generating more valuable insights. Not only does this enable you to widen the group of people you can target, you can also target them with more relevant offers. A ‘Cebu’ plus ‘diving’ searcher may not have been included if you had only used ‘travel’ keywords to define your target market, but now they are on the radar thanks to the vectors. You can then target them with offers to that specific destination, or perhaps other diving destinations in Southeast Asia.
The Current Limits of NLP
Despite some excellent benefits, NLP still needs to bridge some gaps, particularly when it comes to features such as generating description. “The generation performance is less stable than the understanding test,” Sun admits. “Sometimes it will generate unexpected things, so humans do need to double-check it.”
The other challenging area relates to the data source you choose to train the AI on. “The risk is because it is learned directly from your data. If your data is corrupted or wrong then you will have garbage in garbage out,” he adds. This means you will still need humans to verify data sources.
Overall, however, Sun is confident that future generations of deep learning-based NLP will be fine-tuned to require less and less human intervention. It has already “reduced the human workflow and made marketing more scalable,” he says. In the next few years, watch out for greater confidence in chatbots handling sophisticated queries and further automation of marketing activities.
For now, it is clear that what deep learning-based NLP offers businesses and marketing – accurate identification of a wider target audience matched with more relevant offers – equates to a good investment. It will be exciting to see what AI can deliver next.
* Interested in finding out more about how deep learning can help you with customer acquisition and engagement? Read our latest white paper ‘Landing Valuable App Users: How Deep Learning Enhances Acquisition Campaigns’ for more in-depth insights.