Bill Slawski and I recently discussed a new algorithm via email. Bill pointed out a particular research paper and patent that expanded my thoughts beyond Neural Matching and RankBrain.
Recent algorithm research emphasizes understanding content and search queries, which may help clarify certain changes we’ve observed.
The Difference Between RankBrain and Neural Matching
Google has officially explained RankBrain and Neural Matching through Danny Sullivan’s tweets. RankBrain aids Google in associating pages with concepts, mainly by identifying synonyms for words found on a page. In contrast, Neural Matching helps relate words to searches, finding synonyms for search terms. These systems surpass traditional synonym approaches and offer a novel way to understand concepts. For instance, Neural Matching helps recognize that a search for “why does my TV look strange” relates to “the soap opera effect," allowing Google to return relevant results even without exact term matches.
CLSTM and Its Relation to Neural Matching
Bill Slawski highlighted the Contextual Long Short Term Memory (CLSTM) Models for Large Scale Natural Language Processing (NLP) Tasks paper as potentially related. This 2016 paper focuses on understanding context through three sentences using the word "magic." It highlights that incorporating contextual features into the LSTM model can predict word use and sentence structure based on text segment topics, such as literature or music, which can enhance applications like word prediction, next sentence selection, and understanding user intent.
Question Answering Algorithm
A 2019 research paper, "A Hierarchical Attention Retrieval Model for Healthcare Question Answering," appears to refine this algorithm. It proposes a neural network for ranking documents in question-answering, using deep attention at word, sentence, and document levels, to identify relevant information efficiently.
Attention-Based Neural Matching
A 2018 non-Google paper, "aNMM: Ranking Short Answer Texts with Attention-Based Neural Matching Model," may align with Google’s Neural Matching. Although it’s not explicitly tied to Google, it leverages attention over word embeddings for semantic matching, potentially offering insights into Google’s approach.
Neural Matching in Google’s Context
Danny Sullivan referred to Neural Matching as an AI system initiated in 2018 to better relate words to concepts, dubbed a super-synonym system. This suggests that Neural Matching may represent a suite of algorithms working harmoniously.
Takeaways
- Avoid Synonym Spamming: Danny’s description shouldn’t lead to synonym overuse in content. The underlying algorithms are more advanced.
- Focusing on Content Structure: Authors should focus on clear topic assignment within words, sentences, and paragraphs. Planning and writing carefully can aid this.
- Understanding Content Types: While not favoring long-form content, Google’s interest in understanding such content suggests evaluating content types if encountering rankings changes.
The Google Dance
Historically, Google updated once a month, known as the Google Dance. Today, they refresh the index daily, with several annual updates to improve understanding of search queries and content.
More Resources
Images were modified by the author and screenshots were provided accordingly.
