A fundamental goal of search engines is to identify, given a query, documents that have relevant text. This is intrinsically difficult because the query and the document may use different vocabulary, or the document may contain query words without being relevant. We investigate neural word embeddings as a source of evidence in document ranking. We train a word2vec embedding model on a large unlabelled query corpus, but in contrast to how the model is commonly used, we retain both the input and the output projections, allowing us to leverage both the embedding spaces to derive richer distributional relationships. During ranking we map the query words into the input space and the document words into the output space, and compute a query-document relevance score by aggregating the cosine similarities across all the query-document word pairs.
We postulate that the proposed Dual Embedding Space Model (DESM) captures evidence on whether a document is about a query term in addition to what is modelled by traditional term-frequency based approaches. Our experiments show that the DESM can re-rank top documents returned by a commercial Web search engine, like Bing, better than a term-matching based signal like TF-IDF. However, when ranking a larger set of candidate documents, we find the embeddings-based approach is prone to false positives, retrieving documents that are only loosely related to the query. We demonstrate that this problem can be solved effectively by ranking based on a linear mixture of the DESM and the word counting features.
AWS Community Day CPH - Three problems of Terraform
Dual Embedding Space Model (DESM)
1. Dual Embedding Space Model (DESM)
Bhaskar Mitra, Eric Nalisnick, Nick Craswell and Rich Caruana
https://arxiv.org/abs/1602.01137
2. How do you learn a neural embedding?
Setup a prediction task
Source Item → Target Item
(The bottleneck layers are crucial for generalization)
Target
item
(sparse)
Source
item
(sparse)
Source
embedding
(dense)
Target
Embedding
(dense)
Distance
Metric
The bottleneck
Word2vec
Mikolov et. al. (2013)
Word → Neighboring word
I/O: One-Hot
DSSM (Query-Document)
Huang et. al. (2013), Shen et. al. (2014)
Query → Document
I/O: Bag-of-trigrams
DSSM (Session Pairs)
Mitra (2015)
Query → Neighboring query in session
I/O: Bag-of-trigrams
DSSM (Language Model)
Mitra and Craswell (2015)
Query prefix → query suffix
I/O: Bag-of-trigrams
3. Not all embeddings are created equal
The source-target training pairs strictly dictate what notion of
relatedness will be modelled in the embedding space
Is eminem more similar to rihanna or rap?
Is yale more similar to harvard or alumni?
Is seahawks more similar to broncos or seattle?
(Be careful of using pre-trained embeddings as inputs to a different model –
one-hot representations or learning an in situ embedding may be better!)
4. Word2vec
Learning word embeddings based
on word co-occurrence data.
Well-known for word analogy tasks,
[king] – [man] + [woman] ≈ [queen]
What if I told you that everyone
who uses Word2vec is throwing half
the model away?
5. Typical vs. Topical Relatedness
The IN-IN and the OUT-OUT similarities cluster words that occur in the same context
and therefore of the same Type. The overall word2vec model is trained to predict
neighboring words. Therefore the IN-OUT similarity clusters words that commonly co-
occur under the same Topic.
6. Typical embeddings for Web search?
B. Mitra and N. Craswell. Query
auto-completion for rare prefixes.
In Proc. CIKM. ACM, 2015.
7. Which passage is about Albuquerque?
Traditionally in Search we look for evidence of
relevance of a document to a query in terms
of the number of matches of the query
terms in the document.
But there is useful signal in the non-matching
terms in the document about whether the
document is really about the query terms, or
simply mentions them.
A word co-occurrence model can be used to
check if the other words in the document
support the presence of the matching terms.
Passage about Albuquerque
Passage not about Albuquerque
8. Dual Embedding Space Model
• All pairs comparison between query
and document terms
• Document embedding can be pre-
computed as the centroid of all the
unit vectors of the words in the
document
• DESMIN-OUT uses IN-embeddings for
query words and OUT-embeddings
for document words
• DESMIN-IN uses IN-embeddings
document words as well
10. Because Cambridge is not an African mammal
DESM = ✔
BM25 = ✔
DESM = ✘
BM25 = ✔
DESM = ✔
BM25 = ✘
Query: cambridge
11. Telescoping Evaluation
As a weak ranking feature DESMIN-OUT performs better than BM25,
LSA and DESMIN-IN models on a UHRS (Overall) set and a click based
test set.
12. Full retrieval evaluation
The DESM models only a specific aspect of document relevance. In the presence
of many random documents (distractors) it is susceptible to spurious false
positives and needs to be combined with lexical ranking features such as BM25