DYNAMICALLY SUPPORTING UNEXPLORED DOMAINS IN CONVERSATIONAL INTERACTIONS

BY ENRICHING SEMANTICS WITH NEURAL WORD EMBEDDINGS

Yun-Nung (Vivian) Chen and Alexander I. Rudnicky

1. The Task

• We propose an unsupervised

approach for acquiring open domain

knowledge based on a user’s verbal

request.

• We use structured knowledge to

extract slot types as semantic

seeds to obtain domain-related

information, and retrieve more the

most relevant applications without

supervision.

• We enable the system to properly

react to users’ queries, such as

providing relevant applications or

suggesting users install applications

that support uncovered domains,

based on their open domain requests.

probability that user speaks

Q to make the request for

launching the application A

Semantic parsing performs well on a

generic domain, but cannot recognize

domain-specific named entities.

Motivations

o A typical SDS needs a predefined task domain that supports specific functionality; it is not able to

dynamically support functions provided by newly installed or not yet installed apps.

o Structured knowledge resources are available (e.g. Freebase, Wikipedia, FrameNet) and may

provide semantic information that allows new functionality to be linked into the domain.

o Neural word embeddings can provide semantic knowledge via unsupervised training.

Approaches

1. Generating semantic seeds by using knowledge resources

2. Enriching the semantics with neural word embeddings

3. Retrieving relevant applications or dynamically suggesting users install the applications that

support new domain functionality.

Results

o Compared to original queries, using the Freebase knowledge resource (sufficient information

about named entities) to extract slot types for enriching semantics of queries achieves 25% and

18% relative improvement of MAP and P@5 respectively.

2. Framework

• Domain: 13 important application types, accessed the most frequently from Google Play

o Speech data collected from the users with intents described pictorially (WER = 19.8%)

• Main idea: Types of slots help determine semantic meaning of the utterance for expanding domain knowledge.

In an open domain, with spoken queries, how can we dynamically and effectively provide the

corresponding functions to fulfill users’ requests?

Approach ASR

MAP P@5

Original Query 25.50 34.97

Embedding-Enriched 30.42 40.72

Type-Embed.-Enriched

Frame 30.11 39.59

Wikipedia 30.74 40.82

Freebase 32.02 41.23

Hand-Craft 34.91 45.03

“play lady gaga’s bad romance”

1. Semantic Seed Generation

The Semantic Seeds (Slot Types)

2. Semantics Enrichment

Frame-Semantic Parsing

Entity Linking

Wikipedia

Freebase

Structured Knowledge

Word Embeddings

Enrichment Process

3. Retrieval Process

Ranking Model

Ranked Applications

Pandora singer

songwriter song

music :

Application Data

Query Utterance

3. Semantic Seed Generation

• Frame Type of Semantic Parsing

Q: compose an email to alex

Frame: text creation FT LU: compose FE LU: an email

Frame: contacting FT LU: email

Sfrm(Q): frame-based semantic seeds

• Entity Type from Linked Structured Knowledge

。Wikipedia Page Linking 。Freebase List Linking

Q: play lady gaga’s bad romance

… is an American singer, songwriter, and actress.

… is a song by American singer …

Swk(Q): wikipedia-based semantic seeds

celebrity composition

:

composition canonical version musical recording

:

Q: play lady gaga’s bad romance

Sfb(Q): freebase-based

semantic seeds

Enriching semantics improves performance by involving domain-specific knowledge.

Freebase results are better than the embedding-enriched method when |Q’| > 50, especially for P@5, showing that we can

effectively and efficiently expand domain-specific knowledge by types of slots from Freebase.

Hand-crafted mapping shows that the correct types of slots offer better understanding and tells the room of improvement.

4. Semantics Enrichment • Main idea: Use distributed word embeddings to obtain the semantically related

knowledge for each word.

1) Model word embeddings by using application vender descriptions.

2) Extract the most related words by trained word embeddings for

each semantic seed.

5. Retrieval Process

• Query Reformulation (Q’)

。Embedding-Enriched Query: integrates

similar words to all words in Q

。Type-Embedding-Enriched Query:

additionally adds similar words to semantic

seeds S(Q)

• Ranking Model

• Main idea: retrieve the applications that are

more likely to support users’ requests via

vender descriptions

Words with higher similarity suggest that they often occur with

common contexts in the embedding training data.

“text” “message”, “msg”

probability that word x

from Q’ occurs in the

application A

6. Experiments

Lady Gaga Bad Romance

Trailer of Iron Man 3

Alex

Alex

Alex

Alex “I can meet”

…

“I graduated”

…

…

? CMU

?

CMU

English: university

Chinese: ?

1. music listening

2. video watching

7. post to social websites

4. video chat

5. send an email

8. share the photo

6. text

9. share the video

3. make a phone call

10. navigation

11. address request

12. translation

13. read the book

0.22

0.26

0.30

0.34

0.38

0 25 50 75 100 125 150 175 200

MAP

#word / query

Baseline Embedding-Enriched (T)Type-Embedding-Enriched: Frame (T) Type-Embedding-Enriched: Wikipedia (T)Type-Embedding-Enriched: Freebase (T) Type-Embedding-Enriched: Hand-crafted (T)

0.25

0.31

0.37

0.43

0.49

0 25 50 75 100 125 150 175 200

P@5

#word / query

7. Conclusions

yvchen@cs.cmu.edu vivian.ynchen The application with higher P(Q | A) is more likely

to be able to support the user desired functions.