=Paper=
{{Paper
|id=Vol-2068/wii3
|storemode=property
|title=Supporting Real Estate Search Through Automatic Information Suggestion
|pdfUrl=https://ceur-ws.org/Vol-2068/wii3.pdf
|volume=Vol-2068
|authors=Goro Otsubo
|dblpUrl=https://dblp.org/rec/conf/iui/Otsubo18a
}}
==Supporting Real Estate Search Through Automatic Information Suggestion==
https://ceur-ws.org/Vol-2068/wii3.pdf
Supporting Real Estate Search Through Automatic
Information Suggestion
Goro Otsubo
LIFULL Co., Ltd.
Tokyo, Japan
ohtsubogoro@lifull.com
ABSTRACT in most cases the user does not know what type of detailed
Searching real estate property is uncommon task for most of search conditions are available and how to set them. As a
the users. As a result, the user is not familiar with the detailed result, they are not used effectively in the real estate search
search condition which is useful for search. In this paper, we process.
propose to use voice recognition as a support for real estate
To solve this issue, we assumed that we need support system
property search. First user set vague search condition with
which can recommend proper detailed search condition to
GUI. Then system listens conversation between users. From
assist real estate search. To do that, we have chosen to utilize
the conversation, system extracts keyword and suggest detailed
the conversation between users. Asking detailed demand for
search condition and real estate information search results with
new real estate property is sometimes intimidating for the user.
those conditions. We will discuss system design, algorithm
Moreover, the user may not necessary know the proper word to
used to link spoken words and detailed search condition and
search the real state information. Therefore, we assumed that
preliminary test results.
the system which can extract detailed search condition from
ACM Classification Keywords the casual user conversation would be effective to support the
H.5.m. Information Interfaces and Presentation (e.g. HCI): real estate property search.
Miscellaneous; See http://acm.org/about/class/1998/ for the Interfaces for searching information using voice have been
full list of ACM classifiers. This section is required. studied for a long time, and commercial service which uti-
lizes voice interface become popular in recent years due to
Author Keywords popularization of smartphones and home devices. Recently
Voice Recognition; Multi modal; GUI speech recognition on the server side has become widespread,
the accuracy of speech recognition has dramatically improved
INTRODUCTION even for unspecified speakers.
In Japan, people will move to new real estate property only six
times for the lifetime on average. In other words, searching However, in most cases speech recognition is used for simple
real estate property is "An event that happens once in a decade". web search and search in the app store[10]. This is due to the
As a result it is very difficult to become familiar with special fact that even though the precision of speech to text conversion
terms which are used to search real estate property. has improved, the next step which is "understanding meaning"
remains as tough problem yet. According to the research
In recent years real estate information sites on the Internet have by Luger et al.,[8] although the user’s expectation for speech
been widely used. Using those sites, the user can access and recognition is high, current speech recognition interface makes
search real estate information by themselves. However, search the user feel very stressful when using it.
interface of those sites are like query interface of the property
database. They have every function available to search the Considering technical difficulty described above, we have cho-
database, but they are not not necessarily easy to use for all sen to use voice recognition as a support role. The user is
the user who has interest in real estate property. not assumed to directly talk to the search interface. Instead,
system listens the conversation between users. The system
Typical user can set the vague request to search the real estate recognizes the keyword in the conversation, and tries to ex-
property like the town they want to search, layout of property plore the appropriate detailed search condition to search the
and price. Further, for example, it is also possible to set real estate property. Matching spoken word by the user and
detailed search conditions such as "Pets allowed". However, detailed search condition is the first challenge in this research.
The frustration that the user feels when using the voice inter-
face is that the system can not understand the meaning of the
word they have pronounced. We need to match appropriate
detailed search condition to words in users’ conversation.
There are various attempt to incorporate speech interface in
search as a support role for exploratory search. Andolina
©2018. Copyright for the individual papers remains with the authors. et al. proposed systems [4][5] that extract keywords from
Copying permitted for private and academic purposes.
WII’18, March 11, 2018, Tokyo, Japan
usersâĂŹ conversations to stimulating human creative think- solve the word mismatch problem, many different approaches
ing.We assumed similar approach could be effective in real have been proposed.
estate search.
In this research, we tried to utilize Word2Vec [9]. By using
Other reason why an existing system using speech recognition Word2Vec, each word is vectorized and similarity between
makes a user feel frustrated is that the system is not transparent words can be calculated. Basic idea in this research is that we
[8]. In other words, the user has no way of knowing why the tried to match spoken words and detailed search condition not
system returned a response when the misunderstood answer with simple word match, but with match between related words
came back. For this reason, we need to make matching pro- expanded from spoken words and detailed search condition
cess between spoken and recognized word and recommended using Word2Vec. Following is the algorithm we used.
detailed condition as clear as possible. By showing every
keywords searched during the detailed search condition and 1. For each detailed search condition, manually set two to five
relation between them, we will be able to achieve that goal. related keywords.
Also, attempt to avoid keyword input by selecting and manip- 2. For each manually set keyword, search related words us-
ulating suggested keywords by touch has been proposed[7] . ing Word2Vec, and record them as "expanded related key-
By using similar interaction, we assumed that we will be able words".
to increase the effectiveness of the system. Even if system can
not recognize the user’s intention correctly, user may be able 3. Convert conversation voice to text string using speech to
to find and select interesting keywords displayed on the screen text conversion function.
and explore related detailed search condition.
We will describe the design of the system below. 4. Extract noun and verb from the recognized text, and record
them as "spoken words"
SYSTEM DESIGN
5. For each spoken words, search related words using
Screen shot of developed system is shown in Figure 1. Word2Vec, and record them as "expanded related key-
User can set the search conditions which are area to search, words".
layout and price of real estate property via GUI. After set-
ting the search condition, users have free conversation about 6. Tries to find same word from expanded keywords from
their intention for new house between themselves. System spoken words, and expanded keywords from manually set
continuously monitors their conversation and recognized text keywords. If same word can be found, put link between
is shown in the lower part of the screen. System automati- spoken word and detailed search condition.
cally and continuously extract keywords from conversation Example of matching is shown in Figure 2.
and tries to find the related detailed search condition such as
"Within 800 meters from convenience store", "Pet allowed". Next we will discuss the training data set for Word2Vec. Re-
System also display the search result of the real estate property lated words extracted using Word2Vec depends on nature of
specifying each detailed condition. Link is shown between training data set. As an example, we will show the result of re-
detailed search condition and searched property. There are two trieved related words of the word "Noise" using Wikipedia[3]
major challenges in developing the proposed system. First, as training data in Table 1.(Original words are in Japanese)
we needed to develop algorithm which search detailed search Generally, words used in connection with "Noise" are lined
condition using user’s conversation data. Second, we need in- up, but they are different from what the user associates when
teraction interface which will effectively support user to search searching for a real estate property. Users who are concerned
real estate information even if the system does not recognize about "Noise" may choose "Top floor" detailed search condi-
the user’s intention correctly. We will discuss these challenges tion if they are concerned about the noise from the floor above.
next. Users who care about the noise from the roads may choose
"Higher than the second floor". In either case, those detailed
MATCHING ALGORITHM BETWEEN DETAILED SEARCH search conditions has no relations to the words extracted using
Wikipedia as a training data.
CONDITION AND SPOKEN WORDS
One of the major frustrations felt using existing speech inter- We have also gathered text data from the web site called All
face is that the system recognizes only programmed keywords About Japan [1] which has large amount of text related to real
while there is no clue for the user about which word to speak. estate property. We found out that related words extracted
As a result, quite often system does not recognize the word is more suitable for user’s intuition when searching the real
that the user pronounced. Even though the user pronounces the estate property. Retrieved related words of the word "Noise"
word which has similar meaning to the keyword that system using All About Japan as training data is also shown in Table 1.
recognizes, system can not understand the similarity between "Sound leak" is what user may be interested in when searching
those words. real estate property.
Similar difficulty is recognized in the field of question and However, number of words in All About Japan is not necessary
answer retrieval(herein Q&A retrieval) task[6] . Major chal- enough. As a result, in many cases we could not extract related
lenge for Q&A retrieval is word mismatch between the user’s words because spoken word does not exist in All About Japan
question and the question-answer pairs in the archive [11]. To data set.
Figure 1. Screen shot of developed system
Rank Wikipedia All About Japan The other aspect that we should consider is transparency of
1 Soot Water leakage the system which is described earlier. To ease user frustration,
2 Air pollution Shaking we need to make inference process clear to the user.
3 Salinity Suspicious person Considering these factors, we designed the interaction inter-
4 Exhaust gas Sound leak face shown in Figure 1. Not only detailed search condition
5 Interference Exhaust gas and spoken words are displayed, but also manually set key-
Table 1. Related words of "Noise" searched using different training data words and expanded keywords are shown on the screen. All
the words displayed on the screen can be used as an search
keyword by dragging and dropping the word into the text area
shown in the lower part of the screen.
Considering characteristics of each data set described above, As stated before, precision of voice-to-text speech and search
we decided to use both data set for keyword extraction. First of detailed search condition is not necessary high. In that case,
we tries to extract keywords from All About Japan data set. If user may be frustrated if we only display "no results found"
no matching word is found, we tries to extract keywords using or totally irrelevant result. In this research we tried to display
Wikipedia data set. as much words as possible on the screen. By viewing those
words, it it probable that some words on the screen might be
interesting for the user. If so, user can start new search via
interaction on the display, not by voice recognition.
USER INTERACTION
By using the algorithm described above, we can expect that
we will be able to search the detailed search condition better EVALUATION
than simple word match. However, still there are errors and we We have conducted two types of evaluation so far. First, we
don’t expect that we can search the detailed search condition evaluated how effective search algorithm is for various user
with high precision. input. Since we would like to evaluate effectiveness of our
Figure 2. Concept diagram to match spoken words and detailed search condition
algorithm, we used text input rather than voice input to avoid to use the voice interface as an main interface. Even though
error caused by speech recognition. current system only extracts and search with detailed search
condition, the user tried to search with voice sentence like
We have chosen fifty sentences from Q&A site about real
"Search property in Tokyo area". Even though we designed
estate search[2] such as "Good view and well-ventilated" . All
the system so that the user can set those search condition with
the sentences do not include the exact word in detailed search
GUI, the user often forget that.
condition. Therefore, none of the conditions can be searched
using simple word match algorithm. For each sentence, we From the result of user test, we realized that interaction design
have selected corresponding detailed search condition. System should be improved so that user does not misunderstand that
displays up to four detailed search condition. If at least one system can recognize every request that user might have. In
search condition displayed is related to input sentence, we the current design, search condition setting GUI is hidden
evaluated it as a success. while the system listen the conversation. In this case, the
user expects system can understand any word what they say.
As a result, 34 out of 50(68%) sentences can be evaluated as
To avoid such a misunderstanding, we need to show search
success. 9 out of 34 sentences which are evaluated as success condition setting GUI upfront. When the user speaks, and do
include manually defined extend keywords. If we remove not operate the GUI, we can show current voice recognition
them from the result, 25 out of 41 (61%) sentences can be interface over the search condition setting GUIs.
searched successfully using proposed algorithm.
Second, using current system, we conducted simple user eval-
uation. We have explained system’s concept and function, and
put the system beside, we had conversation about what type CONCLUSION AND FUTURE DIRECTION
of real estate property the user is interested in. Five users have We have developed the real estate search system which rec-
participated in the test. During the user test, we got consistent ommends detailed search conditions from users’ conversation.
response from the user. Every participants see the importance Evaluation result shows the mixed results. We could confirm
of system’s assistant role. In some cases, the user can find the the potential of proposed algorithm. However, we also recog-
interesting results. Four out of five participants see detailed nized that user interface design should be improved. Based on
search condition that they have never searched real estate prop- the evaluation results, we will redesign the system interaction,
erty with before. We also observed that the user quite often try and will conduct further user test to evaluate how effective
system can support the search of real estate information.
REFERENCES 7. Khalil Klouche, Tuukka Ruotsalo, Diogo Cabral,
1. 2017. All Abount Japan : House,Real estate property. Salvatore Andolina, Andrea Bellucci, and Giulio Jacucci.
https://allabout.co.jp/r_house/. (2017). 2015. Designing for Exploratory Search on Touch
Devices. In Proceedings of the 33rd Annual ACM
2. 2017. OKWeb-Sumai(Housing).
Conference on Human Factors in Computing Systems
https://okwave.jp/c622.html. (2017).
(CHI ’15). ACM, New York, NY, USA, 4189–4198. DOI:
3. 2017. Wikipedia (Japanese). https://ja.wikipedia.org/. http://dx.doi.org/10.1145/2702123.2702489
(2017).
8. Ewa Luger and Abigail Sellen. 2016. "Like Having a
4. Salvatore Andolina, Khalil Klouche, Diogo Cabral, Really Bad PA": The Gulf Between User Expectation and
Tuukka Ruotsalo, and Giulio Jacucci. 2015. Experience of Conversational Agents. In Proceedings of
InspirationWall: Supporting Idea Generation Through the 2016 CHI Conference on Human Factors in
Automatic Information Exploration. In Proceedings of Computing Systems (CHI ’16). ACM, New York, NY,
the 2015 ACM SIGCHI Conference on Creativity and USA, 5286–5297. DOI:
Cognition (C&C ’15). ACM, New York, NY, USA, http://dx.doi.org/10.1145/2858036.2858288
103–106. DOI:
9. Thomas Mikolov. 2013. word2vec:Tool for computing
http://dx.doi.org/10.1145/2757226.2757252
continuous distributed representations of words. (2013).
5. Salvatore et al Andolina. 2015. IntentStreams: Smart https://code.google.com/word2vec/.
Parallel Search Streams for Branching Exploratory
10. Verto. 2017. WhatâĂŹs the Future of Personal Assistant
Search. In Proceedings of the 20th International
Apps? (2017). http://research.vertoanalytics.com/
Conference on Intelligent User Interfaces (IUI ’15).
what-the-future-of-personal-assistant-apps-webinar-deck.
ACM, New York, NY, USA, 300–305.
11. Xiaobing Xue, Jiwoon Jeon, and W. Bruce Croft. 2008.
6. Jiwoon Jeon, W. Bruce Croft, and Joon Ho Lee. 2005.
Retrieval Models for Question and Answer Archives. In
Finding Similar Questions in Large Question and Answer
Proceedings of the 31st Annual International ACM SIGIR
Archives. In Proceedings of the 14th ACM International
Conference on Research and Development in Information
Conference on Information and Knowledge Management
Retrieval (SIGIR ’08). ACM, New York, NY, USA,
(CIKM ’05). ACM, New York, NY, USA, 84–90. DOI:
475–482. DOI:
http://dx.doi.org/10.1145/1099554.1099572
http://dx.doi.org/10.1145/1390334.1390416