• No results found

An example process of mining user requests linguistic rules

3.4 Automatic user requests elicitation

3.4.3 An example process of mining user requests linguistic rules

How to balance the trade-off between the linguistic rules number and coverage and how to tackle the data sparsity problem can be illustrated in an example process of mining a set of user request linguistic rules in this subsection. In this example, the starting clue is this sentence:

“Miss seeing Lily's antics when I enter a reading or food, and can you add half a second before she starts?”

From this sentence, “can you add” is a set of keywords that can help to elicit the request “half a second before she starts”. Searching this clue “can you add” against the whole 21280 user reviews, gives 4 sentences containing this set of keywords, and all the 4 sentences are explicitly expressing user requests. However, a search for other keyword combinations derived from this clue should be carried out against the whole dataset.

The search can expand to all possible keyword combinations or ones derived from similar keywords. These keyword combinations could be “can you”, “could you”, “would you”, “can I”, “could I”, “would I”. All these keywords can match a pattern: a modal verb + a personal pronoun + a base form verb.

However, inventing keyword combinations is not a good way to create linguistic rules in practice. These combinations have to be verified through the real dataset. The verification results on the 21280 user reviews in dataset one are below. For each keyword combination, a search is performed on dataset one to find how many times they occur. Among the occurrences, percentages of containing user requests are calculated and reported in Table 17.

85

Table 17 Initial keyword combinations and their occurrences in dataset one

Keyword combinations Number of occurrences Percentage of containing user requests

could you + <VB> 11 91%

could I + <VB> 0 0%

would you + <VB> 4 50%

would I + <VB> 2 0%

can you + <VB> 27 59%

can I + <VB> 22 50%

It is obvious that the two combinations that have 0 % user request should be ruled out. But none of the rest of combinations has good enough precision to become a linguistic rule as it is. Therefore, the author took the occurrences of each of the rest combinations for further breaking down. Taking the example of the 11 occurrences of the first combination, “could you + <VB>”, the data are analysed below.

Table 18 Request analysis for “could you + <VB>"

reviewId appId text Request

/ Not

6625 4

It would be great if you could use the fool database for FatSecret, as it has food entries specific to various countries.

It would also be great if the app could access the food data from Google Fit, as you can use other apps to write data to it. Also could you guys look into adding calories burned with exercise please. Including the calories burned from apps such as Google Fit Then the app would get 5 stars. PS I am a paying subscriber

Request

7484 4 Very useful app. Could you please add Basaglar? Request 8222 4 Could you add Persian language please? Request

86

8338 4

Could you install a sick explanation for high numbers? I have been very ill and my numbers were High. I did leave notes, but think it would be clearer to tab sick into my log

with the numbers. Thanks Request

9821 5 This is a great app but 2 things are missing and they are after

dinner and before bed could you please add these ... Request

10053 5

Easy to use and meets all my requirements. My Wish List:

When you export the dates, could you put them in yyyymmdd format please - or at least an option. Not everyone in the world uses mm/dd/yyyy and it can be quite

confusing explaining it to a non-techie doctor.

Request

10422 5

This app is great for me. I'm a newly diagnosed diabetic and it offers everything I need. Granted I use it for just my glucose level, the app itself has numerous other uses. From a food diary to the ability to record blood pressure the app is great! With all this and it's free, how could you need

something else?

Not

16270 6

Hello the app looks good. Haven't had a chance to really try it out. Could you please add the feature of changing the

language to spanish???? Thanks Request

20012 17

I have been using ontrack for more than a year now and have nothing but good things to say. I email and print graphs for my doctor, he has come to expect the reports. My doctor also recommends OnTrack to his patients. And free too! What

else could you want....

Not

20793 17

Could you add a bmi calculation? Then I don't need another program any more ... the graphs seems a bit compressed too,

but I love your program.

Request

87

21051 18

Hi, could you please fix the bug which causes the Previous Readings at Blood Glucose to appear distorted and abnormally large, probably because of the SHARE icon?

Sony Xperia M Not

From these 11 occurrences in Table 18, “could you + <VB>” can be subsumed into the following rule.

Table 19 A sample rule for “could you + <VB>"

(how | What else) Could | could you (guys) (please) add | put | install | look (into) <request>

By explanation, this rule is looking for “Could you” or “could you” followed by a base form verb “add”, “put”, “install” or “look into”. Moreover, “guys” or “please” can optionally appear before the base form verb. But there must not be “how” or “What else” appearing before “Could you” or “could you”. Then the text after the base form verb is the user request to elicit according to this rule. At the time of writing, this rule has 100 % precision against dataset one.

However, the author was trying to group user request linguistic rules as much as possible and balance the trade-offs between the number of rules and the performance of specific rules.

After the consideration for the trade-off, the author decided to subsume the resulting linguistic rules of three keyword combinations, “can you + <VB>", “could you + <VB>", and “would you + <VB>", into one rule (77), but leave “can I + <VB>" to another rule (78).

Therefore, this process eventually produced two rules that cover occurrences of 44 and 22 in the dataset one respectively.

Table 20 One clue leads to two linguistic rules covering 66 occurrences in dataset one

77 (how | nor | why | what | where | what more | what else | how often | not only | no longer) can | could | would you (guys) (please) add | put | make | give | input | change

| create | install | consider | look (into) <request>

78 can I | i have + <NN|NNS|NNP|NNPS> -> <request>

88

The ways of coding these two rules are introduced next.